buildings

Article
Machine Learning for Real-Time Building Outdoor Wind
Environment Prediction Framework in Preliminary Design:
Taking Xinjiekou Area of Nanjing, China as the Case
Lin Sun and Guohua Ji *

School of Architecture and Urban Planning, Nanjing University, 22 Hankou Road, Nanjing 210093, China
* Correspondence: jgh@nju.edu.cn

Abstract: The incorporation of physical environmental performance as a primary consideration in

building design can facilitate the harmonization of the built environment with the surrounding site
and climate, enhance the building’s environmental adaptability and environmental friendliness,
and contribute to the achievement of energy-saving and emission-reduction objectives through the
integration of natural lighting and ventilation. General computational fluid dynamics (CFD) can
help architects make accurate predictions and effectively control the building’s wind environment.
However, CFD integration into the design workflow in the preliminary stages is frequently chal-
lenging due to program uncertainty, intricate parameter settings, and substantial computational
expenses. This study offers a methodology and framework based on machine learning to overcome
the complexity and computational cost barriers in simulating outdoor wind environments of build-
ings. In this framework, the machine learning model is trained using an automated CFD simulation
system based on Butterfly and implemented within the Rhino and Grasshopper environment. This
framework provides real-time simulation feedback within the design software and exhibits promising
accuracy, with a Structural Similarity Index Measure (SSIM) ranging from 90–97% on a training
dataset of 1200 unique urban geometries in Xinjiekou Area of Nanjing, China. Furthermore, we
programmatically integrate various parts of the simulation and computation process to automate
Citation: Sun, L.; Ji, G. Machine multiscenario CFD simulations and computations. This automation saves a significant amount of
Learning for Real-Time Building time in producing machine-learning training sets. Finally, we demonstrate the effectiveness and
Outdoor Wind Environment accuracy of the proposed working framework in the design process through a case study. Although
Prediction Framework in Preliminary our approach cannot replace CFD simulation computation in the later design stages, it can support
Design: Taking Xinjiekou Area of architects in making design decisions in the preliminary stages with minimal effort and immediate
Nanjing, China as the Case. Buildings performance feedback.
2024, 14, 2613. https://doi.org/
10.3390/buildings14092613 Keywords: urban and rural areas; building outdoor wind environment simulation; sustainable
Academic Editors: Francesco Nocera development; machine learning; preliminary design stage
and Chi Yan Tso

Received: 11 June 2024

Revised: 31 July 2024
1. Introduction
Accepted: 21 August 2024
Published: 23 August 2024 A comfortable wind environment within an urban area enhances outdoor space
comfort and minimizes negative impacts on the surrounding environment, such as the
mechanical effects of wind and thermal discomfort. Therefore, including wind-related
performance in building and urban design projects is practical. Computational fluid dy-
Copyright: © 2024 by the authors. namics (CFD) simulations have been extensively used to assess and enhance building wind
Licensee MDPI, Basel, Switzerland. performance [1]. Various CFD simulations and analysis software have been developed,
This article is an open access article including ANSYS Fluent [2], OpenFOAM [3], and CHAM PHOENICS [4]. These software
distributed under the terms and programs can accurately simulate and analyze the wind environment around buildings.
conditions of the Creative Commons Despite the abundance of CFD tools, most are developed by technical researchers
Attribution (CC BY) license (https://
focusing on empirical and analytical validations [5]. However, building design involves an
creativecommons.org/licenses/by/
iterative process that frequently necessitates evaluating building performance efficiency to
4.0/).

Buildings 2024, 14, 2613. https://doi.org/10.3390/buildings14092613 https://www.mdpi.com/journal/buildings

Buildings 2024, 14, 2613 2 of 26

compare various design alternatives. Some research findings indicate that many CFD tools
are challenging to integrate into architect workflows because they ignore their needs [6–8].
According to several architects, most CFD tools are too complex and burdensome [9],
making it challenging to obtain prompt feedback to guide design decisions during the
preliminary design stages. Consequently, the gap between architects using CFD tools and
the tools themselves has never been bridged [10].
This study aims to overcome the deficiency by introducing an auto-wind framework.
This innovative platform provides real-time feedback on building wind environment infor-
mation by seamlessly integrating generative design, CFD simulations, image processing,
and machine learning. Auto-wind can help architects accurately assess and incorpo-
rate the building wind environment during the schematic stage. Therefore, this study
promotes the interface between performance-based design optimization technology and
architectural design workflow, enhancing architects’ design capability and initiative for
performance-oriented building design. In addition, this study encourages the application
and popularization of this technology in design practice. Before presenting the study’s
work, the development of CFD plugins based on the rhino–grasshopper platform is briefly
reviewed to identify research gaps.

1.1. Overview of the CFD Plugins into the Rhino–Grasshopper Platform

CFD enables accurate simulations of microclimate and pollutant dispersion using
computers. Incorporating CFD simulations into the design process for evaluation and opti-
mization can significantly enhance the design quality and comfort [11–13]. However, the
high computational expense of the associated software for conducting simulations impedes
its use in preliminary stage design evaluations. With the growing need for optimization
techniques in building design, many developers have integrated CFD into their design
platforms through plugins. As light-version CFD tools, CFD plugins require significantly
fewer computer resources than CFD software [14]. In addition, they significantly simplify
many aspects of CFD modeling, including numerical schemes [15,16], turbulence model-
ing [17,18], mesh schemes, and mesh generation methods [19–21], making CFD simulations
accessible and understandable. Integrating CFD into design platforms using plugins marks
a landmark development for the interaction between computer-aided design and CFD,
allowing architects to perform CFD simulations in their familiar design environments.
Over the years, there has been a noticeable increase in literature exploring CFD integration
through plugins, indicating the growing popularity of CFD plugin applications [22].
Developed in 2007, Grasshopper is a visual programing plugin for the Rhino design
platform. Architects can maintain full design logic while improving design efficiency
through Directed Acyclic Graph (DAG) technology [23]. Avoiding ornate descriptions and
complex terminology, Grasshopper’s simple sentences and causal connections provide a
clear and concise logical flow of information. Therefore, Grasshopper has quickly become
the most commonly used platform for architectural design optimization in architectural
education and practice. RhinoCFD integrates numerous primary functions derived from
PHOENICS [24], encompassing airflow, pollutant dispersion, and heat/mass transfer sce-
nario simulations. Rhino consistently displays adequate precision and fast computation
speeds when conducting outdoor airflow simulations [25]. Flow-RT is a plugin offered
by the established CFD software Simerics-MP and incorporates state-of-the-art real-time
CFD simulations into the Rhino design platform [26]. Eddy3D streamlines mesh generation
while providing automated mesh segmentation and adaptive mesh refinement capabilities,
enabling architects and researchers to generate high-quality geometric meshes [27]. Waibel
et al. created the GH_Wind plugin in Rhino–Grasshopper using the fast fluid dynam-
ics method to enhance wind-driven design optimization efficiency. A few studies have
implemented this plugin for building shape design optimization [28,29]. Hu (2022) et al.
complemented the reliability validation of CFD plugins, contain GH_Wind, Butterfly and
Autodesk CFD [30].
Buildings 2024, 14, 2613 3 of 26

Another commonly employed performance simulation tool within the Grasshopper

platform is the Ladybug toolset, an open-source plugin with a robust eco-community and
analysis grid written in Python by Mostapha Sadeghipour Roudasri and Chris Mackey. The
Ladybug toolset’s Butterfly, a CFD plugin for Grasshopper, executes linked energy, wind,
and comfort computations for the same parametric model. However, Butterfly necessitates
highly specialized commands and complex workflows for establishing simulation tasks and
is most appropriate for designers with CFD experience [31]. However, Swift has designed
a succinct graphical user interface (GUI) that streamlines the early setup necessary for CFD
simulation workflows [32].
Current CFD plugins can preprocess most parameter settings to simplify the CFD
simulation workflow. Nevertheless, these default operations and parameters may not
satisfy the requirements of actual design projects [33]. While most plugins streamline
the CFD simulation workflow, practical experience and knowledge of basic CFD princi-
ples remain essential [34,35]. Designers must still spend considerable time waiting for
CFD simulation results during the building design process, making it challenging for
designers to incorporate performance-based design optimization techniques into the actual
design process.
With the rapid development of intelligent algorithm technology in recent years,
machine-learning technology has been increasingly used in architecture, and building
performance simulation (BPS) has gradually become one of the most suitable fields for
this new technology. Machine-learning algorithms represented by simple artificial neural
networks can learn the corresponding input–output association patterns from multivariate
high-dimensional datasets to build an input–output association model. The corresponding
data output can be predicted by feeding new data into the trained prediction model. This
property is widely used in building energy prediction to speed up the optimization of
building energy consumption and has been called the surrogate or meta-model. As a subset
of machine learning, deep learning (DL) enables computers to learn from data without
explicitly defining problem-specific rules.
Currently, researchers have employed these techniques in computational design. Guo
et al. proposed a neighborhood model using convolutional neural networks (CNNs) for
the real-time prediction of nonuniform steady laminar flow. CNN has been experimentally
verified to predict wind speed fields at two orders of magnitude faster than the GPU-
accelerated CFD-based CFD solver and four orders of magnitude faster than the CPU-
based CFD solver, with a lower error rate [36]. In 2014, Goodfellow et al. proposed
generative adversarial networks (GANs), which have received widespread attention and
are being increasingly used [37]. Since the algorithm brings artificial intelligence closer
to architectural challenges, such as mapping and image generation, GANs are becoming
increasingly popular in predicting building performance. For instance, Rahmani et al.
proposed an energy model machine in 2017, which uses GANs to automate building energy
consumption data calculations [38]. Sarah et al. used a GAN to learn numerous wind
speed maps generated by performance analysis software, leading to a speedy outdoor wind
speed analysis of building floor plans under restricted conditions [39]. Duering et al. (2020)
from the Austrian University of Technology developed InFra, an integrated simulation and
analysis system using machine learning and generative design. The system significantly
reduces the computational cost of simulating the built environment for optimal design
while offering numerous optimized design solutions to users [40]. Yousif et al. (2021)
introduced the deep performance framework, a novel tool based on the Pix2Pix algorithm,
which automates building performance evaluation for energy simulation and daylighting
analysis [41]. Zhao (2022) proposed a multi-objective optimization design method for wind
environment in severe cold regions integrating GAN [42]. Tan et al. (2023) evaluated the
results and process of the Pix2Pix model and cycleGAN, and explore the difference of
results between these two DL models and professional CFD software [43]. Huang (2023)
used XGBoost and Pix2Pix to realize the “space-environment” correlation mapping in the
design of climate-adaptive urban blocks [44].
Buildings 2024, 14, 2613 4 of 26

This technique eliminates physical computations and provides real-time feedback on

building performance. Designers can use this tool to explore design possibilities extensively
and achieve high-performance designs.
In conclusion, CFD plugins integrated into design platforms have been extensively
implemented in urban and architectural design. However, implantable plugins with limited
functionality have limitations in their application.
• First, although plugins can streamline the CFD simulation process, expertise in CFD re-
mains crucial in practice. Moreover, users face a lack of explicit guidance on workflow
and simulation operations when employing immature GUI plugins.
• Second, the computational cost of current CFD plugins is high, making it unrealistic
for architects to conduct extensive performance analysis and design modifications
during the short program design phase. Consequently, these tools fail to meet the
design practice expectations.
• Third, while machine learning in CFD prediction is a well-established technique, these
studies require a substantial number of wind simulation experiments to generate
paired datasets, and data preparation and labeling are intricate and time-consuming.

