3rd Physics Informed Machine Learning

This workshop continues discussions and explorations started in January 2016 and January 2018 at the first and second editions of the workshop. A revolution in statistics and machine learning (ML) is underway. Modern algorithms can now learn high level abstractions via hierarchical models, leading to break-
through accuracy in benchmarks for computer vision, language, etc. Underlying these advances is a strong and deep connection to various aspects of applied mathematics and statistical physics. For example, proper choice of statistical force allows to screen interaction and learn graphical models governing multi-dimensional distributions eefficiently, gauge transformations from physics guide incorporation of symmetries in the neural network design, dynamical system interpretation helps to understand and improve performance of most eefficient deep learning schemes, etc.

This workshop seeks perspectives on leveraging the deep connection between ML and physics, but now with the goal to better understand and model physical systems, static and dynamic. We invite experts both in machine learning techniques as well as domain science applications. In this third edition of PIML we will focus on applications of interest to our LANL and DOE sponsors, specifically on

• improving scale-reduced, large eddy modeling and simulations of turbulence that arise in various mechanical-engineering, aerospace and climate
  applications;
• building reduced models for infrastructures (energy systems, trac
  ows, etc).
• guiding development of inverse/design approaches in quantum physics, e.g. related to tensor networks;
• designing new computational paradigms (such as related to quantum-, neuromorphic-, etc computers).

The workshop discussions are aimed towards approaches and methods for physical modeling applications where a big-data, black-box approach to ML is only a starting point. We seek participants who may suggest innovative approaches that extend application agnostic ML techniques by incorporating complex constraints imposed by physical principles (e.g. conservation laws, causality, symmetries, entropy principles and related).

The workshop format assumes 7 talks a day, late afternoon discussions and poster session. We encourage application of young researchers for fellowships (see web-site for details).

We plan for active participation of LANL researchers and program managers across directorates and divisions interested in the physics informed learning. This emerging area of research has many aspects of computational co-design, and draws on LANL's strengths in statistical physics, modern applied mathematics, theoretical and applied computer science, infrastructure modeling and
simulations, fluids and materials modeling, and high performance computing. Looking forward, we view physics informed learning as a viable path for LANL and DOE toward truly predictive multi-scale modeling, optimization, inference and leading for plethora of relevant applications in sciences and engineering.

Topics

• Methods (Monday and Tuesday morning)
  • Deep Learning - Theory and Methods
  • Graphical Models
  • Active and reinforcement learning
• Spatio-Temporal Flows: Dynamical Systems, Fluid Mechanics and Turbulence (Thursday & Friday Morning)
• Other Applications (Tuesday afternoon and Wednesday)
  • Cyber, Physical Infrastructures and Communication Networks
  • Novel Computational Paradigms (quantum, neuromorphic, memristors, etc.)
  • Physics (statistical mechanics, condensed matter, biophysics, etc.)