Birds of a Feather Archive

Physics-Informed Machine Learning Meets HPC – SOTA and Challenges for Widespread Adoption

Authors: Felipe Viana (University of Central Florida), Srinivas Tadepalli (Amazon Web Services), Matteo Corbetta (KBR at NASA Ames Research Center), Less Wright (Meta), Geeta Chauhan (Meta), Ankur Srivastava (Amazon Web Services)

Abstract: High-fidelity simulations are increasingly important in the design of complex systems. However, the computational cost of such models hinders their use for design space exploration, optimization, and uncertainty quantification. Alternative approaches, such as projection-based methods, often exhibit limited accuracy and call for collecting simulations at several data points, which is expensive in the first place. Recently, however, research institutions and industry have been collaborating to develop physics-informed neural network frameworks for simulations. This BoF seeks input from the machine learning and HPC communities as well as open participation in the development of useful tools to meet their needs.

Long Description: “Physics-informed machine learning meets HPC - SOTA and challenges for widespread adoption” will be the first of a Bird of a Feather (BoF) series. This BoF seeks to (1) elucidate the current state of the art in the use of neural networks as solvers of partial differential equations; (2) highlight the current experiences in implementation of research and industry-grade software; and (3) enumerate challenges and future research opportunities that will contribute to the adoption of this emerging technology to solve large scale real world problems leveraging accelerated computing platforms (GPUs, TPUs, CPUs). Ideally, the BoF will help the community of researchers and practitioners to start answering the questions: a) what is the target scalability of physics-informed neural networks? and b) given current capability, what does the community need to prioritize to achieve this target? The high computational cost associated with complex simulations of nonlinear and multiphysics systems, such as metamaterial devices and hypersonic aircraft panels, makes their design a daunting task. For example, a high fidelity transient aerothermoelastic simulation of a single hypersonic vehicle trajectory can require from weeks to months, even on recent high-performance computing clusters. As growth in computational power is often used to increase model fidelity, engineers and scientists are in constant need for novel approaches that efficiently reduce the cost of complex simulations without losing accuracy; and therefore, enabling design space exploration, optimization, and uncertainty quantification. One popular approach is to use fast approximations based on either reduced-order models (e.g., equivalent forces or approximate material models) or on surrogate models (e.g., using proper orthogonal decomposition). Unfortunately, the speed gains come at the expense of limited accuracy, preventing exploration of large design spaces. ​​Alternatively, there has been a growing interest in physics-informed neural networks, where neural networks and stochastic gradient descent algorithms are used to solve the governing partial differential equations that describe complex engineering systems. The key promise of such an approach is the scalability that can be achieved through modern ML frameworks (such as Pytorch, TensorFlow, Jax, and others) and the use of distributed and cloud computing, leveraging accelerators. Besides AWS, UCF, KBR@NASA Ames and Meta, key players expected to participate in this BoF include leading academic institutions, research laboratories, and industry from segments ranging from computing platform/solutions to sectors such as energy, aviation, healthcare, transportation, and many others. Participants will demonstrate what work is currently being conducted at their centers to set the context for further discussion.

URL:


Back to Birds of a Feather Archive Listing