Generative Design

• physical
• computational

Design Optimization
• physical
• computational

Parametric Design
• scalar
• repetitive / modular
• algorithmic

Genetic Design
• search
• optimization
• exploration
University of Michigan, TCAUP Optimization Slide 1 / 78

Optimization

What to optimize (levels)

• material
• member
• geometry
• topology

How to optimize (methods)

• physical
• computational

University of Michigan, TCAUP Optimization Slide 2 / 78

Optimization

Material Minimize
Member Weight
Geometry Volume
Topology Cost
Maximize
Strength
Strain Energy
Elasticity
Ductility
Stiffness

University of Michigan, TCAUP Optimization Slide 3 / 78

Optimization

Material Strength (Full Utilization)

Member in section
Geometry in length
Topology Stability (no buckling)
in section
in length
Serviceability
in use
maintenance
Connections
stress transfer

University of Michigan, TCAUP Optimization Slide 4 / 78

Optimization

Material
Member Arrangement of Nodes
Geometry best overall for members
Topology includes optimization of :
• material
• members

University of Michigan, TCAUP Optimization Slide 5 / 78

Optimization

Material
Member Arrangement of Members
Geometry includes optimization of :
Topology • material
• members
• geometry

University of Michigan, TCAUP Optimization Slide 6 / 78

Geometry vs. Topology

definitions

University of Michigan, TCAUP Optimization Slide 7 / 78

Methods
Physical
• structural models
• form finding models
• analog models

Computational
• Linear Programming (Simplex)
• Homogenous – sensitivity
• Stochastic/Probabilistic Algorithms
• GA (Genetic Algorithms)
• ES (Evolutionary Strategies)
• SA (Simulated Annealing)
• Swarming (Particle Swarm Optimization)

University of Michigan, TCAUP Optimization Slide 8 / 78

Methods

Physical
structural models
form finding models
analog models
Louis I. Kahn's structural
model for the Richards
Medical Research Laboratory

Frei Otto’s model of the cable

nets of the 1972 Olympics

Stadium in Munich from

1972 Olympics

University of Michigan, TCAUP Optimization Slide 9 / 78

Methods

Physical
structural models
form finding models
analog models

Heinz Isler, Burgdorf

University of Michigan, TCAUP Optimization Slide 10 / 78

Methods

Physical
structural models
form finding models
analog models

University of Michigan, TCAUP Optimization Slide 11 / 78

Methods

Computational
Linear Programming (Simplex)
Homogenous – sensitivity
Evolutionary Algorithms
ES (Evolutionary Strategies)
GA (Genetic Algorithms)

University of Michigan, TCAUP Optimization Slide 12 / 78

Methods

Computational
Linear Programming (Simplex)
Homogenous – sensitivity
Evolutionary Algorithms
ES (Evolutionary Strategies)
GA (Genetic Algorithms)

University of Michigan, TCAUP Optimization Slide 13 / 78

Methods

Computational
Linear Programming (Simplex)
Homogenous – sensitivity
Evolutionary Algorithms
ES (Evolutionary Strategies)
Rechenberg & Schwefel
GA (Genetic Algorithms)
John Holland

(1 , 1) random walk
(1+1) one parent  one child - select from all
(+1) multi-parent  one child - select from all
(+) multi-parent  multi-child - select from all
( , ) multi-parent  multi-child - select from children
(Bäck, 1996)

University of Michigan, TCAUP Optimization Slide 14 / 78

Methods

Computational
Linear Programming (Simplex)
Homogenous – sensitivity
Evolutionary Algorithms
ES (Evolutionary Strategies)
GA (Genetic Algorithms)
Selectorecombinative
• selection
• crossover
• mutation

University of Michigan, TCAUP Optimization Slide 15 / 78

Genetic Algorithms

crossover

University of Michigan, TCAUP Optimization Slide 16/22

Genetic Algorithms

crossover inversion

University of Michigan, TCAUP Optimization Slide 17/22

Genetic Algorithms - Geometry

half-uniform crossover

University of Michigan, TCAUP Optimization Slide 18/22

 Genetically Enhanced Parametric Design for
Performance Optimization and Design Exploration

The ParaGen Method

Form Exploration in
Early Design Phases

Based on Parametric
Geometry, GA search
with a SQL database

Peter von Buelow, Dr. –Ing.