1.2. Paper Overview

This study presents a parametric working framework built on the Rhino–Grasshopper
platform to assist designers in efficiently, easily, and accurately conducting wind analysis
using CFD in a parametrically modeled environment during the initial design stage. The
framework assesses the wind environment of buildings across various geometrical parame-
ters. Focusing on each component of the working framework, building generation, BPS,
and machine learning are frequently studied or developed separately in current research.
We aim to propose a friendly decision-support framework for architects that incorporates
the preliminary design stages. This framework will integrate various components into a
cohesive whole, improving the efficiency and effectiveness of the design process.
This study has three primary contributions.
• Building auto-wind, a fully automated and user-friendly working framework on the
Rhino–Grasshopper platform, facilitates the integration of machine-learning models,
BPS software, and design platforms in a single working environment.
• This study proposes a framework for assessing the wind environment of buildings
in real time. This framework considers various geometrical parameters, significantly
enhancing architects’ efficiency in the preliminary design stages.
• In the auto-wind framework, we automate dataset production by scripting, freeing
users from the complex and time-consuming data preparation task.
Subsequently, Section 2 describes the development process for the auto-wind frame-
work, including its development strategy and automation workflow design and introduces
QueenAnt, a Grasshopper plugin that we developed. In Section 3, we present a real-world
application that provide experimental validation of the effectiveness of integrating the pro-
posed CFD prediction system into the actual architectural design process. In addition, we
explore its role in supporting architects’ design decisions in the preliminary design phase.
Section 4 analyzes the experimental results and discusses the advantages and limitations of
the proposed method. Section 5 summarizes the primary conclusions of this study.

2. Method
The DL model training workflow typically involves data collection, preprocessing, and
training. This study follows the same framework, but the distinction is that the proposed
approach was devised with the architect as the primary focus and considers the attributes of
the schematic design stage. In a standard application, the user must establish the structure
and general weather conditions to perform parametric modeling and CFD simulations.
Therefore, auto-wind can be preprogrammed with many building cases before the designer
initiates the design process, which is fully automated and requires no human input.
 Buildings 2024, 14, 2613 5 of 26

Upon completing model training, auto-wind can rapidly predict the wind conditions
for a given building. Consequently, designers can quickly acquire BPS results based on the
unique DL model for various design scenarios with distinct geometrical parameters within
the initial few hours of design, offering the necessary support for decision-making in the
preliminary design stages.

2.1. Automated Workflow Design

Figure 1 illustrates the proposed framework’s workflow. In brief, the proposed work-
flow interconnects the generation tool, performance simulation program, and DL model
through a recursive loop. This loop automatically generates building volume, evaluates
performance, constructs the dataset, and trains the DL model. In other words, auto-wind is
an automated framework for conducting complex CFD simulations and DL model training
without requiring expertise in related technologies.
Auto-wind comprises training and application. The training phase follows a specific
workflow that entails generating numerous parameterized models with various morpholo-
gies using the generator tool (Figure 1a), generating the dataset required for training, includ-
Buildings 2024, 14, x FOR PEER REVIEW 6 of 27
ing the CFD simulation work (Figure 1b-1) and the data postprocessing work (Figure 1b-2),
and training the DL model based on the dataset (Figure 1c).

Figure
Figure 1. 1. Workflow
Workflow ofof the
the proposedframework
proposed framework(Source:
(Source:Authors).
Authors).

2.1.1. Building Geometry Generation (a)

This study produces extensive datasets for training by employing the EvoMass tool
to generate building massing models. EvoMass was used due to its capability to produce
building massing designs with considerable design variation and diversity (Figure 2). The
EvoMass generation models are based on the additive and subtractive algorithms, with
this study employing the additive algorithm to create diverse design variations of archi-
tectural massing design [45]. The additive algorithm produces architectural structures by
combining multiple mass elements to create structures with highly diverse topological
Buildings 2024, 14, 2613 6 of 26

The goal in the application phase is to leverage the trained machine-learning model
for genuine building design problems (Figure 1d). The primary steps involve (1) inputting
new building design data into the trained model. The data must have the same format
and features as the input data used during the training phase. (2) The trained model will
be used to predict the input data and obtain the predicted building performance results.
(3) The predicted results will be analyzed to evaluate the design solution performance
based on various performance indicators.

2.1.1. Building Geometry Generation (a)

This study produces extensive datasets for training by employing the EvoMass tool
to generate building massing models. EvoMass was used due to its capability to produce
building massing designs with considerable design variation and diversity (Figure 2).
The EvoMass generation models are based on the additive and subtractive algorithms,
with this study employing the additive algorithm to create diverse design variations of
Buildings 2024, 14, x FOR PEER REVIEW 7 of 27
architectural massing design [45]. The additive algorithm produces architectural structures
by combining multiple mass elements to create structures with highly diverse topological
variations [46].

Figure 2. Random examples of building massing design generated by EvoMass for different typolo-
Figure 2. Random examples of building massing design generated by EvoMass for different typol-
gies (Source: Authors).
ogies (Source: Authors).
2.1.2. Datasets Creation (b)
2.1.2. Datasets Creation (b)
After generating a specific number of building massing variants, the workflow creates
After generating
the dataset. a specific
For this study, number
the dataset of building
will solely massing
comprise variants,
images. theinvolved
The steps workflow
in cre-
ates thethe
creating dataset. For
dataset this study,
involve the dataset
conducting CFD will solely comprise
simulations images.postprocessing.
and performing The steps involved
in creating the dataset involve conducting CFD simulations and performing postpro-
(b-1) CFD Simulation
cessing.
(b-1)Before
CFD delving
Simulationinto the CFD simulation operations, it is noteworthy to mention that the
current CFD plugins on Grasshopper have typically been employed for forecasting airflow
states Before
in urban delving
districtsinto the CFD
[47–53]. simulationusing
It is simulated operations, it isCFD
an external noteworthy to mention that
solver (OpenFOAM)
the current CFD plugins on Grasshopper have typically been employed
to guarantee computational exactness and effectiveness. Moreover, due to its integration for forecasting
airflow states in urban districts [47–53]. It is simulated using
into Rhino and Grasshopper, Butterfly allows concurrent model modification and an external CFDdatasolver
(OpenFOAM)
analysis, to guarantee
effectively computational
resolving issues exactness
that arise from andinteraction.
software effectiveness. Moreover,the
Nevertheless, due to
its integration
platform into Rhino
has limitations. Inand Grasshopper,
a standard workflow,Butterfly allows
architects usingconcurrent
Butterfly model modifica-
for outdoor
wind environment
tion and simulations
data analysis, mustresolving
effectively set severalissues
parameters to manage
that arise the simulation
from software interaction.
and
Nevertheless, the platform has limitations. In a standard workflow, architectsmodel,
results [49]. The setting steps include (1) creating or importing the geometric using But-
(2) defining
terfly the boundary
for outdoor conditions, (3)
wind environment selecting the
simulations appropriate
must mesh
set several type, accuracy,
parameters to manage
the simulation and results [49]. The setting steps include (1) creating or importing the ge-
ometric model, (2) defining the boundary conditions, (3) selecting the appropriate mesh
type, accuracy, and mesh quality requirements, (4) setting the time step, number of solver
iterations, convergence criteria, and stopping conditions, and (5) visualizing the analysis
outcomes.
During this stage, CFD plugins on design platforms require highly specialized com-
 Buildings 2024, 14, 2613 7 of 26

and mesh quality requirements, (4) setting the time step, number of solver iterations,
convergence criteria, and stopping conditions, and (5) visualizing the analysis outcomes.
During this stage, CFD plugins on design platforms require highly specialized com-
mands and complex workflows for CFD simulation of architectural solutions. Only one
solution can be simulated at a time, and the simulation needs multiple human–computer
interactions and parameter selections in Grasshopper. Architects must invest significant
effort in acquiring CFD professional knowledge to use it proficiently. During the initial de-
sign stages, architects frequently need to evaluate numerous scenarios. However, manually
conducting various simulations on parametric models can be tedious and time-consuming.
Unfortunately, Butterfly has yet to develop a sufficient GUI, which limits the effectiveness
of communicating information to users.
dings 2024, 14, x FOR PEER REVIEW Therefore, the workflow proposed in this study enhance the CFD simulation calcula- 8 of
tion process. First, a new preprocessing function was implemented. Second, we developed
a concise GUI to streamline the initial simulation workflow setup, including a set of sim-
ulation templates requiring few input parameters. Finally, an automation function for
human simulation
selectionis added by writing a batch
or intervention, program to unify
automatically different
initiating theparts of the simulation
following simulation on
and calculation. The program independently simulates all imported models without hu-
the first is completed, thus achieving full automation of simulation and calculation (Figu
man selection or intervention, automatically initiating the following simulation once the
3). first is completed, thus achieving full automation of simulation and calculation (Figure 3).

FigureFigure 3. Structure
3. Structure of the
of the proposed automation
proposed automation of CFD simulation
of CFD workflow
simulation (Source: (Source:
workflow Authors). Authors).

(b-2) Postprocessing
(b-2) Postprocessing
The postprocessing stage begins after completing each CFD simulation. The data
are normalized
The postprocessing in this phase, and the after
stage begins dataset necessary for
completing the DL
each CFD model is created.The
simulation. A data a
built-in postprocessor is provided to visualize and display the simulation results in a Rhino
normalized in this phase, and the dataset necessary for the DL model is created. A bu
window next to Grasshopper, with real-time updates as the simulation runs.
in postprocessor
A numerical is provided toofvisualize
interpretation and display
the CFD simulation theissimulation
findings results
included. A pixel is thein a Rhi
window next
tiniest unitto
in Grasshopper, with real-time
an image. We quantitatively analyzeupdates as theresults
the simulation simulation runs. a
by establishing
consistent legend that maps wind speed to the RGB values of each pixel.
A numerical interpretation of the CFD simulation findings is included. A pixel is t
tiniest unitSince the image.
in an Pix2Pix model in this study is aanalyze
We quantitatively DL technique for translating
the simulation images
results byfrom
establishi
one form to another (described in more detail in the following section), building a dataset
a consistent legend that maps wind speed to the RGB values of each pixel.
necessitates a set of paired pictures comprising input and target images. Therefore, the
Since the Pix2Pix
postprocessing phasemodel in this
comprises two study
primaryis steps.
a DL First,
technique for translating
the original 3D model used images fro
one form
in thistosimulation
another (described in more
is reloaded and encodeddetail in the
(Figure following
4 shows section),
the encoding building
process under a data
different simulation tasks). This step converts the input into a
necessitates a set of paired pictures comprising input and target images. Therefore, tdata format compatible
with the image-to-image translation algorithm. Then, the simulation results are loaded
postprocessing phase comprises two primary steps. First, the original 3D model used
synchronously for visualization. Once these two steps have finished, the system merges
this simulation
the input and is target
reloaded andThe
images. encoded (Figure
Pix2Pix model 4 shows
typically thethe
places encoding process
input image on theunder d
ferentright
simulation tasks). This step converts the input into a data format
and the target image on the left in its parameter settings. The merged image has twice compatible w
the width of a single
the image-to-image image andalgorithm.
translation the same heightThen, (Figure
the 5). Once the combined
simulation images
results are loaded sy
have been generated, the system will capture them at a resolution
chronously for visualization. Once these two steps have finished, the system merges t specified by the user
from the current view. The images will then be saved in a format suitable for training the
input and target images. The Pix2Pix model typically places the input image on the rig
and the target image on the left in its parameter settings. The merged image has twice t
width of a single image and the same height (Figure 5). Once the combined images ha
been generated, the system will capture them at a resolution specified by the user fro
the current view. The images will then be saved in a format suitable for training the mod
 Buildings 2024, 14, 2613 8 of 26

model, frequently JPEG or PNG, before being automatically stored in a separate external
folder. The gathered data will be split into training and test sets upon completion of the
Buildings 2024,
data collection process. Generally, approximately 80% of the data will be employed for9 of 27
Buildings 2024, 14,14, x FOR
x FOR PEERREVIEW
PEER REVIEW 9 of 27
training, while the remaining 20% will be used for testing. It is crucial to ensure that there
is no overlap between the training and test sets.

