CAREER: Generative Machine Learning for Scientific Modeling and Discovery

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

Since the advent of human civilization, science has played a pivotal role in addressing societal challenges and provided us with the necessary knowledge, tools, and methodologies to drive progress. However, the process of scientific understanding and discovery can itself be quite slow as it requires meticulous observation and analysis of complex natural phenomena. This can be discouraging as we grapple with pressing sustainability challenges in areas such as climate and energy security that demand a swift response. This project will develop generative artificial intelligence (AI) frameworks to aid scientific reasoning by efficiently analyzing vast streams of scientific data to identify patterns, simulate natural phenomena, and suggest experiments. The resulting algorithms will be grounded in real-world applications aimed at accelerating sustainable development. The project will be complemented by educational and outreach activities aimed at diverse interdisciplinary audiences. The central goal of this project is to develop generative AI systems for scientific workloads that can learn with inexpensive supervision and exhibit strong generalization and reliability across broad domains, akin to foundation models for language and vision. To achieve this goal, the project will encompass innovations in data engineering, large-scale pretraining, and finetuning techniques to enhance reliability. On the data and modeling side, this research will lead to new generative architectures and objectives to fuse heterogeneous multi-modal and multi-scale datasets for scalable training of scientific foundation models. The project will further develop methodologies to efficiently finetune these models to novel scenarios using in-context learning, permit online updates with new experimental evidence, and align inference with relevant domain knowledge. In addition to fundamental advances, the project will be grounded in key scientific tasks spanning simulation, forecasting, and experimental design for climate and energy domains. Educational and outreach efforts will include the development of interdisciplinary courses, open-source software and interactive tutorials, as well as seminars and workshops in AI for Science targeting diverse groups of students, researchers, and practitioners. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.