Sequential Decision-Making for Mesh Refinement and Model Selection in Multiscale, Multi-Physics Applications

PI: Rama Tipireddy
Intelligent automation and decision support are needed to enhance computational efficiency and robustness in multiscale and multi-physics problems, including materials science, manufacturing, and climate and weather modeling. Current scientific computing approaches for enabling decisions by scientists fail to explore the role of learning, reasoning, and probabilistic planning. Decisions about the performance of the computation are usually made prior to run time and do not allow for real time changes to the computation. Current state-of-the-art scientific computing methods consist of rule-based algorithms that cannot automatically adapt to a dynamically changing computing environment. As a result, changes to the computation increases the total computing time and require a human expert to frequently monitor the results.
Although ML methods have proven to be beneficial in many fields, such as robotics, optimal control, etc., these techniques have not been fully leveraged for scientific computing applications. Development of a sequential decision-making framework will automate scientific computing by optimizing the policies for mesh refinement, time-stepping, model and algorithm selection, resource allocation, and pre- and post-processing. Our multi-agent sequential decision-making framework for scientific computing will consist of a data-driven learning approach for evaluating the system’s state, automated reasoning with physics-based contextual knowledge, and probabilistic planning using multi-agent reinforcement learning. In this approach, different scientific computing agents will work in a cooperative manner to achieve a common goal of obtaining efficient, accurate, and stable numerical solutions. Each agent will have a specific role, such as adaptively choosing the optimal time step, adaptive mesh refinement, model selection, or optimal resource allocation.
The model-based and model-free multi-agent reinforcement learning algorithms developed in this project will enhance the existing numerical methods for their numerical stability, accuracy, and computational efficiency. Algorithms developed in this project can be easily adapted to efficiently compute the numerical solutions to mathematical models of systems with relevance to U.S. Department of Energy missions.