ACMD Seminar: Physics-Informed Machine Learning to Infer Dynamics from Data

Biswadip Dey
Siemens Corporate Technology

Thursday, May 21, 2:30-3:30 PM EDT
Online via BlueJeans. Contact acmdseminar [at] nist.gov (acmdseminar[at]nist[dot]gov) for details.

This seminar is a joint production with the AI/Materials WG and the AI COI.

Abstract: In recent years, neural networks and end-to-end representation learning have become very accurate and widely adopted in many application domains. To learn underlying patterns from data and enable generalization beyond the training set, the learning approach incorporates appropriate inductive bias by promoting representations which prefer one hypothesis over another. The success in predicting an outcome for previously unseen data then depends on how well the inductive bias aligns with the ground reality. Therefore, in a variety of applications involving physical systems, generalization in neural networks can be improved by leveraging the underlying physics.

In this talk, we will describe how a generalization of the Hamiltonian dynamics can aid us in developing a learning framework which can infer the dynamics of a physical system from observed time-series data. The design of the computation graph of Symplectic ODE-Net, the proposed learning framework, is informed by the underlying physics governed by Hamiltonian dynamics with external control. Our results show that incorporation of such physics-based inductive bias enables us to learn the underlying dynamics in a transparent, data-efficient way. In addition, we propose a parametrization which can enforce this Hamiltonian formalism even when the generalized coordinate data is embedded in a high-dimensional space or we can only access velocity data instead of generalized momentum. Furthermore, to accommodate energy-dissipation (such as mechanical friction, resistive loss etc.), which is often present in real world applications, we introduce Dissipative SymODEN a variant of the Symplectic ODE-Net framework. We will conclude the talk by showing how these frameworks open up new possibilities for synthesizing model-based control strategies in a data-driven setting.

Bio: I am a Research Scientist at Siemens Corporate Technology. My current research interests lie in the areas of machine learning and its application to physical systems, bio-inspired designs, nonlinear dynamics, multi-agent systems, and optimal control. Prior to joining Siemens, I was an Associate Research Scholar and Lecturer at Princeton University. While at Princeton, I studied nonlinear dynamics in complex, multi-agent systems (with Prof. Naomi E. Leonard) and bio-inspired cognitive architectures (with Prof. Jonathan Cohen). Earlier, in January 2015, I completed my PhD in Electrical Engineering from the University of Maryland, College Park. My dissertation on Reconstruction, Analysis and Synthesis of Collective Motion was supervised by Prof. P. S. Krishnaprasad. I received the M.Tech. degree in Systems & Control Engineering from the Indian Institute of Technology Bombay, India, in 2008, and the B.E. degree in Electrical Engineering from Jadavpur University, India, in 2006.