Professor in Architecture
University of Michigan

Peter von Buelow University of Michigan 19 / 78

Aspects of Early Design

Purposeful
directed – not merely random

Goal Oriented
search to find good solution

Creative
seeks new solutions

Ill Structured Problems

problem space not fully defined
cannot solve directly – cyclic

exploration is needed

Peter von Buelow University of Michigan 20 / 78

Exploration with Evolutionary Computation

deterministic optimization
focuses on one ‘best’ solution

It is better suited for later

design phases

exploration tries to expose a

range of ‘pretty good’ solutions

it aids ideation and creativity in

early design phases

Peter von Buelow University of Michigan 21 / 78

Why GA’s fit well to early design

Early Design GA Design

Explorative Populations of solutions
Recursive Operates in cycles
Serendipitous Random mutations
Knowledge of design is low No knowledge of fitness function

Peter von Buelow University of Michigan 22 / 78

 The ParaGen Exploration Method

Based on an NDDP-GA The ParaGen Framework

Oriented toward Exploration
• Helps designer understand
the range of solutions

Non-Destructive
• All solutions are visible

Dynamic Populations
• Well suited for multi-objectives

Both Interactive and Algorithmic

• Can consider non-computational
criteria, e.g. aesthetics

Peter von Buelow University of Michigan 23 / 78

ParaGen is Non-Destructive
Traditional GAs are destructive – solutions are lost

An NDDP-GA is non-destructive
no solutions are lost

Peter von Buelow University of Michigan 24 / 78

 GA + SQL = Exploration

Seven Points:
1. Store no duplicates
2. Multi-objective
3. Dynamic populations
4. Changeable search
direction - corrective
5. Interactive search
6. Pareto optimization
7. Parallel computation

Peter von Buelow University of Michigan 25 / 78

1. Store All Solutions

without Duplicates
Relational Database Size
• terabyte range (e.g. 256 TB in
MySQL, but limited by system)

SQL unique index

• set at child generation
• checked at data entry

Faster analysis
• no duplicate solutions to analyze

More effective post analysis

• no duplicate solutions to view
• more descriptive graphing

Less storage space needed

• only unique solutions are retained

Peter von Buelow University of Michigan 26 / 78

 2. Use Multi-Objective Fitness Functions

Fitness defined by SQL query

• simple sorts – single variable
e.g. 20 best
• defined range – single variable
e.g. between max. and min.
• Multi-objective query
e.g. Pareto set optimization
e.g.
weight less than x AND
height more than y AND
heat gain less than z

Separate fitness for each parent

• each parent can be obtained
using a different query set

Peter von Buelow University of Michigan 27 / 78

3. Create Dynamic Parent Populations

Independent breeding
population
• Not dependent on
previous population
• Non-generational or
static population
• Evolution through
addition (no loss)

Peter von Buelow University of Michigan 28 / 78

 4. Change Search Direction
Search different parts of solution space
• see range of solutions
• bracket uncertainty
• conflicting design desires

Change fitness without restarting

• changing the fitness criteria merely shifts
the view

Aids in post analysis

• allows for the comparison of conflicting
solutions

Allows correction of objectives

• objectives can be easily refined as more
data is made available

Peter von Buelow University of Michigan 29 / 78

Search different parts of solution space

Change fitness without restarting

Peter von Buelow University of Michigan 30 / 78

 5. Interactive Exploration of Solutions

No delay for analysis

• near immediate response to queries

Variety of graphic depiction

• multiple image views (different information)
• control visual array display

Aids in post analysis

• allows for the comparison of conflicting
solutions

Peter von Buelow University of Michigan 31 / 78

Control visual array display

Peter von Buelow University of Michigan 32 / 78

 6. Versatile Graphing Data Analysis

x-y scatter graphs

• filters to set range of points
• plot any two variables or values
• control axis direction
• Show third parameter with color

parallel coordinate graphs

• filters to set range of points
• highlight any set (box select)

