(CLOSED) Scientific Machine Learning for Complex Systems (DE-FOA-0002958)

The DOE SC program in Advanced Scientific Computing Research (ASCR) hereby announces its interest in research applications to explore potentially high-impact approaches in the development and use of scientific machine learning (SciML) and artificial intelligence (AI) in the predictive modeling, simulation and analysis of complex systems and processes.

Supplementary Information
High-performance computational models, simulations, algorithms, data from experiments and observations, and automation are being used to accelerate scientific discovery and innovation. Recent workshops, report, and strategic plans across the DOE have highlighted the research, development, and use of artificial intelligence and machine learning for science, energy, and security. Relevant domains include materials, environmental, and life sciences; high-energy, nuclear, and plasma physics; and the DOE Energy Earthshots Initiative, for examples. A 2018 Basic Research Needs workshop and report on scientific machine learning (SciML) and AI1 identified six Priority Research Directions (PRDs) for the development of the broad foundations and research capabilities needed to address such DOE mission priorities. The first three PRDs for foundational research are a set of themes common to all SciML approaches and correspond to the need for domain-awareness, interpretability, and robustness and scalability, respectively. Of the other three PRDs for capability research, PRD #5 (Machine Learning-Enhanced Modeling and Simulation) and uncertainty quantification are the subject of this FOA.

Research Opportunities

The focus of this funding opportunity is on basic research and development at the intersection of uncertainty quantification (UQ) and scientific machine learning (SciML) applied to the modeling and simulation of complex systems and processes. PRD #5 is the subject of this topic: Machine learning-enhanced modeling and simulation for predictive scientific computing. Scientific computing within the DOE traditionally has been dominated by complex, resource-intensive numerical simulations. However, the rise of data-driven SciML models and algorithms provides new opportunities. Traditional scientific computing forward simulations often are referred to as “inner loop” modeling. The combination of traditional scientific computing expertise and machine learning-based adaptivity and acceleration has the potential to increase the performance and throughput of inner-loop modeling. Such hybrid modeling and simulation approaches offer the opportunity, for example, to combine the versatility of neural networks for function and operator approximations, the domain-knowledge and interpretability of differential equations and operators, and the robustness of high-performance scientific computing software across these areas.

In the context of this FOA, UQ refers to the processes of quantifying uncertainties in a computed quantity of interest, with the goals of accounting for all sources of uncertainty and quantifying the contribution of specific sources to the overall uncertainty. For hybrid scientific machine learning modeling and simulations, the development and use of UQ will incorporate additional or other sources of uncertainties. Such considerations bring new basic research challenges in UQ beyond those encountered in traditional modeling and simulation approaches.