Upcoming Events

Tuesday, September 16

Sep

DSAI Center Seminar Series: Jacob Hochhalter, University of Utah

Conference/Seminar

Title: Balancing Complexity and Interpretability for High-dimensional SciML Models: Fatigue, Fracture, and Constitutive Modeling Abstract: There is no panacea in machine learning; the problem context must guide the choice of method. Neural networks (NNs) are highly expressive and train efficiently with large datasets, but they require substantial data to avoid overfitting and lack inherent interpretability. In contrast, symbolic regression (SR) is interpretable and effective with smaller datasets, common in science and engineering due to costly data acquisition, but lacks NN expressivity. This presentation examines tradeoffs between expressivity vs. interpretability and purely data-driven vs. generalizable approaches. Examples will be drawn from microstructure-dependent fatigue cracking, fracture mechanics, and materials constitutive modeling. For fatigue cracking, we highlight a hybrid SciML method balancing NN expressivity with interpretability. For fracture mechanics, we present mechanics-based decomposition to enhance SR expressivity. For constitutive modeling, we combine thermodynamic guidance with decomposition to reduce data needs while maintaining expressivity.

3:30 pm - 4:30 pm | Engineering Building | More Info

SUBMIT AN EVENT