DSI Director Kyle Cranmer’s work is highlighted in a recent report from the President’s Council of Advisors on Science and Technology (PCAST) that provides recommendations about responsibly harnessing the power of AI to accelerate scientific discovery in the U.S.
Kyle’s work on simulation-based inference is referenced in a section of the report titled Revealing the Fundamental Physics of the Universe. Specifically, the report cites a 2019 paper — The Frontier of Simulation-Based Inference — authored by Kyle, Johann Brehmer, and Gilles Loupe that was published in the Proceedings of the National Academy of Sciences in 2019. Simulation-based inference harnesses the power of AI and machine learning to transform scientific practice, and the PCAST report highlights the potential for AI models to discover new laws of physics and understand the origins of our universe.
In many scientific domains, researchers create digital simulations of complex phenomena such as airflow around an airplane, infectious disease spreading through a population, or the evolution of the universe. One type of simulation is a digital simulator—a virtual representation of a real-world object or system that is sometimes called a “digital twin”. Digital twins can help scientists understand how physical systems currently function and how they might perform under different conditions. These simulators encapsulate large amounts of expert knowledge, and they often bring together teams of researchers with different types of expertise. For example, one expert might model the extreme environment near the center of a galaxy, a second expert would study how light propagates through the vast expanse of space, a third would be responsible for understanding how the light goes through the atmosphere, and a fourth would provide expertise on a telescope’s optics. The simulator can pull together this varied expertise into a coherent whole, helping scientists relate the observations to the phenomena being studied.
“You can also think of a simulator as a computational version of a hypothesis,” says Kyle. “Scientists might have two competing theories or hypotheses, and they can simulate what would happen for each of these theories.”
Testing hypotheses is a core part of the scientific method. Unfortunately, it’s not easy to test hypotheses when they are expressed as a simulator. It’s ironic, because these digital twins are detailed, accurate representations of complex, real-world phenomena. But that same complexity is what makes this work difficult.
Simulators are good at predicting what observations might look like under a single theory. Inferring which theories are preferred, based on observations, is much harder. In a complex simulation, there are many paths that could lead to the same observation, and inference requires looking at them all. Simulation-based inference uses machine learning and AI to address this complexity. AI/ML is very good at finding patterns in data, and researchers can use simulators to generate virtually unlimited synthetic data. Kyle and his colleagues figured out some clever ways to use AI/ML and the simulators together to make robust statistical statements about hypotheses.
The initial work on simulation-based inference focused on particle physics problems, but it has quickly been adopted in many other fields. It’s now popular in astrophysics and cosmology, and it is being applied to a wide range of topics including neuroscience, biology, genomics, epidemiology, genetics, social science, economics, and finance.
The report cites the potential of simulators to not only transform science, but also guide decision making:
Beyond advances in core science and engineering disciplines, AI methods promise to provide high fidelity models—“digital twins”—of the world that can help us to cut through uncertainty and complexity to predict, to plan, and to guide policymaking, where scarce data and models currently make it difficult to assess potential pathways forward.
Another part of the report addresses automated workflows, which may incorporate AI components:
Virtually every aspect of the laboratory workflow, from experimental design to data collection to data interpretation, could be partially or fully automated through AI assistance, although we view expert human supervision of such automated laboratories to be essential and highly desirable for decades to come.
This recommendation references a 2022 report from the National Academies of Sciences, Engineering and Medicine. The work done by Kyle and his team on simulation-based inference, and their work on automated workflows with AI components for particle physics, is cited in this report.
To view a full copy of the PCAST report, click here.
To view PCAST’s letter to the President and the executive summary of the report, click here.