Scientific machine learning is a core component of artificial intelligence and a computational technology that can be trained, with scientific data, to augment or automate human skills.
Major research advances will be enabled by harnessing DOE investments in massive amounts of scientific data,
credit:
software for predictive models and algorithms, high-performance computing (HPC) and networking platforms, and the national workforce.
The crosscutting nature of machine learning and AI provides strong motivation for formulating a prioritized research agenda.
Scientific Machine Learning and Artificial Intelligence (Scientific AI/ML) will have broad use and transformative effects across the research communities supported by DOE.
Accordingly, a 2019 Basic Research Needs workshop report identified six Priority Research Directions.
The first three PRDs describe foundational research themes that are common to the development of Scientific AI/ML methods and correspond to the need for domain-awareness, interpretability, and robustness.
The other three PRDs describe capability research themes and correspond to the three major use cases of massive scientific data analysis (PRD #4), machine learning-enhanced 4 modeling and simulation (PRD #5), and intelligent automation and decision-support for complex systems (PRD #6).
The principal focus of this FOA is on Scientific AI/ML for modeling and simulations (PRD #5).
Foundational research (PRDs #1, 2, and 3) will be needed for strengthening the mathematical and statistical basis in developing predictive AI/ML-based computational models and adaptive algorithms for scientific advances.
Also, new techniques, software tools, and approaches will likely be needed to reap scientific benefits from the extreme heterogeneity of scientific computing technologies (e.g, processors, memory and interconnect systems, sensors) that are emerging.
Scientific computing within DOE traditionally has been dominated by complex, resourceintensive numerical simulations.
However, the rise of data-driven Scientific AI/ML models and algorithms provides exciting opportunities.
Traditional forward simulations in scientific computing often are referred to as “inner-loop” modeling.
The combination of traditional scientific computing knowledge with ML-based adaptivity and acceleration has the potential to increase the performance and throughput of inner-loop modeling.
Additionally, ML-based algorithms must be scalable and efficient in order to handle massive amounts of data using high performance computing resources.
The scientific computing community has decades of expertise involving numerical analysis and algorithm development that can benefit the inner loop of training in ML.
Therefore, an opportunity exists to advance high-performance ML by entraining more involvement from the computational mathematics community.