About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Modeling, AI Applications, and Method Development in Nuclear Materials
|
| Presentation Title |
Actively Learned Physically Constrained Gaussian Process Regression for TRISO Particle in Particle Compact |
| Author(s) |
George John Joseph, Apurba Roy, Jacob Hirschhorn, Michael Tonks, Daniel Schwen, Michael Shields |
| On-Site Speaker (Planned) |
George John Joseph |
| Abstract Scope |
Accurately modeling fission-product release from TRISO-based fuel is essential for reactor safety analysis. Concurrent multi-scale coupling can capture this behavior by representing individual TRISO particles as lower-length-scale (LLS) models and the fuel compact as the higher-length-scale (HLS) model. In this framework, LLS models provide heat and fission-product fluxes to the HLS model, while the HLS model supplies particle surface temperatures to the LLS models. However, running LLS models for thousands of particles concurrently with the HLS model is computationally intractable. To reduce this cost, we develop an actively learned, physically constrained Gaussian process regression (GPR) surrogate. Inequality constraints enforce non-negative outward heat and fission-product fluxes to the compact, while GPR predictive uncertainties guide active learning and targeted retraining over the relevant operating range. This approach requires only a limited number of LLS model evaluations, substantially reducing computational cost while maintaining accurate flux predictions, demonstrated through coupled BISON runs. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Modeling and Simulation, Machine Learning, Other |