AI/ML and Simulation in the

Context of Product Innovation

Prith Banerjee, Ph.D.

Chief Technology Officer, Ansys

November 4, 2020
 Ansys is the Simulation Leader

MISSION: Empower our customers to design and deliver transformational products

FOCUSED TRUSTED More than

PROVEN GLOBAL
100
This is all we do. 45,000 Member of the 4,500+
Leading product technologies in all
physics areas. Largest development
97100
of the
customers worldwide
prestigious
500
employees globally

industrials
ISO 9001
team focused on simulation
CERTIFIED $30B+ market capitalization

LARGEST INDEPENDENT COMMITTED DRIVEN

3x the size of our nearest
competitor (revenue)
Long-term financial stability

CAD/PLM/IoT agnostic
Overall customer satisfaction
globally is at
89%
Helping customers address new
market challenges: digitalization,
Open Architecture 5G, autonomy, electrification,
$1.5 Billion additive and digital twins

2
 Journey From Physics-Leadership to Engineering

ANSYS Transformational Journey

Operation /
Maintenance
50
3B$+ Technology
Partners
Investments Virtual
prototype

Design
Exploration

Physics | Process | Software | Controls | Engineering Analysis | Digital Technologies

3
Ansys Offers the Only True Simulation Platform

COMPREHENSIVE SOLUTIONS: Autonomy / Electrification / 5G / IIot

Digital Twins
SYSTEMS & Embedded
Software
EMBEDDED SOFTWARE

Fluids Structures Electromagnetics Semiconductor Design Exploration Optical

MATERIALS INFORMATION

PLATFORM AND CLOUD

4
Simulation Impacts Top-Line Growth And Bottom-Line Savings
Revenue growth

Simulation impact Offer more products

Launch right products

Rapid innovation
Faster time to market
Lower cycle time
PRODUCT
LIFECYCLE Reduced risks Cost savings
Increased quality
Improved R&D efficiency
Manage complexity
Fewer physical prototypes

Lower warranty costs

5
Fluids R&D directions

PRODUCTIVITY GAS TURBINES EXTERNAL AERODYNAMICS ELECTRIFICATION

• Workflows • Combustion physics • Speed • Conjugate Heat Transfer
• Customer driven • Numerics • Accuracy Electro-Chemistry
• Speed and robustness • Spray physics • Robustness • Battery Life
• GPU based post processing • Turbulence • Hypersonic physics • Capacity fade
• Usability and repeatability • Best practices • Multiphase flows • Process and Chemical Industries

6
Structures R&D Directions
APPLICATION SPECIFIC ELECTRONICS RELIABILITY
• Medical Initiative: heart simulation • Best-in-class pre-processing for seamless model exchange
• Electrophysiology, Incompressible flow, FSI, Adv. Mat. • Coupling Ansys Mechanical and Sherlock
• Additive Manufacturing
MORPHING-BASED OPTIMIZATION
• Machine learning for faster thermal solutions
• Better topology optimization
• Microstructure & property predictions
• Automatic creation of clean CAD
• Battery modeling in crash analysis
SOLVER METHODS
MATERIALS
• Iso Geometric Analysis (IGA)
• Multiscale – Data Driven framework, Short fibers
• Inverse methods
• Crystal plasticity
• Dual Lagrange Multiplier methods for contact
• Elastomers fatigue, inelastic fracture
• Manufacturing constraints in Topo Optimization
• Thermo mechanical fatigue with damage
• ICME SOFTWARE
• DCS:
EQUATION SOLVERS
• Distributed Log Service
• Data driven means to assess resource needs
• HPC 2.0 – HPC on demand
• Low Rank meth. for acoustic wave equations
• System Coupling for multi-physics
• Efficient scheduling for HPC performance

AUTOMOTIVE APPLICATIONS
• Streamlined Welding for Automotive Chassis
• New Meshing and geometry Enable Optimization Workflows

