Explore a flexible and generalizable explainable AI (XAI) framework for scientific deep learning in this 53-minute talk. Delve into the challenges of interpretability and generalization in modeling physical systems, and discover an innovative approach inspired by human thinking processes. Learn how this framework probes trained neural networks, fits interpretable models using integral equations, and improves out-of-distribution (OOD) generalization. Examine applications in solid mechanics, fluid mechanics, and transport, showcasing the framework's potential for providing analytical representations of deep neural networks and enhancing OOD performance. Gain insights into theory-guided machine learning, functional data analysis, and the synergistic integration of interpretable models with black-box deep learning techniques.
DDPS - A Flexible and Generalizable XAI Framework for Scientific Deep Learning
Inside Livermore Lab via YouTube
Overview
Syllabus
Intro
Deep learning in physics-based systems
Scientific modeling in the age of Al
Lots of data BUT not enough!
Out-of-distribution (OOD) generalization
Interpretable and explainable Al (XAI)
Key idea: Integral equations!
Theory-guided machine learning
Functional data analysis (FDA)
Applications
Motivation example: 1D
Predicting strain energy from heterogeneous material property
Predicting velocity field from heterogeneous permeability
Summary: towards interpretable and generalizable deep learning
Acknowledgments
Example 3: Predicting high-fidelity wall shear stress (WSS) from low-fidelity velocity
Example 5: Local interpretation: Predicting velocity field from permeability fields
Appendix: Library
Generalized functional data analysis (gFDA)
Taught by
Inside Livermore Lab
