Title: Data Science for Expensive Simulations: Toward Trustworthy Digital Twins at Exascale

 

Abstract:  Digital twins—live feedback loops between models and observations—promise to transform scientific discovery and engineering design. But foundational data-science challenges remain: How do we calibrate expensive high-fidelity simulations at exascale when each run costs days to weeks? How do we certify when a twin is trustworthy enough to drive critical decisions? How do we design robust systems when models are inevitably incomplete?

This talk presents two case studies for scalable data-science methods to address these questions. We combine physics-informed dimensionality reduction, multi-fidelity surrogates, and adaptive sampling to turn expensive simulations into actionable science. We demonstrate this framework on two high-consequence applications: exascale model calibration, where surrogate-accelerated multi-objective optimization revealed a hidden structural tension in E3SM; and fusion reactor design, where multi-fidelity uncertainty quantification workflows reduced in design exploration cost by orders of magnitude while improving safety-critical margins and performance metrics.

We conclude by outlining a vision for building a data-science program that trains the next generation of scientists to extract trustworthy, interpretable science toward digital twins at scale, turning expensive simulations into robust, evidence-based decisions.