7
Electromagnetics R&D Directions
AUTOMATION AND PERFORMANCE
• Streamlined electronics workflows for package and board
• Integration with thermo-mechanical for new applications
ANTENNA
• Advances in phased array antenna analysis with scalable HPC.
SEMICONDUCTORS
• RaptorH combines Helic’s RaptorX with Ansys HFSS,
• Gold standard verification for sensitive 5G integrated circuits
ELECTRIFICATION – RELIABILITY AND NVH
• New thermal and vibrational analysis coupled with Maxwell
ANSYS HFSS
• Enhanced distributed direct matrix solver available on Cloud.
• Significantly reduced memory
• Scales to 100’s of cores.
NEW ANSYS EMA 3D PRODUCT
• Models EMI/EMC between complex cables
NEW CAPABILITIES IN ANSYS MAXWELL
• Speeds simulation by creating/solving model for rotating machines.
Setup in full model Solved by partial model Visualized in full model

8
Ansys Long-Term Technology Strategy Dimensions

ARTIFICIAL INTELLIGENCE AND MACHINE LEARNING DIGITAL AND PHYSICAL WORLDS

• Simulation used to train AI methods • AR/VR for Simulation brings digital world to physical
• AI used to improve simulation • IOT and Connectedness brings physical world to digital

PLATFORM FOR MULTIPHYSICS SIMULATION DIGITAL TRANSFORMATION

• Robust Multiphysics, Multi-disciplinary Optimization • Digital Threads, Digital Continuity, Digital Twins
• Micro-services for simulation • Model Based Systems Engineering

HYPERSCALE SIMULATION, COLLABORATION ON CLOUD COMPUTATIONAL METHODS IN NEW AREAS

• GPU, SMP, MPI, task based • Integrated Computational Materials Engineering
• Quantum computing • Computational Chemistry, Biology, Healthcare, Photonic IC

PREDICTIVE AND ROBUST DESIGN FUTURE OF SIMULATION

• High accuracy, adaptive numerical methods • ML based PDE -Solver, generative design
• Materials variation and uncertainties • Integrated synthesis and verification
• Automated mixed mode (0D-4D), MF, MS simulation

9
 AI/Machine Learning and Simulation
Machine Learning, defined by Arthur Samuel, 1959:
“Field of study that gives computers the ability to learn without being explicitly programmed”.

AI/ML Use Cases Engg. Simulation Methods

Machine Learning Methods • Customer Productivity • Geometry/Meshing
• Supervised Learning • Augmented Simulation • 3D/4D Finite Element/Volume
• Unsupervised Learning • Engineering Design • Physical/Math Models
• Semi-supervised Learning • Business Intelligence • Reduced Order & System
• Reinforcement Learning Models, Digital Twins
• Postprocessing/Visualization

10
Why is Machine Learning Relevant today?
• Data storage is cheap
• Almost infinite compute capacity with rise in GPUs/TPUs

Data Compute

Open source datasets which accelerated the pace of research

Machine
Learning

• Open source and ML Frameworks and datasets

• Growing community through top AI/ML conferences

Open source ML frameworks that survived test of time

Top ML conferences

11
 How does Deep Learning Work?

• Single neuron
- Outputs an activation(a) of weighted sum of
input (x)
- Simplest activation: ReLU
Activation Function: ReLU
(Rectified Linear activation)
• Universal Approximation Theorem
(Cybenko, 1989): 𝑗𝑡ℎ neuron: 𝜎(𝑤𝑗𝑇 𝑋 + 𝑏𝑗 )
- A neural network with one hidden layer
containing a finite number of neurons can
approximate any continuous function to a
reasonable accuracy. ≈ ≈
𝑓 𝑋 NN X = σ𝑁 𝑇
𝑗=1 𝑎𝑗 𝜎(𝑤𝑗 𝑋 + 𝑏𝑗 )
- >1D, no problem, but must increase neurons
Explaining universal approximation theorem in 1D: 𝑓 ≈ 𝑁𝑁

• If we increase the numbers of layers

- Can approximately represent many complex
functions, say one function per layer
- Can handle any non-linearity through ReLU
- Why didn’t we have perfect learning systems?