Peter von Buelow University of Michigan 33 / 78

x-y scatter graphs

Peter von Buelow University of Michigan 34 / 78

 x-y scatter graphs

Peter von Buelow University of Michigan 35/ 78

parallel coordinate graphs

Peter von Buelow University of Michigan 36 / 78

 Pareto optimization

Define conflicting objectives

e.g.
weight and stiffness

Determine non-dominated solutions

e.g.
For a given solution there is
no lower weight with a higher
stiffness

Define the Pareto set of non-dominated

solutions

Sort and breed neighbors

or
Select and mutate individuals

Peter von Buelow University of Michigan 37 / 78

Pareto optimization

Selection can contain any number of

objectives

e.g.
weight and deflection and stress

Peter von Buelow University of Michigan 38 / 78

 7. Utilize Parallel Hardware

Dedicated cluster (cloud)

or

ad hoc cluster
• Simple web browser connection

Peter von Buelow University of Michigan 39 / 78

Exploration as Database Mining

Google maps
ParaGen maps

Peter von Buelow University of Michigan 40 / 78

 How it works
1. Select
csv choose set

SQL
database
2. Breed
5. Upload to SQL

Server Side

The ParaGen Cycle

Excel

3. Grow Client Side (parallel)

4. Evaluate
DXF

weight

Peter von Buelow University of Michigan 41 / 78

1. Selection
Interactive selection
client side
via web interface

GA selection
server side

Peter von Buelow University of Michigan 42/ 78

2. Breeding
• Create population from data
• Crossover of data variables
• Half Uniform Crossover
• Output to Excel

SQL database

HUX

Peter von Buelow University of Michigan 43 / 78

3. Grow the Form

• Run parametric software
with Excel input.
• Generate form.
• Output dxf file.
• Capture image.

Peter von Buelow University of Michigan 44 / 78

4. Evaluate Form

• Read into STAAD.Pro

• Add material, support and load
conditions
• Find force and deflection
• Size members
• Determine weight
• Output data to Excel

Peter von Buelow University of Michigan 45/ 78

4. Evaluate Form

Other performance evaluations

can be made with any
Windows software package.

Example from Ecotect

Peter von Buelow University of Michigan 46 / 78

5. Upload Files and Data to Web Server
All variables and performance values are stored in a SQL database

Useful files are retained for later inspection: e.g. DXF, JPG, VRLM,
and input files for FEA and Ecotect

Peter von Buelow University of Michigan 47 / 78

5. Store and Rank Data

Multi level sort options + multi level filters, together with image browsing of
solutions provides robust exploration potential.

Peter von Buelow University of Michigan 48 / 78

How to use it
Post processing
Exploration
Selection

Peter von Buelow University of Michigan 49 / 78

Examples - Shukhov Tower

lightest weight least members Dead + Wind Load

Peter von Buelow University of Michigan 50 / 78

 Examples - Shukhov Tower

Plot of weight vs. members

VRML perspective

Peter von Buelow University of Michigan 51 / 78

Examples (with Michela Turrin & Maria Van Embden Andres)

Delaunay Domes

Voronoi Domes

Peter von Buelow University of Michigan 52 / 78

Examples (with Michela Turrin & Maria Van Embden Andres)

Geometry Generation Structural Performance

Peter von Buelow University of Michigan 53 / 78

Examples (with Andre Chaszar)

Soccer Stadium - variables

oculus size

dome height dome

curvature

Peter von Buelow University of Michigan 54 / 78

Examples (with Andre Chaszar)
Soccer Stadium - performance

• Geometry: curvature, oculus, mesh density, number of elements

• Structural: weight, deflection, modal frequency
• Daylighting: light intensity and distribution on field

Peter von Buelow University of Michigan 55 / 78

Examples (with Andre Chaszar)

Soccer Stadium - geometry

Peter von Buelow University of Michigan 56 / 78

Examples (with Andre Chaszar)
Soccer Stadium - modal frequencies

Peter von Buelow University of Michigan 57 / 78

Examples - Foam
Examples - Foam & Fabric
& Fabric - Wright
- with M. with M. Wright
& M. Jensen& M. Jensen

