DOE Provides Overview of New AI for Science Report

Last week, the Department of Energy (DOE) Advanced Scientific Computing Advisory Committee (ASCAC) met to approve a new report from the DOE on the application of artificial intelligence (AI) and machine learning to science. The report was requested by the DOE last year to look at the outputs and analyze the opportunities and challenges for DOE’s Office of Advanced Scientific Computing Research (ASCR) and the DOE Office of Science associated with AI and machine learning. Specifically, the subcommittee was asked to identify stratifies that ASCR could use to address identified challenges.
 
The report from Cross-Cutting AI Subcommittee was presented by committee member Tony Hey, who spoke to work done to pursue the goals set by DOE last year, which came about as a result of the earlier executive order to pursue U.S. leadership and excellence in AI.
 
The first draft of the report was produced in June and feedback was incorporated and the report was reworked through August. The subcommittee held its last meeting on September 1 to discuss and agree on the final report. The structure of the report includes key findings and recommendations with specific sections throughout the report on AI applications, AI algorithms and foundations, AI software infrastructure, new hardware technologies for AI, instrument to edge computing, workforce training and retention, and collaboration with universities, industry, and inter-agency.
 
Hey also discussed why the report makes clear that DOE should be the lead agency on this initiative. He points out that National Science Foundation (NSF) has identified AI as cutting across its priorities and Big Ideas. He explains DOE should be the lead agency because it focuses on date sets and powerful computing for science. He also says that DOE should be lead in AI for Science because the agency is uniquely able to bring together individuals from many different areas, including software engineers and data scientists. In implanting AI for Science effectively, DOE expects to accelerate the design, discovery, and evaluation of raw materials, along with advancing the development of new hardware and software systems.
 
The AI Applications section of the report makes up the biggest section and includes priority research directions including:

• Basic energy sciences
• Biological and environmental sciences
  • Climate and environmental sciences
  • Biological system sciences
• Fusion energy sciences
• High energy sciences
• Nuclear physics 

Key findings include:

• The growing convergence of AI, data, and HPC provides a once in a generation opportunity to profoundly accelerate scientific discovery, create synergies across scientific areas, and improve international competitiveness.
• Science can greatly benefit from AU methods and tools. However, commercial solutions and existing algorithms are not sufficient to address the needs of science automation and science knowledge from current and future DOE facilities and data.
• Adoption AI for Science technologies throughout the Office of Science will enable US scientists to take advantage of the tremendous new advances in the DOE’s scientific user facilities.
• Realizing the potential for generational shift in scientific experimentation at DOE Laboratories due to science-driven AI/ML technologies requires far more than simply compute power and encompasses the full spectrum of computing infrastructures, ranging from ubiquitous sensors and interconnectivity across devices to real-time monitoring and data analytics, and will require a concentrated and coordinated R&D effort on AU/ML algorithms, tools, and software infrastructure.

Recommendations:

1. Creation of a 10-year AI for Science Initiative
2. Structure of SC AI for Science Initiative
3. An Instrument-to-Edge Initiative
4. Training, focusing, and retention of the AI/ML workforce
5. Inter-agency collaboration
6. International collaboration