A deep neural network

 How does Deep Learning Work?

• Is there a relationship between input X W2 W3

W4
and observation Y?
xk
• Yes, Given pairs(X,Y), Neural Network aj
yk
tries to fit a function ℱ(), which is
parameterized by weights(w), such that : Ŷ(Prediction)
Ŷ= 𝓕(𝒘, 𝑿).
X(Input)
• Key Steps
- Initialize weights randomly or transfer from a previous learned task of a
similar network
- Feed Forward:
• In this step NN predicts the Ŷ based on the current weights 𝒘 and inputs.
• Calculate Error(𝒥(𝑤)) = Norm of (Y- Ŷ)
- Backpropagation:
• Based on the Error(𝒥), gradient of Error w.r.t. weights i.e. ∇(𝒥(𝒘)) is calculated.
• Gradient Descent: 𝑤 = 𝑤 − 𝛼. ∇ 𝒥(𝑤) , where 𝛼 is learning rate

13
AI/Machine Learning at Ansys
Build internal expertise and support organic R&D
• Hired a small team of AI/ML software developers and leader in the Office of the CTO
• Working with a distributed team of AI/ML software developers in the Business Units
• Support, complement, accelerate BU projects short term
• Canonical customer use cases, drive results across BUs long term
• Customer Productivity
• Augmented Simulation
• Engineering Design
• Business Intelligence
Use partnerships to augment:
• Larger companies
• Startups
• Universities

14
 AI/ML for Engineering Simulation
R&D Projects in BUs • Materials BU

CTO Office • ML Analysis for Additive Manufacturing data (Revolutionizing design)

• ML based PDE Solver (Augmented simulation) • ML methods in Granta Materials (Customer productivity)
• Physics Informed Neural Network (Augmented sim)
• Optimization for AMG Solver (Customer productivity) Semiconductor BU
• ML Framework as a service ((Customer productivity) • Data-driven fast static/transient thermal solver (Augmented simulation)
Mechanical BU
• Nvidia: Physics Informed Neural Network Solvers (Augmented
•Auto-identification of contacts (Customer productivity) simulation)
•Google: Reinforcement Learning with Py-ANSYS
(Customer productivity) • AutoML for SI and PI tools (Customer productivity)
•ML based surrogate modeling for VVUQ (Augmented System and Platform BU
simulation) • Automated ROM Creation for non-linear systems (Augmented simulation)
Fluids BU
• Creation of Path Planners for autonomous vehicles (Customer
• Geminus: Turbulence model tuning using ML (Customer productivity)
productivity)
• Geminus: Augmented flow for Twinbuilder (Augmented simulation)
• AI/ML-enabled flow solver (Augmented simulation)
Electronics BU Cloud and Platform BU
•Runtime Predictor (Business intelligence) • Resource prediction compute and storage on cloud (Business intelligence)
•Auto-identify HFSS 3D regions in SIwave (Customer
productivity) ACE
Design BU - Virtual Validation: Fusion modeling + less tests (Augmented simulation)
•Fast Topology Optimization (Revolutionizing design) - Generative design for OEM Concept modeling (Revolutionizing design)

15
Augmented Simulation: Data-Driven Fast On-Chip Thermal Solver
Providing Inferencing by Striding Templates with One Tile for Predicting Temperature at all Tile Locations

P1 P2 P3

P4 P4 P5 Temp

P6 P6 P6 x j
k
y
k

For each tile location, a set of tile

powers and Theta-JA are input to
the trained DeltaT Predictor to
predict the temperature at the
corresponding tile and added to
the Temp from rough thermal
profile

