Grant: Discovery Grants Program - Individual
University of Toronto
$59,000.00
Robust Neural Simulation on Real-World Geometry
FY 2024-25
Summary
Machine learning and data--driven methods have proven to be powerhouses in fields such as Computer Vision wherein they have exhibited the ability to generate a myriad of static and dynamic results quickly and accurately. However, for complex large--scale mechanics, these methods have had little enduring success. This is due to several factors. First, there is a critical lack of available data for such applications. Second, non-differentiable algorithmic components within simulations complicate optimization and training of neural representations. And finally, lack of simulation-specific neural architectures limits generalization and computational performance.
There are two major difficulties in creating data for data-driven simulation methods. First is the cumbersome, but necessary, conversion from a geometric representation to a simulatable representation. For instance, volumetric objects are often represented geometrically, by just a surface, but a simulation will require a volumetric discretization to correctly produce output. Converting between these representations can be time-consuming and unreliable. This conversion process can take hours for particularly complicated examples. Worse still, for some types of surface data, reliable and accurate conversions to a volumetric representation do not exist. I will develop algorithms that are robust and geometry agnostic -- that can act directly on any surface or implicit representation, including surface representations that come from noisy input sources such as 3D scanners or lidar.
The second difficulty with leveraging machine learning approaches in simulation algorithms is that many operations inside of a simulator are not differentiable. Fundamental operations like collision detection and resolution are non-smooth, leading to theoretical and technical hurdles. In many other areas, these problems are overcome by relying on soft functions (softmax, softmin etc.) that provide smooth, differentiable approximations. All suffer from one or more limitations in terms of performance, not being geometry agnostic, or being inaccurate. I will explore and develop new soft function approximations for these problems to address these issues.
I will leverage the new technologies above to create new functional representations for learning-based simulation of complex three-dimensional geometries. The promise of applying machine learning to mechanical simulation is often hampered by representations that are (1) slow to train, (2) overfit to a single geometry/material combination, and (3) computationally inefficient to evaluate. I will explore the use of locally interpolated weights to drive deep neural networks. This not only improves generalization but alleviates the need to train the neural model for new examples -- conserving computational power and energy. This approach will both enable a game changing shift in simulation performance and accuracy.
Program
Discovery Grants Program - Individual
Recipient
Location
Ontario, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data
As published
Application id
RGPIN-2023-05120
Name nom
Levin, David
Department département
Computer Science
Organization id
31
Institution établissement
University of Toronto
Province en
Ontario
Province fr
Ontario
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2024
Competition year année de concours
2023
Award amount
59000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
801
Area of application group en
Information and communication services
Area of application group fr
Services d'information et de télécommunications
Area of application en
Computer software
Area of application fr
Logiciel
Research subject code
2700
Research subject group en
Information technology
Research subject group fr
Technologie de l'information
Research subject en
Information technology
Research subject fr
Technologie de l'information
Field of research list names en
Not available
Field of research list names fr
Non disponible
Socioeconomic objective list names en
Not available
Socioeconomic objective list names fr
Non disponible
Application title
Robust Neural Simulation on Real-World Geometry
Keywords
physics simulation; deformation; contact; differentiable programming; neural networks
Application summary
Machine learning and data--driven methods have proven to be powerhouses in fields such as Computer Vision wherein they have exhibited the ability to generate a myriad of static and dynamic results quickly and accurately. However, for complex large--scale mechanics, these methods have had little enduring success. This is due to several factors. First, there is a critical lack of available data for such applications. Second, non-differentiable algorithmic components within simulations complicate optimization and training of neural representations. And finally, lack of simulation-specific neural architectures limits generalization and computational performance.
There are two major difficulties in creating data for data-driven simulation methods. First is the cumbersome, but necessary, conversion from a geometric representation to a simulatable representation. For instance, volumetric objects are often represented geometrically, by just a surface, but a simulation will require a volumetric discretization to correctly produce output. Converting between these representations can be time-consuming and unreliable. This conversion process can take hours for particularly complicated examples. Worse still, for some types of surface data, reliable and accurate conversions to a volumetric representation do not exist. I will develop algorithms that are robust and geometry agnostic -- that can act directly on any surface or implicit representation, including surface representations that come from noisy input sources such as 3D scanners or lidar.
The second difficulty with leveraging machine learning approaches in simulation algorithms is that many operations inside of a simulator are not differentiable. Fundamental operations like collision detection and resolution are non-smooth, leading to theoretical and technical hurdles. In many other areas, these problems are overcome by relying on soft functions (softmax, softmin etc.) that provide smooth, differentiable approximations. All suffer from one or more limitations in terms of performance, not being geometry agnostic, or being inaccurate. I will explore and develop new soft function approximations for these problems to address these issues.
I will leverage the new technologies above to create new functional representations for learning-based simulation of complex three-dimensional geometries. The promise of applying machine learning to mechanical simulation is often hampered by representations that are (1) slow to train, (2) overfit to a single geometry/material combination, and (3) computationally inefficient to evaluate. I will explore the use of locally interpolated weights to drive deep neural networks. This not only improves generalization but alleviates the need to train the neural model for new examples -- conserving computational power and energy. This approach will both enable a game changing shift in simulation performance and accuracy.