Figure4.
Figure Samples of
4. Samples of geometry
geometryencoding
encoding(Source: Authors).
(Source: Authors).
Figure 4. Samples of geometry encoding (Source: Authors).

Figure
Figure5.
5. Samples
Samples ofof processing
processingvisualization
visualization paired
paired images
images (Source:
(Source: Authors).
Authors).
Figure 5. Samples of processing visualization paired images (Source: Authors).
Although Pix2Pix
Although Pix2Pixperforms
performswellwellonon many
manyimage
imagetranslation tasks,
translation it hasitlimitations,
tasks, has limitations,
suchAlthough
as strict Pix2Pix
data performs
pairing well
requirements, on many
high image
costs translation
associated with tasks,
manual
such as strict data pairing requirements, high costs associated with manual annotation, a
it has limitations,
annotation,
a large
such as sample
strict size,
data and a requirements,
pairing requirement forhigh highly consistent
costs data.with
associated Therefore,
manual this study
annotation,
large sample
proposes an
size, and
approach
a
that
requirement
connects CFD
for highly
simulation
consistent
and
data.
postprocessing
Therefore,
operations
this study pro-a
through
large
posessample size, and
an approach thata requirement
connects CFD for highly consistent
simulation data. Therefore,
andgenerates
postprocessing this study pro-
a recursive loop (Figure 6). In this loop, the system buildingoperations
volumes and through
poses an
aperforms approach
recursiveevaluations that
loop (Figure connects CFD simulation
6). In this saving
loop, the system and postprocessing
generates operations through
automatically, them in a paired image building
format. Involumes and per-
other words,
aforms
recursive loop
evaluations (Figure 6). In this loop, the system generates building volumes
In otherand per-
this method enablesautomatically, saving them
the creation of datasets with ain a paired
single click, image format.
improving the efficiency words,
of
forms
this evaluations
method
dataset automatically,
enables
creation the creation
in Pix2Pix saving
of datasets
and freeing them in a paired
with manually
the user from image
a single click, format.
improving
creating In other words,
datasets. the efficiency
this methodcreation
of dataset enablesinthe creation
Pix2Pix andoffreeing
datasets
thewith
user afrom
single click, improving
manually the efficiency
creating datasets.
of dataset creation in Pix2Pix and freeing the user from manually creating datasets.
uildings 2024, 14, x FOR PEER REVIEW
Buildings 2024, 14, 2613 9 of 26

Figure 6. Workflow of automated production of datasets (Source: Authors).

Figure 6. Workflow of automated production of datasets (Source: Authors).
2.1.3. The DL Model Training (c)
The following step in the workflow is training the deep neural network based on the
2.1.3. TheTherefore,
dataset. DL Model Training
we use (c) image translation model based on the conditional
the Pix2Pix
GAN (cGAN).
The Figure 7step
following showsinthe
theworkflow
workflow of this
is model for the
training thetraining data. It matches
deep neural network bas
the input image with the target image to acquire knowledge of the mapping relationship
dataset. Therefore, we use the Pix2Pix image translation model based on the co
between the input and output. The model exhibits remarkable performance in numerous
GAN
image(cGAN).
translationFigure 7 shows
tasks. We the workflow
implemented of this
the Python-based model
network for the training
architecture of Pix2Pix.data. It
the
Theinput image
Pix2Pix with
model’s the target
generator image
uses the UNETto acquire knowledge
architecture, of the mapping rel
which is an encoder–decoder
containing skip connections to conserve detailed information from the input image. The
between the input and output. The model exhibits remarkable performance in n
model’s discriminator follows the PatchGAN architecture: a deep CNN is intended to
image
classifytranslation
the NxN patches tasks. We implemented
(generator the image
outputs) of the input Python-based network
as true or false and then archit
Pix2Pix.
average theTheoutcomes.
Pix2Pix model’s generator uses the UNET architecture, which is an
In this study,
decoder containing we chose
skip TensorFlow,
connections which is currentlydetailed
to conserve one of theinformation
most widely usedfrom the
frameworks, as our primary tool for implementing DL algorithms. TensorFlow can train
age. The model’s discriminator follows the PatchGAN architecture: a deep CN
and execute complex neural networks, and its architecture is highly adaptable, making it
tended to classify
an excellent the
choice for NxN patches
supporting various (generator
network models. outputs) of the
In addition input image
to TensorFlow, we as tru
and then average
leveraged the outcomes.
libraries such as NumPy, OpenCV, and Matplotlib to provide algorithms with
additional functionalities.
In this study, we chose TensorFlow, which is currently one of the most wid
frameworks, as our primary tool for implementing DL algorithms. TensorFlow
and execute complex neural networks, and its architecture is highly adaptable, m
an excellent choice for supporting various network models. In addition to Tensor
leveraged libraries such as NumPy, OpenCV, and Matplotlib to provide algorith
additional functionalities.
Buildings 2024, 14, x FOR PEER REVIEW 11 of 27
Buildings 2024, 14, 2613 10 of 26

Figure 7. The framework of procedural training (Source: Authors).

2.1.4. CFD Results Prediction (d)

After the Pix2Pix model completes training, the system saves it in an external folder
The saved model is then loaded in Grasshopper. Subsequently, after the user inputs th
design, the system automatically captures and generates image data that can be used a
input for the Pix2Pix model. Ultimately, the trained Pix2Pix model can promptly predic
the wind environment for a specific building scheme.

2.2.
7. QueenAnt
Figure
Figure 7. The
The Toolkit
framework
framework ofofprocedural
procedural training
training(Source: Authors).
(Source: Authors).
The Results
2.1.4. CFD earlier Prediction
processes(d) have been automated and programed using C# and Python
2.1.4.Visual
CFD Results Prediction (d)
After the Pix2Pix model completes usage
Studio was used to simplify training, andthedevelop
system savesthe program as standalone
it in an external folder. Rhin
v7
The and
After
saved Grasshopper
the Pix2Pix
model plugins
model
is then loaded (GHA
completes format
training,
in Grasshopper. files)thethat seamlessly
system
Subsequently, savesthe
after integrate
it user withthe
in aninputs
externalthefolder
desig
The saved model is then loaded in Grasshopper. Subsequently, after the user inputscom
workflow
design, the (Figure
system 8). The
automatically software
capturescreated
and in this
generates research
image is
data named
that can QueenAnt.
be used as It the
input
prises
design, for(1)
the theaPix2Pix
system generationmodel.
automatically Ultimately,
component capturesthe trained
for building Pix2Pix
and generates blocks model
using
image can promptly
EvoMass’s
data predict
that can additive
be used algo
as
the wind(2)
rithm, environment
a component for atospecific
automate building
CFD scheme.
simulations, (3) an image preprocessing compo
input for the Pix2Pix model. Ultimately, the trained Pix2Pix model can promptly predict
nent, environment
the wind (4) a component
2.2. QueenAnt
to capture and export images, and (5) a Pix2Pix modeling compo
Toolkit for a specific building scheme.
nent.
The earlier processes have been automated and programed using C# and Python.
2.2. Visual
QueenAntThe
Studio
workflow
Toolkit is completed by linking these components with some native compo
was used to simplify usage and develop the program as standalone Rhino
nents
v7 on the Rhino–Grasshopper
andearlier
Grasshopper plugins (GHA platform
format (Figure
files) that 9). In integrate
seamlessly practice,with once the fundamenta
The processes have been automated and programed usingthe C#design
and Python
design conditions, such as building model parameters
workflow (Figure 8). The software created in this research is named QueenAnt. It comprises and site, have been established i
Visual Studio
Grasshopper, was used
the user canto simplify usage and develop the program as standalone Rhino
(1) a generation component forset the appropriate
building blocks using simulation
EvoMass’sparameters within the
additive algorithm, (2) aprovided
v7 and Grasshopper
interactive
component tointerface
automateplugins
ofCFDthe(GHA
componentsformat(3)
simulations, files)
(Figure that
an image10).seamlessly
Following this,
preprocessing integrate with(4)
the work
component, the
a design
frame can b
workflow
component
started. (Figure
to 8).
capture Theand software
export created
images, and in
(5) this
a research
Pix2Pix is
modeling named
component.QueenAnt. It com-
prises (1) a generation component for building blocks using EvoMass’s additive algo-
rithm, (2) a component to automate CFD simulations, (3) an image preprocessing compo-
nent, (4) a component to capture and export images, and (5) a Pix2Pix modeling compo-
nent.
The workflow
Figure
Figure 8.8.Component
Component isimplemented
completed
implemented inby inlinking these (Source:
Rhino-Grasshopper
Rhino-Grasshopper components
(Source:
Authors). with some native compo-
Authors).
nents on the Rhino–Grasshopper platform (Figure 9). In practice, once the fundamenta
The
design conditions,workflow
In this toolkit,
suchis the
completed
asPix2Pix
building by linking
model
model these
is programed components
parameters andwith
using site,some
Python have 3.5native compo-
and relies
been on numer
established in
nents on the Rhino–Grasshopper platform (Figure 9). In practice, once the fundamental
ical computation
Grasshopper, the userlibraries
can set (NumPy and SciPy)
the appropriate and DL parameters
simulation frameworkswithin (TensorFlow). Grass
the provided
design conditions, such as building model parameters and site, have been established in
hopper
interactive solely
interface supports
of the IronPython, a variant of Python that integrates with Microsoft’s
Grasshopper, the user can components
set the appropriate (Figure 10). Following
simulation parameters this, thethe
within work frame can be
provided
interactive interface of the components (Figure 10). Following this, the work frame can canno
NET
started. framework. Furthermore, TensorFlow, which is used to create the DL models,
bestarted.
be called through IronPython. QueenAnt simplifies creating DL models by externall
launching them the
In this toolkit, andPix2Pix
starting a Python
model 3.5 process
is programed externally,
using Python 3.5 and providing a user-friendl
relies on numeri-
interface
cal computation to call TensorFlow
libraries (NumPyfrom within
and SciPy) andGrasshopper.
DL frameworksQueenAnt (TensorFlow). andGrasshop-
TensorFlow ex
per solelydata
change supports
via a IronPython,
command-line a variant of Python
interface, that integrates
allowing with Microsoft’s.
for the deployment NET networ
of neural
framework. Furthermore,
models and real-time predictions. TensorFlow, which is used to create the DL models, cannot be
Figure 8. Component
called implemented
through IronPython. in Rhino-Grasshopper
QueenAnt simplifies creating(Source: Authors).
DL models by externally launch-
ing them and starting a Python 3.5 process externally, providing a user-friendly interface
toIncall
this toolkit, the
TensorFlow Pix2Pix
from withinmodel is programed
Grasshopper. QueenAntusing
and Python 3.5 and
TensorFlow reliesdata
exchange on numer-
via a command-line interface, allowing for the deployment of neural network
ical computation libraries (NumPy and SciPy) and DL frameworks (TensorFlow). Grass- models and
real-time predictions.
hopper solely supports IronPython, a variant of Python that integrates with Microsoft’s
NET framework. Furthermore, TensorFlow, which is used to create the DL models, cannot
be called through IronPython. QueenAnt simplifies creating DL models by externally
launching them and starting a Python 3.5 process externally, providing a user-friendly
interface to call TensorFlow from within Grasshopper. QueenAnt and TensorFlow ex-
change data via a command-line interface, allowing for the deployment of neural network
models and real-time predictions.
Buildings 2024, 14, x FOR PEER REVIEW
REVIEW 1212ofof27