16
 Augmented Simulation: Physics-Informed Neural Nets (PINN)
/ 3D transient heat transfer with source

/ 2D heat transfer on a chip with nonlinear power map

• PINN uses upfront inference modeling

• Runtimes in fractions of seconds (100X)

Power map Numerical PINN Comparison

17
 Augmented Simulation: ML-Based Partial Differential Equation Solver

• Motivation: Example Results: Flow over cylinder: Velocity

- Geometries and physics have lots of patterns.
Fluent Predicted
- Do we need to solve from scratch?
• No! Create solution instances on patches using a generative network
• Key insights:
- Take a general domain and voxelize into patches
- Decide boundary conditions for each patch
- "Solve” for the latent vectors on each patch.
- Requires learning a consistency condition between adjacent patches
- More than 100X speedup in time to solution

Generate solutions on Enforce locally

Divide into finer voxel each finer voxel consistent solutions
Voxelize

Each voxel can be

8*8*8 point
voxels or more
Reiterate until convergence

18
Augmented Simulation: Hybrid Physics AI Digital Twin Architecture and Workflow

Software, Data-Based
Test Data
Control Models ROM
Twin Model for
Offline Compute
System Model Hybrid
Physics-AI
+ ROM
+ Solver
Modeling
Data Integration,
Auto-Calibration, …
Twin Model for
Cloud/Edge Deployment
LTI & LPV Static Dynamic
System Libraries ROM Extraction Capabilities

Twin Builder

IoT Gateway FMU for Simulation

Workflows

Validated 3D Physics Models Asset Performance Data

19
 Customer Productivity : Using AI/ML to Improve Convergence
• Problem:
- Solution convergence takes a lot of time and applicable to all solvers
- Can we learn from the history of convergence to find the ideal control parameters?
• Solution:
- We applied Bayesian optimization to a linear solver such as AMG
- Results shown below

Test case 1 @ 7200 sec:

• Default parameters : ~190 time steps
• Best parameters : ~295 time steps
Test case 1 Test case 2 Test case 3

Default parameters 11280 (sec) 11934 29323

Learned Parameters 8260 9658 23416

-27% -19% -20%

Engineering Design: Machine Learning-Driven Generative Design

• Embeddings: Prior designs are parameterized Design Corpus

Learning the Learning the Design Learning the Performance
• Future design style changes embodied in parametric design space
Designs
• Variant designs are constructed from the parametrized model
Styling CFD Models Tests
• Using Fusion Modeling technique, prior CAE and test data are fused to fast Variants
Learning the and accurate surrogate model STL Files
Performance Low
fidelity
High
fidelity
Wind
Tunnel
• Surrogate models are used for rapid evaluation of design variants Geometry Parameters
extraction using ML
CL, CD, Cm, Vel, a
(csv format)
Parameter Fusion model for Performance
Verification (Lift, Drag, Moment)

ANSYS proprietary
parameterization algorithm

Model executable to Model executable to predict

generate geometry variants aerodynamic characteristics
Design Exploration Predict
New aerodynamic
Base performance
Design
Generate Design
styling Verify for engineering
variants Review performance requirements

Quality & Engineering performance Delivered through ANSYS Minerva

21
 Summary
• Engineering simulation solves laws of physics in our world
- Physics captured by equations (Navier Stokes’ equations for fluids, Maxwell’s equations for electro-magnetics)
- Equations solved using numerical methods such as finite element analysis and finite difference
• Artificial Intelligence has progressed from Expert Systems to Machine Learning to Deep Learning
• Discussed the use of AI/ML to improve engineering simulation around four use cases:
- Improving Customer Productivity
- Augmented Simulation
- Revolutionizing engineering design
- Business Intelligence.
• Highlighted six AI/ML applications in engineering simulation
- Data-driven thermal solver
- Physics informed neural network for simulation
- Machine learning based PDE solvers
- Hybrid AI and Physics based Digital Twins
- Using AI/ML to Improve Convergence
- ML based Generative networks

22