Designer selected
Designer selected
populations based on
populations based on
visual criteria.
visual criteria.
Generated with a Python
Generated with a Python
script in Rhino
script in Rhino
Full solution set of 2200
Full solution set of 2200

P. von Buelow
Peter von Buelow
58 of 49
University of Michigan
Taubman College
58 / 78
 Examples - Foam
Examples - Foam & Fabric
& Fabric - Wright
- with M. with M. Wright
& M. Jensen& M. Jensen

Post process
Post process
exploration with SQL
exploration with SQL
sorts and filters
sorts and filters
Reduced set of 32
Reduced set of 32

P. von Buelow
Peter von Buelow
59 of 49
University of Michigan
Taubman College
59 / 78

Examples (with Michela Turrin & Andreas Falk) SolSt

Shading / Ventilation

Branching Columns

Peter von Buelow University of Michigan 60 / 78

Examples (with Michela Turrin & Andreas Falk) SolSt

Peter von Buelow University of Michigan 61 / 78

Examples (with Michela Turrin & Andreas Falk) SolSt

Peter von Buelow University of Michigan 62 / 78

Examples – Geodesic Domes (with Ted Hall)

Snow Loading

Variable parameters
• Outer shell polyhedral type
• Power – subdivisions
• Frequency
• Base trim
• Distance between layers

Peter von Buelow University of Michigan 63 / 78

Examples – NetZeroEnergy (with Studio [Ci] + Constance Bodurow)

Peter von Buelow University of Michigan 64 / 78

Examples – NetZeroEnergy (with Studio [Ci] + Constance Bodurow)

PV on all panels PV on 6 panels

Peter von Buelow University of Michigan 65 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Tree Crotch Joints

Used in 18th century ship

building

Replaces mortise and tendon

joint with a single element

Grain of wood follows the

force path

Can be milled to fit

dimensions of a given
structure

Peter von Buelow University of Michigan 66 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Tree Crotch Joints

Final joints milled from the raw

tree crotch

Peter von Buelow University of Michigan 67 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Based on hexagonal mesh

Joints connect 3 struts

Designed as a catenary vault

Form adjusted to fit given

joints

Parametric exploration of
topology and geometry
options using
Rhino/Grasshopper with
Kangaroo and Karamba

Peter von Buelow University of Michigan 68 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Design Parameters

• Minimize out of plane

curvature
• Naturally occurring
bifurcation angles (30-90°)
• Axially aligned joints and
connectors
• Minimize number of joints
• Smooth shell surface
• Entrance and center height
• Weight
• Member forces
• Deformation

Peter von Buelow University of Michigan 69 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Form Generation

• Using ParaGen exploration

method with Grasshopper
• Each solution analyzed in
Grasshopper
• Solutions saved with images
in MySQL
• Initial generation ca. 500
• Breeding for:
entrance > 2 m (6.5 ft)
peak < 3 m (10 ft)
minimum joints

Peter von Buelow University of Michigan 70 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)
Collected Data

Peter von Buelow University of Michigan 71 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Peter von Buelow University of Michigan 72 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Peter von Buelow University of Michigan 73 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Constructed Prototype

Peter von Buelow University of Michigan 74 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Load Testing

• Load test to failure

• Hung sand buckets from

nodes

• Deflection measurements
on cards

• Maximum load: 311 N/node

= 0.67kN/m2
• (70 lbs/node = 14 psf)

• Wood cracked at node

Peter von Buelow University of Michigan 75 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Load Testing

Peter von Buelow University of Michigan 76 / 78

Examples – LIMB (with Steven Mankouche & Kasey Vliet)

Load Testing

• Load test to failure

• Hung sand buckets from

nodes

• Deflection measurements
on cards

• Maximum load: 311 N/node

= 0.67kN/m2
(70 lbs/node = 14 psf)

• Wood cracked at node

Peter von Buelow University of Michigan 77 / 78

Genetically Enhanced
Parametric Design for
Performance Optimization
and Design Exploration

Additional Information:
pvbuelow@umich.edu
www.umich.edu/~pvbuelow

Peter von Buelow, Dr. –Ing.

Professor in Architecture
Taubman College
University of Michigan

Peter von Buelow University of Michigan 78 / 78