Buildings 2024, 14, 2613 11 of 26

Figure 9. Building CFD prediction system established based on the toolkit (Source: Authors).
Figure9.9.Building
Building CFD prediction system established based on the(Source:
toolkit Authors).
(Source: Authors).
Figure CFD prediction system established based on the toolkit

Figure 10.The
Figure10. Thegraphics
graphicsuser interface
user (Source:
interface Authors).
(Source: Authors).
Figure 10. The graphics user interface (Source: Authors).
3. Showcase
3. Showcase
3. Showcase
We showcase the effectiveness of the proposed work framework by providing a case
studyWe
We
showcase the
demonstrating
showcase how
effectiveness of the proposed
Auto-Wind framework
the effectiveness
work
can be used
of the proposed forframework
work
by providing
performance-based
framework tasks aacase
by providing cas
study to
related demonstrating
building design.how Auto-Wind
An illustration framework
of the can be used
proposed framework for performance-based
is given in Figure 11,
study demonstrating how Auto-Wind framework can be used for performance-base
tasksstep
each related to building
of which is detaileddesign. An illustration
in the following sections.of the proposed framework is given in
tasks related to building design. An illustration of the proposed framework is given i
It should
Figure 11, eachbe step
notedofthat in the
which is case study,inwe
detailed thefocus on quantitative
following sections.analysis through
Figure
two 11, each
metrics: step ofefficiency
simulation which isand detailed in the following sections.
accuracy.
It should be noted that in the case study, we focus on quantitative analysis through
It should be noted that in the case study, we focus on quantitative analysis throug
two metrics: simulation efficiency and accuracy.
two metrics: simulation efficiency and accuracy.
Buildings 2024,14,
Buildings2024, 14,2613
x FOR PEER REVIEW 1312ofof27
26

Figure 11. Proposed

Figure 11. Proposed framework to utilize
framework to utilize Queen_Ant for outdoor
Queen_Ant for outdoor wind
wind environment
environment prediction
prediction
(Source: Authors).
(Source: Authors).

To
To assess
assess the
the efficiency
efficiency of of the
the CFD
CFD tool,
tool, the
the time
time required
required for
for each calculation was
each calculation was
recorded and compared. It is well established that discrepancies will always
recorded and compared. It is well established that discrepancies will always exist in the exist in the
results of CFD numerical simulations and even physical wind tunnel tests.
results of CFD numerical simulations and even physical wind tunnel tests. Therefore, the Therefore, the
objective of this study is not to achieve absolute accuracy. The objective of this
objective of this study is not to achieve absolute accuracy. The objective of this study is to study is
to determine
determine thethe degreeofoferror
degree errorbetween
betweenCFD CFDnumerical
numericalsimulation
simulationtools
tools and
and machine
machine
learning predictions and to analyze the extent to which computational
learning predictions and to analyze the extent to which computational efficiency can efficiency can bebe
improved when machine learning predictions reach an acceptable level
improved when machine learning predictions reach an acceptable level of accuracy. In of accuracy. In this
section, we compare
this section, we comparethe computational
the computational speedspeed
of Butterfly, Eddy3d,
of Butterfly, and GH_Wind
Eddy3d, and GH_Wind tools,
which are commonly used CFD plugin on the Grasshopper platform,
tools, which are commonly used CFD plugin on the Grasshopper platform, with the
with the prediction
speed of the predicted model to observe the simulation efficiency of different tools. We also
prediction speed of the predicted model to observe the simulation efficiency of different
verify the accuracy of the machine learning-based CFD prediction method proposed for
tools. We also verify the accuracy of the machine learning-based CFD prediction method
this study to ascertain the extent to which the method can replicate the established setup
proposed for this study to ascertain the extent to which the method can replicate the
procedures and results. This was accomplished by comparing the discrepancy between the
established setup procedures and results. This was accomplished by comparing the
results predicted by the machine learning model and those calculated by the CFD numerical
discrepancy between the results predicted by the machine learning model and those
simulation tool. As the CFD simulation module in the QueenAnt tool developed in this
calculated by the CFD numerical simulation tool. As the CFD simulation module in the
study performs simulation calculations by calling the OpenFOAM platform, as does the
QueenAnt tool developed in this study performs simulation calculations by calling the
Butterfly tool, and both utilize the visualization tools in the Ladybug toolset for mapping
OpenFOAM platform, as does the Butterfly tool, and both utilize the visualization tools in
from simulation data to images. Consequently, in order to validate the accuracy of the
the Ladybug toolset for mapping from simulation data to images. Consequently, in order
results, the wind speed maps generated by Butterfly were compared with the predictions
to validate the accuracy of the results, the wind speed maps generated by Butterfly were
made by QueenAnt.
compared with the predictions made by QueenAnt.
The case study selected the Xinjiekou area in Nanjing, Jiangsu Province, China. Nan-
The case study selected the Xinjiekou area in Nanjing, Jiangsu Province, China.
jing is located in the subtropical monsoon climate zone, between 31◦ 14′ and 32◦ 37′ north
Nanjing is located in the subtropical monsoon climate zone, between 31°14′ and 32°37′
latitude and 118◦ 22′ and 119◦ 14′ east longitude. The predominant wind direction in the
 contrast, the dominant wind direction in the winter is northeast, with an average wind
speed of 2.8 m per second. The average annual temperature in Nanjing is 15.4 degrees
Celsius, with a hot and humid summer season. The average temperature and relative
humidity
Buildings 2024, 14, x FOR PEER REVIEW in the hottest month are 28.56 °C and 80.8%, respectively. This makes14natural of 27
Buildings 2024, 14, 2613 ventilation in summer especially important in buildings. The EPW weather file and wind
13 of 26
environment parameters of Nanjing are shown in Figure 12.
Figure 13 depicts
north latitude the building
and 118°22′ siteseast
and 119°14′ andlongitude.
surrounding urban environments
The predominant wind direction for inthe
summer
case. In months is southeast, with an average wind speed of 3.5 m per second.
the summer months is southeast, with an average wind speed of 3.5 m per second. In a
this case, several high- and middle-rise buildings are surrounding, In contrast,
creating
the dominant
challenging
contrast,and wind direction
complex
the dominant wind in the
design winter
condition
direction in theisfor
northeast, with an average
performance-based
winter is northeast, with an wind
building
average speed of
design.
wind
2.8 m per second. The average annual temperature in Nanjing is 15.4 degrees Celsius,
Relying solely
speed on simulation
of 2.8 m per second. software to determine
The average the impact of
annual temperature in a designissolution
Nanjing on the
15.4 degrees
with a hot and humid summer season. The average temperature and relative humidity in
Celsius, with a hot and humid summer season. The average temperature
outdoor wind environment can◦ be extremely labor-intensive and time-consuming. This and relative
the hottest month are 28.56 C and 80.8%, respectively. This makes natural ventilation in
humidity
underscores inneed
the hottest month are 28.56 °C and 80.8%, respectively. This makes
studynatural
summerthe especially forimportant
applyinginthe working
buildings. Theframework proposed
EPW weather file andin this
wind to such
environment
ventilation
designparameters
tasks. in summer especially important in buildings. The EPW weather file and wind
of Nanjing are shown in Figure 12.
environment parameters of Nanjing are shown in Figure 12.
Figure 13 depicts the building sites and surrounding urban environments for the
case. In this case, several high- and middle-rise buildings are surrounding, creating a
challenging and complex design condition for performance-based building design.
Relying solely on simulation software to determine the impact of a design solution on the
outdoor wind environment can be extremely labor-intensive and time-consuming. This
underscores the need for applying the working framework proposed in this study to such
design tasks.

Figure 12. The wind rose of Nanjing [left], the local yearly wind speed [top right] and the local yearly
Figuretemperatures
12. The wind [down
roseright] (Source:[left],
of Nanjing Authors).
the local yearly wind speed [top right] and the local
yearly temperatures [down right] (Source: Authors).
Figure 13 depicts the building sites and surrounding urban environments for the case.
In this case, several high- and middle-rise buildings are surrounding, creating a challenging
and complex design condition for performance-based building design. Relying solely on
simulation software to determine the impact of a design solution on the outdoor wind
environment
Figure 12. Thecan berose
wind extremely labor-intensive
of Nanjing and
[left], the local time-consuming.
yearly This
wind speed [top underscores
right] the
and the local
need for
yearly applying the
temperatures working
[down framework
right] (Source: proposed in this study to such design tasks.
Authors).

Figure 13. The location of the case study within Xinjiekou (Source: Authors).

3.1. Parameter Setups of Geometry Massing (Step 1)

Here, we specify the design object as a high-rise office building with limited mass
Figure
elements. 13. The
FigureWe
13. location
location of
distinguish
The of the case
the
the study
caseformal within Xinjiekouof
attributes
study within Xinjiekou (Source:
eachAuthors).
(Source: design by manipulating
Authors).

3.1. Parameter Setups of Geometry Massing (Step 1)

Here, we specify the design object as a high-rise office building with limited mass
elements. We distinguish the formal attributes of each design by manipulating
 Buildings
Buildings 14, x 14,
2024, 2024, FOR 2613
PEER REVIEW 15 of 27
14 of 26

3.1. Parameter Setups of Geometry Massing (Step 1)

parameters, such as the quantity of mass elements and their dimensional restrictions. This
Here, we specify the design object as a high-rise office building with limited mass
case study specifically
elements. We distinguishaims
theto predict
formal the wind
attributes patterns
of each designfor regular orthogonal
by manipulating building
parameters,
units quickly.
such as the quantity of mass elements and their dimensional restrictions. This case study
Based on
specifically theto EvoMass
aims tool, patterns
predict the wind we created four orthogonal
for regular groups comprising
building units1200 building
quickly.
massing models.
Based on theWe distinguished
EvoMass the formal
tool, we created features
four groups for each1200
comprising design set massing
building according to
models. We
parameter distinguished
settings such asthetheformal
number features for each
of cells and design set according
dimensions. to parameter
First, building massing
settings such as the number of cells and dimensions. First, building
models were created within a defined spatial boundary of grid cells in the massing models were (X),
east–west
created within
north–south (Y),aand
defined spatial
vertical (Z)boundary
directions.of The
grid number
cells in the
ofeast–west
summed(X), north–south
volumes was varied
(Y), and vertical (Z) directions. The number of summed volumes was varied to control the
to control the number of vertical cells, ranging from 2 to 4, and the number of cells in the
number of vertical cells, ranging from 2 to 4, and the number of cells in the Z direction was
Z direction was at
held constant held
25. constant
The height at of
25.the
The height
floors wasof4.5
the
m,floors was
and the 4.5 m,ofand
spacing thethe spacing of
columns
thewas
columns
maintained at 8 m. Figure 14 displays the specific parameter settings and resultant and
was maintained at 8 m. Figure 14 displays the specific parameter settings
resultant
geometrygeometry
massing massing
models. models.

