Scientific Machine Learning

Scientific machine learning bridges physics-based modeling with data-driven methods, enabling accurate predictions with quantified uncertainties for complex physical systems.

Nuclear Physics Applications: Developing ML methods to invert nuclear response functions, extracting fundamental physics from experimental measurements. Our uncertainty-quantification approaches published in Physical Review C provide rigorous error estimates critical for nuclear physics research at facilities like Argonne’s ATLAS.

Uncertainty Quantification: Neural architecture search for automated ensemble construction (AutoDEUQ), enabling deep learning models to provide calibrated uncertainty estimates for scientific predictions. This is essential for trustworthy AI in high-stakes scientific applications.

Particle Physics: Classification methods for analyzing alpha-induced reactions in the MUSIC detector at ATLAS facility, combining statistical and ML approaches for event identification in nuclear astrophysics experiments.

Computational Fluid Dynamics: Forward gradient methods for data-driven wall modeling in CFD simulations, bridging high-fidelity simulations with efficient surrogate models for turbulence prediction.

Optimization: Bilevel optimization frameworks for handling imbalanced scientific datasets common in rare event detection, and stochastic methods for generating space-filling experimental designs that maximize information from limited experimental budgets.

Publications