Multi-Fidelity Machine Learning Approach to Material Modeling for Digital Twin Framework

The material properties of structures are critical for developing digital twin technologies in the nuclear field. However, gathering sufficient experimental data for building a mathematical model that accounts for irradiation effects is challenging due to time and financial c onstraints. A s a result, there are often scenarios where only limited mathematical models and sparse experimental data are available in nuclear science and engineering. To address this issue, this study proposes the use of a multi-fidelity deep neural network (MFDNN) to combine a limited mathematical model and sparse experimental data and generate a new high-fidelity model. Unlike the traditional multi-fidelity method, which can only handle linear correlations between different fidelity data sets, the proposed method uses multiple deep neural networks to build a model that considers both linear and non-linear correlations. The effectiveness of MFDNN is verified b y m odeling radiation-induced volume swelling for chemical vapor deposition silicon carbide (CVD SiC) using two sets of data: (1) an empirical model with a limited valid temperature range and (2) sparse experimental data. As a result, new high-fidelity models were generated, and their valid temperature range was extended from 473 K to 1800 K. Moreover, the MFDNN approach reproduced and evaluated a transition point between saturatable point-defect swelling and non-saturated void swelling, which the empirical model did not represent. These results highlight the usefulness of MFDNNs and their potential for many applications beyond nuclear energy.