Modeling and Simulation Tools for Industrial and Societal Research Applications:
Digital Twins, and Genome-based Machine-learning
Abstract:
A variety of seemingly disparate physical processes can be treated with similar modeling and simulation tools. In this talk, I discuss the machine-learning enabled modeling and rapid digital-twin simulation of
Part 1: advanced manufacturing-3D printing, robotics and laser-processing
Part 2: complex energy management systems next-generation solar farms and data centers
Part 3: fire propagation, ignition and ember flow
Part 4: multiple unmanned aerial vehicle control for complex tasks
Part 5: pandemics, transmission, decontamination
Part 6: protective armor systems-impact, ballistics and lightweight materials,
as well as aspects of genomic/evolutionary computing for system optimization, utilizing multiphysics paradigms. The tools range from discrete element methods, computational optics, voxel-based computation to agent-based modeling-all connected together via machine-learning algorithms. For more information see https://cmmrl.berkeley.edu/zohdi-publications/ and http://www.me.berkeley.edu/people/faculty/tarek-i-zohdi.
Bio:
Tarek I. Zohdi http://www.me.berkeley.edu/people/faculty/tarek-i-zohdi/ received his Ph.D. in 1997 in Computational and Applied Mathematics from the University of Texas at Austin. He was a post-doctoral fellow at the Technical University of Darmstadt in Germany from 1997 to 1998 and then a lecturer (C2-Oberingenieur) at the Gottfried Leibniz University of Hannover in Germany from 1998 to 2001, where he received his Habilitation in General Mechanics (Allgemeine Mechanik). Approximately one out of every twenty doctoral degree holders in Germany is allowed to proceed with a Habilitation. It is the highest academic degree in Germany and is usually required to obtain the rank of full Professor there and in other parts of Europe. In July 2001, he became an Assistant Professor at the University of California, Berkeley, in the Department of Mechanical Engineering. He was promoted to Associate Professor in July 2004 and to Full Professor in July 2009. He has held a number of administrative posts at UC Berkeley
Summary:
Team simulates many processes, develops custom simulation codes for each
Initial motivation:
3D laser printer
Open loop: no sensors on the laser itself
Huge thermal gradient on the object being cut with the laser
Laser cutting followed by milling and drilling using additional tool
Managing this thing requires complex multi-physics, multi-scale simulation
Coupled simulation:
Maxwell solver of the energy deposition
Fluid flow of the melted material
Heat conduction
Many unknown parameters of the robot position, environment and material being cut
Domains: (they’ve implemented these themselves)
3D Printing robots
Complex mixtures
Lasers
Solid processing
Next-gen drones and multi-aerial vehicles/swarms
Pandemic/COVID
Irradiation of surfaces to kill pathogens
Contact tracing (agent-based)
Explosions and inverse forensics of explosions
Power grid
Ballistics
Data Centers
Solar farms
Vaccine design
Complex fires
Optimization framework:
Digital twin model of the device
Control loop
Measures state of system
Computes error between model and sensor measurements
Adjusts model state
Take next best action based on current model state and goal for device
Emphasis is on the optimization:
Describe system using a “genomic” representation
Use genetic algorithms to tune configuration
Optimize by trying many random genomes (random strings)
Mutation, crossover operation to create next generation of genomes
Techniques:
Strong gene breeding: mate strong candidates
Weak gene elimination: remove poorly performing strings
Adaptively re-focus search at better options
Good for non-convex, non-smooth optimization spaces
Hierarchical search:
Have hierarchy of models with different performance/accuracy tradeoffs
Constrain parameter space using man runs of cheap model
Estimate model error to know whether we’re likely off
Given regions that look promising and/or where error is high, run more expensive models
Hierarchical
Novel approach: replace mesh-based codes with voxel-based codes for continuous systems
No need for meshing
Easy interface for implementing ODE/PDE logic
Simple way to change cost/precision tradeoff of the code
Agent-based models is a good choice for discrete systems
Big picture: combine digital twins with ML