Physics-regularized neural networks for predictive modeling of silicon carbide swelling with limited experimental data

Kazuma Kobayashi; Syed Bahauddin Alam

doi:10.1038/s41598-024-78037-7

Physics-regularized neural networks for predictive modeling of silicon carbide swelling with limited experimental data

Sci Rep. 2024 Dec 28;14(1):30666. doi: 10.1038/s41598-024-78037-7.

Authors

Kazuma Kobayashi¹, Syed Bahauddin Alam^{2

3}

Affiliations

¹ Nuclear, Plasma & Radiological Engineering, University of Illinois Urbana-Champaign, Urbana, USA.
² Nuclear, Plasma & Radiological Engineering, University of Illinois Urbana-Champaign, Urbana, USA. [email protected].
³ National Center for Supercomputing Application, University of Illinois Urbana-Champaign, Urbana, USA. [email protected].

Abstract

This study introduces a physics-regularized neural network (PRNN) as a computational approach to predict silicon carbide's (SiC) swelling under irradiation, particularly at high temperatures. The PRNN model combines physics-based regularization with neural network methodologies to generalize the behavior of SiC, even in conditions beyond the traditional empirical model's valid range. This approach ensures continuity and accuracy in SiC behavior predictions in extreme environments. A key aspect of this research is using nested cross-validation to ensure robustness and generalizability. The PRNN model effectively bridges empirical and sparse experimental data by integrating prior knowledge and refined tuning procedures. It demonstrates its SiC's predictive power in high-irradiation conditions essential for nuclear and aerospace applications.