Enabling Predictive Scale-Bridging Simulations through Active Learning

Designing effective methods for multiscale simulation is a longstanding challenge. Our goal is to advance the state of the art for Machine Learning (ML) beyond sequential training and inference and facilitate scale bridging through novel techniques. Active Learning (AL) is a special case of semi-supervised ML in which a learning algorithm is able to interactively use the fine-scale model to obtain the desired outputs at new data points, making it ideal for concurrent scale-bridging. Our AL procedure will dynamically assess uncertainties of the ML model, query new fine scale simulations as necessary, and use the new data to incrementally improve our ML models. This capability will be demonstrated on two applications: transport in nanoporous media (e.g., for hydraulic fracturing) and inertial confinement fusion (ICF), validating against experimental data. Although the physics is quite dissimilar, both applications represent problems that suffer from inaccurate macro-scale predictions due to subscale physics that are ignored.

 

In this presentation, I will present recent, initial results from this effort, which is supported by a LDRD-DR at LANL, and will focus more on implementation of the AL framework for the ICF application. In particular, we focus on using this model to connect MD simulations with a kinetic model and a DNS hydro code to study interfacial mixing.

Research presented in this presentation was supported by the Laboratory Directed Research and Development program of Los Alamos National Laboratory under project number(s) 20190005DR. Los Alamos National Laboratory is operated by Triad National Security, LLC, for the National Nuclear Security Administration of U.S. Department of Energy (Contract No. 89233218CNA000001).