Figure
Figure 14.14.
TheThe parametersetups
parameter setups of
of geometry
geometrymassing
massingmodels
models(Source: Authors).
(Source: Authors).
3.2. Scripts of CFD Simulation Tools
3.2. Scripts of CFD Simulation Tools
A CFD simulation is typically initiated with the input of the geometry to be analyzed.
A CFD simulation
Subsequently, a domain,is typically
analogousinitiated withwind
to a virtual the input
tunnel,ofisthe geometryaround
constructed to be analyzed.
the
Subsequently,
aforementioned geometry. The domain is then meshed, resulting in a three-dimensional the
a domain, analogous to a virtual wind tunnel, is constructed around
representation geometry.
aforementioned of the space.TheThis representation
domain is utilized
is then meshed, by the solver
resulting to iteratively
in a three-dimensional
calculate flow dynamics until convergence is achieved. Critical factors, such
representation of the space. This representation is utilized by the solver to iteratively as tunnel cal-
culate flow dynamics until convergence is achieved. Critical factors, such as tunnel to
dimensions, mesh resolution, and turbulence models, are meticulously set according dimen-
established CFD practices. These settings play a pivotal role in determining the accuracy of
sions, mesh resolution, and turbulence models, are meticulously set according to estab-
the simulation’s outcome.
lished CFD practices. These settings play a pivotal role in determining the accuracy of the
simulation’s outcome.
3.2.1. Butterfly
Figure 15 illustrates the layout of the Butterfly script. The process commences with the
3.2.1. Butterfly
generation of the initial Grasshopper geometry, which is then employed to construct the
Butterfly
Figuregeometry. Subsequently,
15 illustrates the layout theofaforementioned geometries
the Butterfly script. The are linkedcommences
process to the wind with
tunnel component, thereby defining the simulation domain as a
the generation of the initial Grasshopper geometry, which is then employedvirtual wind tunnel. The
to construct
dimensions of the tunnel are established in relation to the height of the building, taking
the Butterfly geometry. Subsequently, the aforementioned geometries are linked to the
into account windward, leeward, lateral, and upper extensions.
wind tunnel component,
Subsequently, thereby
the data definingtothe
is transmitted thesimulation domain as An
meshing components. a virtual
initial wind tunnel.
cell size
Theis dimensions of the
established, and tunnel
these cells are
are established
automaticallyingraded,
relationexhibiting
to the height of the
a gradual building,
increase in tak-
ingsize
into account windward, leeward, lateral, and upper extensions.
as they move away from the building. Subsequently, the cells within the geometry are
Subsequently,
eliminated and thethe data is
requisite transmitted
refinements areto the meshing components. An initial cell size
implemented.
is established, and these cells are automatically graded, exhibiting a gradual increase in
size as they move away from the building. Subsequently, the cells within the geometry
are eliminated and the requisite refinements are implemented.
For the analysis of the wind environment of buildings, the parameter setting in But-
terfly has a relatively fixed process, and the manual setting of Butterfly parameters can be
roughly divided into “defining the solver”, “setting the materials”,” boundary condi-
 Subsequent to the application of the solution, the values derived from the prob
utilized to color the mesh in accordance with the pressure or velocity vectors. In th
analysis, the results section presents the colored mesh with pressure representation
Buildings 2024, 14, 2613
arrows depicting wind flow velocity vectors. 15 of 26

Buildings 2024, 14, x FOR PEER REVIEW 16 of 27

Subsequent to the application of the solution, the values derived from the probes are
utilized to color the mesh in accordance with the pressure or velocity vectors. In the final
analysis, the results section presents the colored mesh with pressure representations and
arrows depicting wind flow velocity vectors.

Figure Figure
15. Butterfly script
15. Butterfly (Source:
script Authors).
(Source: Authors).

For the analysis of the wind environment of buildings, the parameter setting in
3.2.2. GH_Wind
Butterfly has a relatively fixed process, and the manual setting of Butterfly parameters
can be roughly divided into “defining the solver”, “setting the materials”, “boundary
Figure 16 illustrates the layout of the GH_Wind script. GH_Wind enables u
conditions”, and “computational solving”.
define several parameters,
Subsequent including
to the application the
of the wind the
solution, speed,
valuesstep size,
derived and
from thedimension
probes are of th
ulationutilized
space.toAdditionally,
color the mesh in GH_Wind
accordance withincorporates a velocity
the pressure or componentvectors.that automatical
In the final
analysis,
cretizes input the results
building section presents the colored mesh with pressure representations and FFD is
Figure 15. Butterfly scriptvolumes and transforms them into voxels. Although
(Source: Authors).
arrows depicting wind flow velocity vectors.
accurate as CFD, it can still generally differentiate between high-performing desig
3.2.2. GH_Wind
3.2.2.
ants and GH_Wind
those that are less optimal.
Figure
Figure1616 illustrates
illustratesthe
the layout
layout ofof the
the GH_Wind
GH_Wind script.
script. GH_Wind
GH_Windenables
enablesusers
userstoto
define
define several parameters, including the wind speed, step size, and dimension ofsim-
several parameters, including the wind speed, step size, and dimension of the the
ulation space.
simulation Additionally,
space. GH_Wind
Additionally, GH_Windincorporates a component
incorporates that automatically
a component dis-
that automatically
cretizes
discretizes input building volumes and transforms them into voxels. Although FFDasis
input building volumes and transforms them into voxels. Although FFD is not
accurate as CFD,asitCFD,
not as accurate can still generally
it can differentiate
still generally between
differentiate high-performing
between designdesign
high-performing vari-
ants and those that are less optimal.
variants and those that are less optimal.

Figure 16. GH_Wind script (Source: Authors).

Figure16.
Figure GH_Windscript
16.GH_Wind script(Source:
(Source:Authors).
Authors).
3.2.3. Eddy3D
3.2.3.Eddy3D
3.2.3. Eddy3D
Figure 17 shows
Figure 17shows
showsthe
thelayout
layoutof ofEddy3D
of Eddy3D script.
script. The Eddy3D
TheEddy3D
Eddy3D forplugin
pluginfor for Grasshop
Grasshopper
Figure 17 the layout Eddy3D script. The plugin Grasshopper isis
an airflow and
an airflow microclimate
and microclimate simulation
simulation tooltool
that that employs
employs the
the OpenFOAM OpenFOAM
an airflow and microclimate simulation tool that employs the OpenFOAM platform and
platform andplatfor
the Finite
the Finite VolumeVolume Method (FVM) toto
Method predict annual wind comfort. The EPW weather files
the Finite Volume Method (FVM) predict
(FVM) to predict annualannual wind comfort.
wind comfort. The EPW
The EPW weather files weath
are utilized to generate the analysis, which is accelerated by dividing the EPW winds into
are utilized to generate
are utilized to generatethe
larger sectors analysis,
the analysis,
and interpolating which
which
local wind is accelerated
based on windby
is accelerated
speeds dividing
by dividing the the into
EPW winds
attenuation factors. EPW
The wind
larger sectors and interpolating local wind speeds based on wind attenuation factors. The
larger sectors and interpolating local wind speeds based on wind attenuation factor
results are highly customizable in Grasshopper and Rhino, allowing for precise control
resultsover
arethehighly customizable in Grasshopper and Rhino, allowing for precise c
visualization.
over the visualization.
 Buildings 2024, 14, 2613 16 of 26

uildings 2024, 14, x FOR PEER REVIEW 17 of

results are highly customizable in Grasshopper and Rhino, allowing for precise control
over the visualization.
Buildings 2024, 14, x FOR PEER REVIEW 17 of 27

Figure 17. Eddy3D script (Source: Authors).

Figure 17. Eddy3D
Figure 17. Eddy3Dscript
script(Source: Authors).
(Source: Authors).

3.3. Parameter
3.3. Parameter Setups
Setups of
of CFD
CFD Simulation
Simulation (Step
(Step 2)
2)
3.3. Parameter Setups of CFD Simulation (Step 2)
3.3.1. Tunnel Setup
3.3.1. Tunnel Setup
3.3.1. Tunnel
SinceSetup
Since finite computer
computer resources
resources constrain
constrainthe
theextent
extentofofcomputational
computationalcapabilities,
capabilities,
it
itis isnot
Since not feasible
feasible
finite toto infinitely
infinitely
computer expand
expand
resources thethe range.
range.
constrain In In related
related
the studies,
studies,
extent researchers
of researchers
computational typically
typically es-
capabilities,
establish the calculation range by considering the size of the simulated building and their
is nottablish
feasible the calculation
to infinitely range by considering
expand the range. the Insizerelated
of the studies, researchers their typically e
experience.
experience.For Forthis
thisexperiment,
experiment,we weset
setthe
thecalculation
calculationrangerangebased
basedononthe building’s
the building’soverall
over-
tablish the calculation
dimensions:
all dimensions: 5 H5for
H for
range
the by height
vertical
the
considering
vertical (H (H
height
the sizethe
representing
representing
ofthe
the simulated
highest building
highest
building
height),
building 5 H5and the
height),
experience.
for
H for thethe For
lateral this
lateral experiment,
andand
inlet boundaries
inlet we
boundaries set
andandthe
15 H calculation
15for
H forthethe
outlet range
boundary
outlet based
boundary on
(Figure the building’s
18).18).
(Figure ove
all dimensions: 5 H for the vertical height (H representing the highest building height),
H for the lateral and inlet boundaries and 15 H for the outlet boundary (Figure 18).

Figure 18.
Figure 18. Selected
Selected domain’s
domain’s dimensions
dimensions criteria
criteria (Source:
(Source: Authors).
Authors).

3.3.2. Meshing
In Butterfly, the smallest grid length was 0.0054 H = 0.987552 m at the building with
grading in the X and Y direction based on the meshing method described in the related
study by Meng et al. [54]. The grid size is set to 150, 150, and 120 in the x, y, and z axes,
Figure 18. Selected domain’s dimensions criteria (Source: Authors).
respectively. These values are reasonably set in combination with the overall dimensions
of the building. For GH_Wind, a grid size of 5 m was used as Waibel et al. [55] in a similar
3.3.2.validation
Meshingof a rectangular building. The wind vector parameters, including wind speed
and direction, are
In Butterfly, theset according
smallest to the
grid Epw file
length was for0.0054
the project
H =location.
0.987552In m
Eddy3d,
at thethe cell
building wi
 Buildings 2024, 14, 2613 17 of 26

3.3.2. Meshing
In Butterfly, the smallest grid length was 0.0054 H = 0.987552 m at the building with
Buildings 2024, 14, x FOR PEER REVIEW 18 of 27
grading in the X and Y direction based on the meshing method described in the related
study by Meng et al. [54]. The grid size is set to 150, 150, and 120 in the x, y, and z axes,
respectively. These values are reasonably set in combination with the overall dimensions of
size the building.
of the entireFor GH_Wind,
virtual windatunnel
grid sizeisofspecified
5 m was used
as 20asm.
Waibel et al.
As the grid[55]isinapproached,
a similar the
validation of a rectangular building. The wind vector parameters, including wind speed
cell size is refined to 3 levels, with each cell having a side length of 2.5 m.
and direction, are set according to the Epw file for the project location. In Eddy3d, the cell
size of the entire virtual wind tunnel is specified as 20 m. As the grid is approached, the
3.3.3.cell
Turbulence Models
size is refined andwith
to 3 levels, Solution
each cell having a side length of 2.5 m.
Although the LES and DES methods yield higher accuracy than the RANS method,
3.3.3. Turbulence Models and Solution
they typically require longer computational time. We balanced simulation precision and
Although the LES and DES methods yield higher accuracy than the RANS method,
computational efficiency using the more efficient RANS method across 1200 cases in our
they typically require longer computational time. We balanced simulation precision and
CFDcomputational
simulations.efficiency
Consequently,
using thethe RNG
more k–ε model
efficient RANS was
methodimplemented for calculating
across 1200 cases in our
CFD simulations. Consequently, the RNG k–ε model was implemented for calculating. was set at
In Butterfly, we set residual control convergance value at 0.0001, the step
In Butterfly,the
500. In GH_Wind, we step
set residual
was setcontrol
at 600.convergance
Eddy3D’s value at 0.0001, the
max iterations was step
setwas set In this
at 800.
section, the initial residuals for the key values in the CFD numerical equationsInare less
at 500. In GH_Wind, the step was set at 600. Eddy3D’s max iterations was set at 800.
this section, the initial residuals for the key values in the CFD numerical equations are less
than 1e-3 and are considered converged once the calculated residuals are constant.
than 1e-3 and are considered converged once the calculated residuals are constant.
At aAtreference
a referenceheight
heightofof 10
10 mm above
abovethethe ground,
ground, the the input
input reference
reference windwas
wind speed speed was
10 m/s, which
10 m/s, represents
which representstypical
typical gusts resulting
gusts resulting from
from dynamic
dynamic urbanurban
terrainterrain patterns.
patterns. This This
experiment
experiment utilized
utilizedananIntel
Intel i9-10900F
i9-10900F @@2.80 2.80 GHz
GHz CPU
CPU withwith 6 threads.
6 threads. FigureFigure
19 shows 19 shows
the CFD
the CFD analysis
analysis ofofdifferent
different tools.
tools.

