IBM-Illinois Discovery Accelerator Institute

Abstract

Inspired by the remarkable success of machine learning in fields such as computer vision and language modeling, the scientific community is increasingly seeking to harness its power to address grand societal challenges. Yet scientific discovery demands more than recognizing patterns in observed data. Models must extrapolate reliably and remain consistent with established scientific principles. Purely empirical “black box” approaches often struggle to meet these requirements in complex scientific systems.

This talk introduces Knowledge-Guided Machine Learning (KGML), a class of methods in which scientific knowledge is embedded directly into model architectures, objective functions, and training procedures. By integrating data-driven learning with governing principles, KGML introduces explicit inductive bias that complements purely empirical models. It is particularly well suited for problems traditionally addressed using mechanistic or first-principles formulations, whose predictive performance may be limited by incomplete knowledge or structural simplifications. By combining the strengths of theory and data, KGML enables models that are both expressive and scientifically coherent. Although illustrated through applications in ecology, hydrology, agronomy, and climate science, the paradigm applies broadly wherever domain knowledge is available to inform model design.

KGML becomes even more critical as generative AI systems are applied to scientific discovery, where models rely heavily on scale to learn complex statistical patterns. But scale alone does not ensure principled extrapolation. Embedding scientific knowledge into generative models provides the inductive bias necessary for reliable out-of-distribution generalization, while reducing reliance on brute-force scaling and its associated computational and energy demands.

More broadly, the effort to make AI robust and reliable for scientific discovery signals a broader paradigm shift in artificial intelligence. In this sense, advancing AI for science may also help redefine the foundations of AI itself.

Biography

Vipin Kumar is a Regents Professor and William Norris Land Grant Chair in the Department of Computer Science and Engineering at the University of Minnesota. He is internationally recognized for foundational contributions to artificial intelligence, data science, and high-performance computing, and for advancing AI-driven scientific discovery.

His early work on scalable parallel AI search algorithms introduced the isoefficiency framework for analyzing scalability and laid foundations for later advances in graph partitioning, including the widely used METIS family of software that underpins large-scale engineering simulations and multi-physics applications worldwide. From 1998 to 2005, he served as Director of the Army High Performance Computing Research Center, then the DoD’s largest extramural HPC research program, leading interdisciplinary research in scalable scientific computing.

Kumar was among the first computer scientists to apply modern machine learning to planetary-scale Earth science. He led the NSF Expeditions in Computing project “Understanding Climate Change: A Data-Driven Approach,” which advanced data-driven methods as a complement to traditional climate modeling and produced globally adopted Earth system data products. He later developed Knowledge-Guided Machine Learning (KGML), a framework that integrates scientific principles into AI models to improve generalization, robustness, and scientific consistency. 

Kumar is a Fellow of AAAI, AAAS, ACM, IEEE, and SIAM. His honors include the ACM SIGKDD Innovation Award (2012), the IEEE Computer Society Sidney Fernbach Award (2016), and the IEEE Taylor L. Booth Education Award (2025), reflecting his leadership in data science, high-performance computing, and computing education. In 2026, he was appointed to the United Nations’ Independent International Scientific Panel on Artificial Intelligence, the first global scientific body established by the UN General Assembly to provide independent, evidence-based assessments of AI.