About this Abstract |
Meeting |
2025 TMS Annual Meeting & Exhibition
|
Symposium
|
Algorithms Development in Materials Science and Engineering
|
Presentation Title |
Multi-Fidelity Models for Time-Dependent Full-Field Predictions |
Author(s) |
Aditya Venkatraman, Ryan Katona, David Montes de Oca Zapiain, Philip Noell |
On-Site Speaker (Planned) |
Aditya Venkatraman |
Abstract Scope |
Predicting full-field solutions of materials degradation phenomena is highly challenging due to the computational expense of high-fidelity, physics-based numerical simulations. We present a multi-fidelity framework to accelerate time-dependent, full-field predictions using a hierarchy of complex physics-based models. Initially, we create a reduced-order representation of the governing equations through physics-informed autoencoders, defining the latent dynamics of the physics-based model. These latent dynamics are characterized using Sparse Identification of Nonlinear Dynamics (SINDy), with parameterization obtained via a Gaussian Process (GP) surrogate model. By operating on the latent space GPs, we develop a hierarchy of latent dynamics at various fidelities, allowing for seamless uncertainty propagation across different fidelities and timeframes. This framework provides rapid and accurate full-field predictions of metal corrosion, validated against five increasingly complex physics-based models. Although developed for corrosion, this framework is applicable to a broad range of materials degradation phenomena. SAND2024-07843A |
Proceedings Inclusion? |
Planned: |
Keywords |
Computational Materials Science & Engineering, Machine Learning, Other |