About this Abstract |
| Meeting |
MS&T24: Materials Science & Technology
|
| Symposium
|
Machine Learning and Simulations
|
| Presentation Title |
Multi-Fidelity Gaussian Process Models for Time-Series Outputs |
| Author(s) |
Aditya Venkatraman, Ryan Michael Katona, David Montes de Oca Zapiain, Philip Noell |
| On-Site Speaker (Planned) |
Aditya Venkatraman |
| Abstract Scope |
We present a multi-fidelity Gaussian Process Regression method to emulate time-series outputs from a hierarchy of complex physics-based models. Our aim is to create an accurate emulator of the highest-fidelity model while minimizing computational costs, especially for the highest fidelity model. Initially, we establish a GPR model to emulate responses from the lowest-fidelity model. The predicted low-fidelity responses are treated as additional input features for higher-fidelity models. Taylor-series approximations, obtained through automatic differentiation, are used to propagate predictions and uncertainties across hierarchies and time. We employ Active Learning techniques using predictive standard deviation to enhance accuracy while minimizing high-fidelity simulation runs. We demonstrate the efficacy of our framework using a hierarchy of five electrochemical models for corrosion, by accurately predicting the cathodic current evolution of corroding galvanic couples. This versatile formulation shows promise for Bayesian calibration, surrogate-based optimization, and model fusion in diverse engineering disciplines. |