Verification and Validation for Trustworthy Scientific Machine Learning

Presentation at the "UQ and Trustworthy AI Algorithms for Complex Systems and Social Good" workshop, Institute for Mathematical and Statistical Innovation (IMSI), University of Chicago, 3–7 March, 2025.

Abstract

This talk introduces a framework for establishing trustworthiness in Scientific Machine Learning (SciML) models through rigorous verification and validation practices. While SciML increasingly permeates computational modeling across many domains, the development of good modeling practices has lagged behind its application, creating a critical trust gap. We developed a framework with recommendations in four areas: Problem Definition, Verification, Validation, and Continuous Credibility Building. The 16 specific recommendations adapt established computational science standards while addressing the unique challenges of machine learning models. An illustrative example of ice-sheet modeling helps demonstrate how these practices can be applied to enhance model reliability, transparency, and alignment with scientific objectives. This work aims to catalyze community dialogue toward consensus-based standards for predictive SciML, providing researchers with practical guidance for developing models that can be trusted for high-consequence scientific and engineering applications.