Collaborative Research: OAC Core: Smart Surrogates for High Performance Scientific Simulations

点击翻译按钮获取中文摘要

High-fidelity computer simulations underpin discovery in a broad range of scientific domains. However, their computation cost limits their full potential. There have been increasing efforts in approximating scientific simulations with deep neural networks, to accelerate simulation workflows by orders of magnitude. Current practice, however, largely relies on fixed network architectures and offline simulation data -– predefined by experience, rather than optimized by quantitative metrics. This leads to an empirical, subjective, and laborious practice, yet with a suboptimal outcome. This research addresses the above critical gaps with a new conceptual, mathematical, and infrastructure framework for developing Smart Surrogates. As a domain-agnostic framework, Smart Surrogates will deliver timely support for an increasing but yet-to-be-met demand for surrogate modeling for scientific simulations. The prototype surrogates created in this project will also directly enable long-term follow-on research in each of the domains involved. This collaborative research provides multidisciplinary training at the intersection of artificial intelligence, high-performance computing, and scientific simulations in a variety of domains, helping prepare next-generation researchers adept at transdisciplinary thinking and skill. It plans to proactively recruit students from underrepresented groups, and develop a hands-on workshop on Smart Surrogates for dissemination to a broader student body. Finally, the dissemination of ROSE as an open-source toolkit will impact HPC simulation workflows in a broad range of social applications, including but not limited to drug design and the study of climate change.The development of Smart Surrogates includes three parallel but interwoven methodological, infrastructure, and domain evaluation thrusts: 1) Thrust I – Methodological Innovations: This thrust develops fundamental innovations in deep active learning to jointly optimizes training-data selection and neural architectures, in a Bayesian setting equipped with uncertainty quantification. This allows Smart Surrogates to support the intelligent active selection of training simulations along with dynamic adjustment of neural architectures; 2) Thrust II – Infrastructure innovations): This thrust designs, implements, and disseminates the RADICAL Optimal & Smart-Surrogate Explorer (ROSE) toolkit to support the concurrent and adaptive executions of simulation and surrogate training and selection tasks.; 3) Thrust III – Scientific innovations: This thrust grounds the developments and evaluation of Smart Surrogates in two domain problems: surrogates for 1) diffusion equations with singular initial conditions and 2) personalized virtual heart simulations, built on the team’s past works with established domain collaborators. This allows fast prototyping, while setting the basis for a continuum of follow-up research to adopt Smart Surrogates in a larger range of complex scientific simulations.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.