Figure 19. 19.

Figure CFD analysis
CFD analysisofofdifferent tools((Left):
different tools ((Left): Butterfly,
Butterfly, (Mid):
(Mid): GH_Wind,
GH_Wind, (Right):(Right):
Eddy3D)Eddy3D)
(Source: Authors).
(Source: Authors).

3.4. Datasets (Step 3)

3.4. Datasets (Step 3)
The simulation results reveal that buildings draw high-velocity airflow from the upper
The simulation
atmosphere down to results reveal
the ground thatthey
when buildings draw high-velocity
reach a certain airflow from
height. This phenomenon par- the up-
per atmosphere down to the ground when they reach a certain height. This phenomenon
ticularly affects the corners of the building, accelerating the flow and creating a stationary
particularly affects
vortex in front thestructure.
of the cornersConsequently,
of the building, accelerating
the pedestrian the flow and
wind environment creating
around the a sta-
building deteriorates, endangering those passing by. Therefore, we determined
tionary vortex in front of the structure. Consequently, the pedestrian wind environment the wind
speed to be approximately 1.5 m high and analyzed the wind environment surrounding
around the building deteriorates, endangering those passing by. Therefore, we deter-
buildings of various heights to investigate the correlation between building height and
mined the wind speed to be approximately 1.5 m high and analyzed the wind environ-
wind environment.
ment surrounding buildings
In the case, the of various
machine learning modelheights to investigate
of Auto-Wind will use thethe correlation
obtained analyticsbetween
building height and wind environment.
from OpenFOAM for dataset construction. We produced dataset comprising 1200 pairs of
floor
In theplans andthe
case, wind speed simulation
machine images for
learning model wind speed forecasting
of Auto-Wind (Figure
will use the 20). We
obtained analytics
then partitioned the dataset into training and test sets using an 8:2 ratio.
from OpenFOAM for dataset construction. We produced dataset comprising 1200 pairs of
floor plans and wind speed simulation images for wind speed forecasting (Figure 20). We
then partitioned the dataset into training and test sets using an 8:2 ratio.
Buildings 2024,
Buildings 14,14,
2024, x FOR
2613PEER REVIEW 1819 of 27
of 26

Figure
Figure20.
20.AAsample
sampleof
ofthe
thetraining
training datasets (Source:Authors).
datasets (Source: Authors).

3.5.Parameter
3.5. ParameterSetups
SetupsofofTraining
Training (Steps
(Steps 4&5)
4&5)
InInthis
thisstudy,
study, aa learning
learning rate
rate of
of 0.0005
0.0005was
wasset,
set,Adam
Adamserved
servedasasthe optimizer
the forfor
optimizer thethe
generator and the discriminator, the momentum term parameter was 0.65, and the input
generator and the discriminator, the momentum term parameter was 0.65, and the input
and output resolutions were 256 × 256 pixels. The batch size was 1, and the generator
and output resolutions were 256 × 256 pixels. The batch size was 1, and the generator
network used the UNET 9-blocks network architecture. We reduced the model oscillations
network used the UNET 9-blocks network architecture. We reduced the model oscillations
by setting the pool size parameter to 50. The dataset underwent training for 800 epochs
byatsetting
50 epochtheintervals,
pool sizeandparameter to 50.
the trained The dataset
weights underwent
were saved training
to track for 800
the training epochs at
progress.
50The
epoch intervals, and the trained weights were saved to track the training
training was completed on a computer with a GeForce RTX 3070 graphics card and progress. The
training was completed
averaged 25 s per epoch. on a computer with a GeForce RTX 3070 graphics card and aver-
aged 25 s per
Once theepoch.
model is trained, the designer can interact with the CAD environment and
Once
create the massing
urban model ismodels
trained,forthe
thedesigner
predictor.can interact with the CAD environment and
create urban massing models for the predictor.
3.6. Results
3.6. Results
This section is divided into three parts for the purpose of discussing the efficiency of
the This
simulation,
sectionthe loss incurred
is divided into during the training
three parts for the of the Pix2Pix
purpose model, andthe
of discussing theefficiency
accuracy of
of the predictions. Firstly, with regard to the efficiency of the simulation, four scenarios
the simulation, the loss incurred during the training of the Pix2Pix model, and the accu-
were selected for analysis by Butterfly, GH_Wind, Eddy3D, and QueeAnt, respectively. The
racy of the predictions. Firstly, with regard to the efficiency of the simulation, four scenar-
specific parameters for each tool are presented in the table below. As illustrated in the
ios were selected for analysis by Butterfly, GH_Wind, Eddy3D, and QueeAnt, respec-
Figure 21, for expansive urban areas, tools such as Butterfly and Eddy3D, which employ
tively. The specific
conventional CFDparameters for each tool are
analysis methodologies, presented
typically in the table
necessitate morebelow. As illustrated
than one hour to
incomplete
the Figure 21, for expansive urban areas, tools such as Butterfly and
the calculations for each site. In contrast, the GH_Wind tool employs the FFDEddy3D, which
employ
method to resolve the NS equations, thereby markedly enhancing the efficiency of theone
conventional CFD analysis methodologies, typically necessitate more than
hour to complete
simulation. the calculations
Nevertheless, for eachtime
the simulation site. still
In contrast,
exceeds the GH_Wind
40 min. QueeAnt toolenables
employs thethe
FFD methodoftopredicted
generation resolve the NSimages
wind equations, thereby markedly
for large-scale areas in a enhancing
fraction of athe efficiency
second (or a of
the
fewsimulation. Nevertheless,
seconds, depending thesize
on the simulation time still
of the analyzed exceeds
area). This 40 min.that
shows QueeAnt enables
Queen_Ant
can
the achieve real-time
generation feedback
of predicted windof images
simulation
for results with areas
large-scale equal in
accuracy compared
a fraction to the(or
of a second
a computing
few seconds, speed of other plugins
depending (Figure
on the size 22).analyzed area). This shows that Queen_Ant
of the
can achieve real-time feedback of simulation results with equal accuracy compared to the
computing speed of other plugins (Figure 22).
ldings 2024, 14, x FOR PEER REVIEW 20 of
Buildings 2024, 14, 2613 19 of 26

Buildings 2024, 14, x FOR PEER REVIEW 20 of 27

Figure 21. Settings

Figure
Figure21. and
21.Settings
Settings time
and
and timeresults (Source:
results (Source:
results (Source: Authors).
Authors).
Authors).

Figure22.
Figure Theprediction
22.The prediction speed
speed of
ofdifferent
differentCFD
CFDplugins
plugins(Source: Authors).
(Source: Authors).
Second, Figure 23a indicates that the overall training loss stabilizes after 300 epochs,
Second, Figure 23a indicates that the overall training loss stabilizes after 300 epochs,
Figure indicating that the lossspeed
22. The prediction function has foundCFD
of different an optimal
pluginsvalue, and Authors).
(Source: the prediction accuracy
indicating
cannot be that further theimproved
loss function has foundthe
by increasing annumber
optimalofvalue,
trainingandsessions.
the prediction
In Figure accuracy
23b,
cannot be further improved by increasing the number of
it is observable that the dis_loss steadily declines to a lower value, around 50 epochs,training sessions. In Figure 23b,
Second,
itsignifying
is observableFigure 23a indicates that the overall training loss stabilizes50 epochs, sig-epoc
after 300
that that
the imagethe dis_loss
produced steadily
by thedeclines
generator tobecomes
a lower value, around
challenging to differentiate
indicating
from a that
nifying that the
genuine theimageloss function
image by produced hasby found
the discriminator. an optimal
the generator becomes value, and thetoprediction
challenging differentiate accura
cannotfrombeaThird,
further
genuine improved
image
regarding by theby
prediction increasing
accuracy, wethe
discriminator. number
created 50 new of plan
training sessions.
and elevation In Figure 2
images
thatThird,
it is observable
the trainedregarding
that modelthe prediction
dis_loss
had not seen accuracy,
steadily wepredicted
declines
before and created 50
to their new
a lower plan and
value,
outcomes. elevation
Wearound 50images
then captured epochs, s
both
that thethe predicted
trained modelimages hadand notthose
seencomputed
before
nifying that the image produced by the generator becomes challenging to differenti by
and Butterfly.
predicted A python
their script
outcomes. utilizing
We then
the PIL and
captured both cv2the image libraries
predicted was used
images and to compare
those the images
computed by via two methods.
Butterfly. A python The
script
fromfirsta genuine
method
image
is the
by the discriminator.
utilizing the PIL anddifference
cv2 imageapproach,
libraries was which usedsubtracts
to compare the the
twoimages
imagesvia to two
obtain the
methods.
Third,
difference regarding
part (relative prediction
error). Along accuracy,
with this we created
approach, we 50 new
assess plansimilarity
image and elevation
using imag
The first method is the difference approach, which subtracts the two images to obtain the
thatdifference
theStructural
the trained model error).
part Similarity
(relative had
Approachnot
Along seen
withbefore
(SSIM), which
this and predicted
compares
approach, weimages
assessbased theiron
image outcomes.
three main
similarity usingWe th
capturedmetrics: brightness,
both the contrast,
predicted and image
images structure
and which (Figure
those compares 24).
computed This process employs
by Butterfly. a more
A python
the Structural Similarity Approach (SSIM), images based on three main scr
comprehensive approach to image evaluation by examining a set of pixels, rather than
utilizing
metrics: thebrightness,
PIL and cv2 imageand
contrast, libraries
imagewas used to
structure compare
(Figure the images
24). This process via two metho
employs a
comparing them based on similarity, mimicking how humans perceive images. The SSIM
The more
first
method
comprehensive
method
generates is thea value
approach toapproach,
difference image evaluation
ranging from −1 to 1, with which by examining
subtracts
closer thea two
to 1 signifying
set ofimages
pixels, to
greater similarity
rather
obtain t
than
difference comparing
between part
two(relative them
images. These based
error). on similarity,
Along with
techniques mimicking
this approach,
were utilized how
during thewehumans perceive
assess
diverse image
training images. The
similarity
phases of usi
SSIM method generates a value ranging from −1
the Structural Similarity Approach (SSIM), which compares images based on three ma
each model to analyze their learning progress and to 1, with
determine closer
how to 1 signifying
individual parameters greater
similarity between
affectbrightness,
their learning two images.Figure
curves. These25 techniques
present were evaluated
cases utilized duringthroughthe diverse training
metrics: contrast, and image structure (Figure 24). Thisinterpolation
process employ
phases
methods of and
eachSSIM. model to analyze
Meanwhile, their26learning
Figure depict the progress
statisticsand determine
of SSIM values forhow allindividual
the test
more comprehensive
parameters affect approach
their learning to image
curves. evaluation
Figure 25 by
present
results. The SSIM values of the case model’s predictions range from 0.90 to 0.98, as shown
examining
cases a set
evaluated of pixels, rath
through
thaninterpolation
comparing
in Figure 26. The them
methods based
relative anderror onplots
SSIM. similarity,
Meanwhile,
indicate that mimicking
Figure howexist
26 depict
the differences humans
the only in perceive
statistics of SSIM
small images. T
values
regions,
SSIM formethod
all the test generates
results. The a value ranging
SSIM values from
of the case−1 to 1, with
model’s closerrange
predictions to 1 signifying
from 0.90 togrea
0.98, as shown in Figure 26. The relative error
similarity between two images. These techniques were utilized during the diverseplots indicate that the differences exist onlytraini
phasesin small
of each regions,model ensuring
to analyze high prediction
their learning accuracy overall.and
progress Specifically,
determine thehow accurate
individu
parameters affect their learning curves. Figure 25 present cases evaluated isthrou
prediction of wind speed distribution on the windward side’s strong wind region
dings 2024, 14, x FOR PEER REVIEW 21 of
Buildings 2024, 14, 2613 20 of 26
Buildings 2024, 14, x FOR PEER REVIEW 21 of 27

