There is a great interest in scientific communities for harnessing the power of AI in applications ranging from climate science to quantum chemistry. The common theme in many of these applications is that the data are spatiotemporal with governing physics. Unfortunately, today's ML approaches are mostly purely data-driven, i.e., they solely rely on (labeled) data for learning statistical patterns. Collecting labeled data can be quite expensive in real-world applications. Moreover, the resulting black-box AI models are difficult to interpret for domain scientists.
Many scientific applications contain valuable domain knowledge such as laws of physics or symmetry. On its own, black-box AI may ignore known physical laws, or spend tremendous training time only to re-discover them. This can lead to solutions that may violate physical principles or predictions that generalize poorly to unseen test scenarios. For example, energy conservation is well-understood in climate science but existing ML models predictions often fail to follow such a principle.
Physics-guided or physics-informed AI is an emerging area spanning several disciplines to principally integrate physics in AI models and algorithms. The goal of this tutorial is to (1) provide an overview of spatiotemporal data analysis and its central role in science (2) survey development in physics-guided AI and their connection to existing techniques in scientific fields, and (3) identify the benchmark datasets, open problems and future directions in physics-guided AI and its broader impact.
We believe this is a very timely and highly impactful topic. This tutorial will draw attention from the data science community to emerging applications in science. It will also bring new audiences from the scientific fields to data science. Currently, many techniques for analyzing scientific data have been developed in isolation. This tutorial aims to bridge the gap and facilitate cross-learning from different domains.
Rose Yu, Paris Perdikaris, Anuj Karpatne: Physics-Guided AI for Large-Scale Spatiotemporal Data. KDD 2021: 4088-4089
- Date of publication:
- August 14, 2021
- ACM SIGKDD International Conference on Knowledge Discovery and Data Mining
- Page number(s):