Canada Spends Logo

Explorez notre analyse du budget 2026

Grant: Discovery Grants Program - Individual

University of Alberta

$30,000.00

Geometric mechanics, physics-informed neural networks and applications

EF 2023-24

Résumé
Machine learning (ML) and artificial intelligence methods have provided valuable insights into unstructured data. The applications of ML methods have become ubiquitous in the world around us. However, when the amount of data is insufficient to use the big-data approach, one has to rely on additional insights to make accurate predictions. Physics Informed Neural Networks (PINNs) have been developed to provide further insights into the systems. PINNs train the network to match the data to actual physical models describing the equations, such as Partial Differential Equations (PDEs) describing the systems and their boundary conditions. PINNs, in some sense, combine the recent progress in the power ML methods and the centuries of insights of `traditional' science. They may also allow for exceptional computational efficiency compared to the regular methods of numerical solutions of PDEs. Powerful as they are, the `basic' PINNs have difficulties describing systems with little or no friction, such as Hamiltonian systems. Recently, generalizations of PINNs were proposed to deal with the canonical systems using (SympNets) and their extensions for non-canonical brackets by reducing the systems into the canonical form. We build on these insights and develop an alternative approach based on the reduction of mechanical systems by continuous symmetries (Lie groups), as many systems are difficult to write and analyze in canonical variables. The symmetry-reduced systems only include the `essential' variables in the resulting equations of motion, coupled with the structure derived from the Lie group symmetries. The resulting equations are known as the Lie-Poisson equations. The idea is to build up the PINNs-like dynamics from elementary `motions', preserving the Lie-Poisson structure using machine learning from data. We call the resulting network LPNets. This approach allows for the exact conservation of some properties of the original system (Casimirs) and the preservation of other conserved quantities like energy with high accuracy. The preservation of equation structure and the resulting conservation laws allow for the construction of highly accurate, efficient, long-term systems predictions even starting from a limited data. We provide a unified geometric approach for implementing this approach for various important, practical problems, from satellite dynamics to fluids, further extending these methods for the interaction of fluids and elastic media, systems with multiple interacting parts involving friction, thermodynamics, and/or constraints. We also study how to discretize the problems of continuum mechanics to practically implement LPNets. We will use the insights from these variational methods to construct the building blocks of the Lie-Poisson versions of PINNs for each particular problem and subsequently implement these methods for fast, efficient, and accurate PINNs-based numerical solutions.
Programme
Discovery Grants Program - Individual
Bénéficiaire
Emplacement
Alberta, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data

Tel que publié

Application id
RGPIN-2023-03590
Name nom
Putkaradze, Vakhtang
Department département
Mathematical and Statistical Sciences
Organization id
9
Institution établissement
University of Alberta
Province en
Alberta
Province fr
Alberta
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
30000
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
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
1205
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Mathematical sciences
Area of application fr
Sciences mathématiques
Research subject code
2958
Research subject group en
Applied mathematics
Research subject group fr
Mathematical physics
Research subject en
Mathématiques appliquées
Research subject fr
Mathématiques de la physique
Application title
Geometric mechanics, physics-informed neural networks and applications
Keyword
Modeling choices; PINNs (Physics Inspired Neural Networks); Variational methods
Application summary
Machine learning (ML) and artificial intelligence methods have provided valuable insights into unstructured data. The applications of ML methods have become ubiquitous in the world around us. However, when the amount of data is insufficient to use the big-data approach, one has to rely on additional insights to make accurate predictions. Physics Informed Neural Networks (PINNs) have been developed to provide further insights into the systems. PINNs train the network to match the data to actual physical models describing the equations, such as Partial Differential Equations (PDEs) describing the systems and their boundary conditions. PINNs, in some sense, combine the recent progress in the power ML methods and the centuries of insights of `traditional' science. They may also allow for exceptional computational efficiency compared to the regular methods of numerical solutions of PDEs. Powerful as they are, the `basic' PINNs have difficulties describing systems with little or no friction, such as Hamiltonian systems. Recently, generalizations of PINNs were proposed to deal with the canonical systems using (SympNets) and their extensions for non-canonical brackets by reducing the systems into the canonical form. We build on these insights and develop an alternative approach based on the reduction of mechanical systems by continuous symmetries (Lie groups), as many systems are difficult to write and analyze in canonical variables. The symmetry-reduced systems only include the `essential' variables in the resulting equations of motion, coupled with the structure derived from the Lie group symmetries. The resulting equations are known as the Lie-Poisson equations. The idea is to build up the PINNs-like dynamics from elementary `motions', preserving the Lie-Poisson structure using machine learning from data. We call the resulting network LPNets. This approach allows for the exact conservation of some properties of the original system (Casimirs) and the preservation of other conserved quantities like energy with high accuracy. The preservation of equation structure and the resulting conservation laws allow for the construction of highly accurate, efficient, long-term systems predictions even starting from a limited data. We provide a unified geometric approach for implementing this approach for various important, practical problems, from satellite dynamics to fluids, further extending these methods for the interaction of fluids and elastic media, systems with multiple interacting parts involving friction, thermodynamics, and/or constraints. We also study how to discretize the problems of continuum mechanics to practically implement LPNets. We will use the insights from these variational methods to construct the building blocks of the Lie-Poisson versions of PINNs for each particular problem and subsequently implement these methods for fast, efficient, and accurate PINNs-based numerical solutions.