achievable,
ensuringwith the region
high prediction exhibiting
accuracy high relative
overall. Specifically, error prediction
the accurate mainly observed
of wind in t
speed distribution
building’s wake area.onThe
the windward
reason side’s
for strong
this is wind
that theregion is achievable,
training set with the
shares manyregionsimilarit
achievable, with the region exhibiting high relative error mainly observed in the
with exhibiting
the windhigh relative
speed error mainly in observed in the building’s wakewhile
area. The reason for traini
building’s
this is thatwake area.distribution
the training The reason
set shares many
theisstrong
for this
similarities thewind
that with area,
training
the wind set shares
speed
insufficient
many
distribution similarities
in the
samples
with lead
thewind
strong to higher
windarea,
speed influence
while of
distribution building shape
in the strong
insufficient training sampleswind on the weak
lead toarea,
higher wind
while region
insufficient
influence (e.g.,
of building the wa
training
zone).
samples
shape on lead
thetoweak
higher influence
wind of building
region (e.g., the wakeshape
zone).on the weak wind region (e.g., the wake
zone).

(a) gen_total_loss
(a) gen_total_loss (b)dis_loss
(b) dis_loss

Figure 23.23.
Figure
FigureTraining loss
23.Training
Traininglossdiagram
loss diagram of
diagram Pix2Pix
ofPix2Pix
of model
Pix2Pixmodel
model (Source:
(Source:
(Source: Authors).
Authors).
Authors).

Figure 24.24.Diagram
Figure Diagramof
of the structuralsimilarity
the structural similarity (SSIM)
(SSIM) measurement
measurement system
system [56] (Source:
[56] (Source: FiguresFigures
modified
modifiedfrom
fromWang
Wang Z,
Z, et al.,
al., 2004@IEEE,
2004@IEEE,reprinted
reprinted with
with permission
permission of IEEE).
of IEEE).
Figure 24. Diagram of the structural similarity (SSIM) measurement system [56] (Source: Figu
modified from Wang Z, et al., 2004@IEEE, reprinted with permission of IEEE).
Buildings
Buildings2024, 14,14,
2024, 2613
x FOR PEER REVIEW 22 of 26
21 of 27

Figure 25. Four cases of wind-speed analysis comparison of prediction and ground truth
Figure Authors).
(Source: 25. Four cases of wind-speed analysis comparison of prediction and ground truth (Source:
Authors).
Buildings 2024, 14, x FOR PEER REVIEW 23 of 27
Buildings 2024, 14, 2613 22 of 26

Figure 26. Statistics of SSIM values for the 50 cases (Source: Authors).
Figure 26. Statistics of SSIM values for the 50 cases (Source: Authors).
4. Discussion
4. Discussion
This research presents a novel framework for automating wind environment simula-
tionsThis researchdesigns
for building presentsbya novel framework
incorporating for automating techniques
machine-learning wind environment simula-
into parametric
tions for building
performance designs
analysis tools.by
Theincorporating machine-learning
auto-wind framework techniques
enables real-time into
wind parametric
environment
performance
analysis analysis
during tools. Thedesign
the preliminary auto-wind
stage,framework enables for
making it helpful real-time wind environ-
architectural applica-
ment analysis during the preliminary design stage, making it helpful for architectural
tions. The benefits and limitations of the auto-wind framework are discussed as follows: ap-
plications. The benefits and limitations of the auto-wind framework are discussed as fol-
4.1.
lows:Advantages
The effectiveness of the auto-wind framework was assessed through a case. As previ-
4.1. Advantages
ously explained, the auto-wind framework automated the creation of the design–simulation
dataset
Theand trained a Pix2Pix
effectiveness model forframework
of the auto-wind predicting was
the building’s wind environment.
assessed through It
a case. As pre-
was substantiated that the prediction can be made in seconds by providing
viously explained, the auto-wind framework automated the creation of the design–simu- the building
model information.
lation dataset In particular,
and trained auto-wind
a Pix2Pix model forcanpredicting
provide highly accurate predictions
the building’s for
wind environ-
buildings
ment. It waswithsubstantiated
varying geometric parameters,
that the predictionimproving
can be made simulation efficiency
in seconds and saving
by providing the
computational
building modelresources. TheIn
information. framework
particular,considers
auto-windwind-related
can provideperformance exclusively
highly accurate predic-
in this study, but after adjusting the relevant parameters, it can be seamlessly
tions for buildings with varying geometric parameters, improving simulation efficiency integrated
with other building
and saving performance
computational tools,
resources. Thesuch as Ladybug
framework and Honeybee
considers wind-related in Grasshopper.
performance
This feature significantly reduces the need for design optimization techniques
exclusively in this study, but after adjusting the relevant parameters, it can be and shortens
seamlessly
the design cycle.
integrated with other building performance tools, such as Ladybug and Honeybee in
Grasshopper. This feature significantly reduces the need for design optimization tech-
4.2. Limitations
niques and shortens the design cycle.
First, we generated 1200 building cases to train and evaluate the machine-learning
model while generating building blocks. However, the dataset is still small. In this study,
4.2. Limitations
we only examined rectangular blocks because they are the predominant building type.
First, we generated 1200 building cases to train and evaluate the machine-learning
Additionally, we simplified the buildings to reduce modeling and meshing complexity and
model while generating building blocks. However, the dataset is still small. In this study,
enhance computational efficiency because some architectural details only minimally impact
we only examined rectangular blocks because they are the predominant building type.
the wind environment. However, the simplified buildings differ significantly from their
Additionally, we simplified the buildings to reduce modeling and meshing complexity
realistic counterparts. For future research, it is crucial to classify and analyze representative
and enhance
building forms computational efficiency
before conducting CFD because some
simulations and architectural details only minimally
training machine-learning models.
impact the wind environment. However, the simplified buildings
Furthermore, the accuracy of the DL model should be enhanced by acquiring differ significantly from
more training
their realistic counterparts. For future research, it is crucial to classify
data, especially by augmenting the geometric categories to diversify the input geometries. and analyze
Buildings 2024, 14, 2613 23 of 26

For instance, the CFD dataset might benefit from further simulations to include specific
building forms that are currently under-represented.
Second, in a realistic urban environment, wind flow could be substantially affected by
the complexity of the context. And Terrain models and tree cadastres can be processed in
both Butterfly and Eddy3D. However, such input data is not currently available in urban
tree cadastral datasets. Creating an accurate mesh at the intersection of buildings and
terrain requires extensive manual model cleanup. Due to the limited availability of relevant
tree data and to ensure acceptable simulation times, the experiment does not include terrain,
vegetation and water bodies. The simulation scope and time should be balanced when
considering the broader urban context in future studies.
Third, when analysing the urban wind environment, it has been established as a best
practice to run several individual simulations from different wind directions. Limited by
high computational costs, summer wind directions of Nanjing were used in case study.
Finally, the experiments demonstrate the advantages and potential of the auto-wind
framework in improving CFD simulation efficiency; however, it cannot replace CFD simu-
lation. CFD simulation is founded on mathematically described realistic physical models,
whereas machine-learning techniques rely on empirical predictive models. In this study,
machine-learning methods have produced less accurate predictions than CFD simula-
tions. Future research will focus on the limitations discussed above and should yield more
compelling outcomes.

5. Conclusions
This study presents a methodology that significantly reduces simulation time by in-
troducing machine-learning techniques to simulate a building’s wind environment. The
methodology uses numerical analysis methods instead of dynamic simulation software to
achieve time reduction of several orders of magnitude. This study presents an objective and
comprehensible methodology for evaluating the wind environment of a specific building
during the preliminary design stages on the Rhino–Grasshopper platform. The auto-wind
framework integrates parametric modeling, CFD simulation, batch image processing, and
training of machine-learning models, creating an easy-to-use design tool and working
framework. After training, the framework’s machine-learning models can predict a build-
ing’s wind environment in real time. Although training incurs some time costs, it remains
feasible for the entire design process. Additionally, the auto-wind framework can perform
CFD simulations with minimal computational time for architects and promptly produce
analysis results. Thus, the CFD analysis results of design solutions with different geometric
parameters can be quickly compared, providing helpful information for adjusting param-
eters. Although CFD cannot and should not be replaced in the final design phase, our
proposed method incorporates building wind environment analysis into the preliminary
design with ease and rapid feedback of the results. Consequently, it minimizes the duration
of the design cycle and increases the probability of better design outcomes.

Author Contributions: Conceptualization, L.S.; methodology, L.S.; software, L.S.; validation, L.S.;
formal analysis, L.S.; investigation, L.S.; resources, L.S.; data curation, L.S.; writing—original draft
preparation, L.S.; writing—review and editing, G.J.; visualization, L.S.; supervision, G.J.; project
administration, L.S.; funding acquisition, G.J. All authors have read and agreed to the published
version of the manuscript.
Funding: This research is funded by National Natural Science Foundation of China (52178017) and
International Science and Technology Cooperation Fund Project (SJXTGJ2104).
Data Availability Statement: The datasets presented in this article are not readily available because
the data are part of an ongoing study.
Conflicts of Interest: The authors declare no conflict of interest.
Buildings 2024, 14, 2613 24 of 26

References
1. Blocken, B.; Janssen, W.D.; van Hooff, T. CFD simulation for pedestrian wind comfort and wind safety in urban areas: General
decision framework and case study for the Eindhoven University campus. Environ. Model. Softw. 2012, 30, 15–34. [CrossRef]
2. ANSYS, 2021. Fluent. Available online: https://www.ansys.com/products/fluids/ansys-fluent (accessed on 6 June 2022).
3. OpenFOAM, 2021. OpenFOAM V9 User Guide: 7.2 Turbulence Models. Available online: https://cfd.direct/openfoam/
userguide/v9-turbulence/ (accessed on 6 June 2022).
4. CHAM, 2021. Phoenics. Available online: https://www.cham.co.uk/phoenics.php (accessed on 6 June 2022).
5. Hong, T.; Chou, S.K.; Bong, T.Y. Building simulation: An overview of developments and information sources. Build. Environ.
2000, 35, 347–361. [CrossRef]
6. Van Dijk, E.; Luscuere, P. An architect friendly interface for a dynamic building simulation program. In Proceedings of the
Sustainable Building 2002, Oslo, Norway, 23–25 September 2002.
7. Lam, K.P.; Wong, N.H.; Henry, F. A study of the use of performance-based simulation tools for building design and evaluation in
Singapore. Architecture 1999, 1, 11–13.
8. Gratia, E.; De Herde, A. A simple design tool for the thermal study of an office building. Energy Build. 2002, 34, 279–289.
[CrossRef]
9. Hong, T.; Zhang, J.; Jiang, Y. IISABRE: An integrated building simulation environment. Build. Environ. 1997, 32, 219–224.
[CrossRef]
10. Warren, P. Bringing Simulation to Application; FaberMaunsell Limited: London, UK, 2002.
11. Spengler, J.D.; Chen, Q. Indoor air quality factors in designing a healthy building. Annu. Rev. Energy Environ. 2000, 25, 567–600.
[CrossRef]
12. Mochida, A.; Tominaga, Y.; Murakami, S.; Yoshie, R.; Ishihara, T.; Ooka, R. Comparison of various ke models and DSM applied to
flow around a highrise building-report on AU cooperative project for CFD prediction of wind environment. Wind Struct. 2002, 5,
227–244. [CrossRef]
13. Jiang, Y.; Alexander, D.; Jenkins, H.; Arthur, R.; Chen, Q. Natural ventilation in buildings: Measurement in a wind tunnel and
numerical simulation with large-eddy simulation. J. Wind Eng. Ind. Aerodyn. 2003, 91, 331–353. [CrossRef]
14. Han, T.; Huang, Q.; Zhang, A.; Zhang, Q. Simulation-based decision support tools in the early design stages of a green building—A
review. Sustainability 2018, 10, 3696. [CrossRef]
15. Bosselmann, P.; Arens, E.; Dunker, K.; Wright, R. Urban form and climate: Case study, Toronto. J. Am. Plan. Assoc. 1995, 61,
226–239. [CrossRef]
16. Sousa, J.P.M.; Moya, R.A.C.; Prohasky, D.; Vaz, C.E.V. Empirical analysis of three wind simulation tools to support urban planning
in early stages of design. Blucher Des. Proc. 2015, 2, 363–370. [CrossRef]
17. Si, B.; Tian, Z.; Jin, X.; Zhou, X.; Tang, P.; Shi, X. Performance indices and evaluation of algorithms in building energy efficient
design optimization. Energy 2016, 114, 100–112. [CrossRef]
18. Wortmann, T. Opossum-introducing and evaluating a model-based optimization tool for grasshopper. In Proceedings of the 22nd
CAADRIA Conference, Suzhou, China, 5–8 April 2017; pp. 283–292. [CrossRef]
19. Marcum, D.L.; Weatherill, N.P. Unstructured grid generation using iterative point insertion and local reconnection. AIAA J. 1995,
33, 1619–1625. [CrossRef]
20. Löhner, R. Automatic unstructured grid generators. Finite Elem. Anal. Des. 1997, 25, 111–134. [CrossRef]
21. Nakahashi, K.; Ito, Y.; Togashi, F. Some challenges of realistic flow simulations by unstructured grid CFD. Int. J. Numer. Methods
Fluids 2003, 43, 769–783. [CrossRef]
22. Hu, Y.; Peng, Y.; Gao, Z.; Xu, F. Application of CFD plug-ins integrated into urban and building design platforms for performance
simulations: A literature. Front. Archit. Res. 2023, 12, 148–174. [CrossRef]
23. Aish, R.; Woodbury, R. Multi-Level Interaction in Parametric Design. In Smart Graphics; Springer: Berlin/Heidelberg, Germany,
2005; pp. 151–162.
24. CHAM, 2017. RhinoCFD. Available online: https://www.cham.co.uk/rhinoCFD.php (accessed on 6 June 2022).
25. Chronis, A.; Dubor, A.; Cabay, E.; Roudsari, M.S. Integration of CFD in computational design. In Proceedings of the eCAADe
2017, Rome, Italy, 20–22 September 2017; pp. 601–610. [CrossRef]
26. food4Rhino, 2021. Rhino Flow-RT. Available online: https://www.food4rhino.com/en/app/rhino-flow-rt (accessed on
6 June 2022).
27. Kastner, P.; Dogan, T. A cylindrical meshing methodology for annual urban computational fluid dynamics simulations. J. Build.
Perform. Simul. 2020, 13, 59–68. [CrossRef]
28. Zhang, R.; Waibel, C.; Wortmann, T. Aerodynamic Shape Optimization for High-Rise Conceptual Design. In Proceedings of the
eCAADe, 2020, Berlin, Germany, 16–17 September 2020. [CrossRef]
29. Shen, Y.; Wang, L.; Zhang, R.; Tong, Z.; Ji, G. EvoMass+ GH_Wind An agile wind-driven building massing design optimization
framework. In Proceedings of the 39th eCAADe Conference, Novi Sad, Serbia, 8–10 September 2021; Volume 1, pp. 477–486.
[CrossRef]
Buildings 2024, 14, 2613 25 of 26

30. Hu, Y.; Xu, F.; Gao, Z. A Comparative Study of the Simulation Accuracy and Efficiency for the Urban Wind Environment Based
on CFD Plug-Ins Integrated into Architectural Design Platforms. Buildings 2022, 12, 1487. [CrossRef]
31. Roudsari, M.S.; Pak, M. Ladybug: A Parametric Environmental Plugin for Grasshopper to Help Designers Create an Environmentally-
Conscious Design; IBPSA: Rapid City, SD, USA, 2013.
32. ODS, 2017. Swift. Available online: https://www.odsengineering.com (accessed on 6 June 2022).
33. Jo, Y.S.; Choi, I.S. A Case of Nasal Septal Abscess Localized at Keystone Area, Suspected of Occurring Following Cauterization
for Epistaxis. J. Clin. Otolaryngol. Head Neck Surg. 2020, 31, 74–77. [CrossRef]
34. Kaijima, S.; Bouffanais, R.; Willcox, K. Computational fluid dynamics for architectural design. Archit. Des. 2013, 83, 118–123.
[CrossRef]
35. Kim, D. The Application of CFD to Building Analysis and Design: A Combined Approach of an Immersive Case Study and Wind
Tunnel Testing. Ph.D. Thesis, Virginia Tech, Blacksburg, VA, USA, 2014.
36. Guo, X.; Li, W.; Iorio, F. Convolutional neural networks for steady flow approximation. In Proceedings of the 22nd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp. 481–490.
[CrossRef]
37. Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative adversarial
nets. Commun. ACM 2014, 63, 139–144. [CrossRef]
38. Daut, M.A.M.; Hassan, M.Y.; Abdullah, H.; Rahman, H.A.; Abdullah, P.; Hussin, F. Building electrical energy consumption
forecasting analysis using conventional and artificial intelligence methods: A review. Renew. Sustain. Energy Rev. 2017, 70,
1108–1118. [CrossRef]
39. Mokhtar, S.; Sojka, A.; Davila, C.C. Conditional generative adversarial networks for pedestrian wind flow approximation. In
Proceedings of the 11th Annual Symposium on Simulation for Architecture and Urban Design, Virtual, 25–27 May 2020; pp. 1–8.
40. Duering, S.; Chronic, A.; Koenig, R. Optimizing Urban Systems: Integrated optimization of spatial configurations. In Proceedings
of the 11th Annual Symposium on Simulation for Architecture and Urban Design, Virtual, 25–27 May 2020; pp. 1–7.
41. Yousif, S.; Bolojan, D. Deep-Performance: Incorporating Deep Learning for Automating Building Performance Simulation
in Generative Systems. In Proceedings of the 26th Association for Computer-Aided Architectural Design Research in Asia
(CAADRIA), Hong Kong, China, 29 March–1 April 2021; Volume 1, pp. 151–160. [CrossRef]
42. Zhao, X. A Multi-Objective Optimization Design Method Integrating CGAN for Wind Environment in Severe Cold Regions.
Ph.D. Thesis, Harbin Institute of Technology, Harbin, China, 2022.
43. Tan, C.; Zhong, X. A Rapid Wind Velocity Prediction Method in Built Environment Based on CycleGAN Model. In Proceedings of
the Computational Design and Robotic Fabrication (CDRF) 2022, Shanghai, China, 25–26 June 2022; Springer: Singapore, 2023.
[CrossRef]
44. Huang, C.Y. Study on Environmental Performance-driven Multi-objective Optimization Design Method for Climate-adapted
Urban Blocks. Ph.D. Thesis, North China University of Technology, Beijing, China, 2023.
45. Wang, L.; Chen, K.W.; Janssen, P.; Ji, G. Algorithmic generation of architectural massing models for building design optimisation:
Parametric modelling using subtractive and additive form generation principles. In Proceedings of the 25th International
Conference on Computer-Aided Architectural Design Research in Asia, CAADRIA 2020, Bangkok, Thailan, 5–6 August 2020;
Volume 1, pp. 385–394. [CrossRef]
46. Wang, L.; Janssen, P.; Chen, K.W.; Tong, Z.; Ji, G. Subtractive building massing for performance-based architectural design
exploration: A case study of daylighting optimization. Sustainability 2019, 11, 6965. [CrossRef]
47. Elwy, I.; Ibrahim, Y.; Fahmy, M.; Mahdy, M. Outdoor microclimatic validation for hybrid simulation workflow in hot arid climates
against ENVI-met and field measurements. Energy Procedia 2018, 153, 29–34. [CrossRef]
48. Estrado, E. Optimisation of Complex Geometry Buildings Based on Wind Load Analysis. Master’s Thesis, Delft University of
Technology, Delft, The Netherlands, 2019. [CrossRef]
49. Brook-Lawson, J.; Holz, S. CFD comparison project for wind simulation in landscape architecture. J. Digit. Landsc. Archit. 2020, 5,
318–329.
50. Xu, X.; Wu, Y.; Wang, W.; Hong, T.; Xu, N. Performance-driven optimization of urban open space configuration in the cold-winter
and hot-summer region of China. In Building Simulation; Tsinghua University Press: Beijing, China, 2019; Volume 12, pp. 411–424.
[CrossRef]
51. Sun, C.; Rao, J. Study on performance-oriented generation of urban block models. In Proceedings of the 2019 Digital FUTURES:
The 1st International Conference on Computational Design and Robotic Fabrication (CDRF 2019), 1st ed.; Springer: Singapore, 2020;
pp. 179–188. [CrossRef]
52. Jia, B.; Liu, S.; Ng, M. Air quality and key variables in high-density housing. Sustainability 2021, 13, 4281. [CrossRef]
53. Sakiyama, N.R.M.; Frick, J.; Bejat, T.; Garrecht, H. Using CFD to evaluate natural ventilation through a 3D parametric modeling
approach. Energies 2021, 14, 2197. [CrossRef]
54. Meng, F.-Q.; He, B.-J.; Zhu, J.; Zhao, D.-X.; Darko, A.; Zhao, Z.-Q. Sensitivity analysis of wind pressure coefficients on CAARC
standard tall buildings in CFD simulations. J. Build. Eng. 2018, 16, 146–158. [CrossRef]
Buildings 2024, 14, 2613 26 of 26

55. Waibel, C.; Bystricky, L.; Kubilay, A.; Evins, R.; Carmeliet, J. Validation of Grasshopper-based Fast Fluid Dynamics for Air Flow
around Buildings in Early Design Stage. In Proceedings of the 15th International Conference of IBPSA-Building Simulation 2017,
San Francisco, CA, USA, 7–9 August 2017. [CrossRef]
56. Wang, Z.; Bovik, A.C.; Sheikh, H.R.; Simoncelli, E.P. Image quality assessment: From error visibility to structural similarity.
IEEE Trans. Image Process. 2004, 13, 600–612. [CrossRef] [PubMed]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.