Transfer Learning Technique Using Similar Phenomena Training for Fuel Digital Twins of Nuclear Systems

Transfer Learning (TL) is a procedure used for training Neural Networks (NN) by using pretrained models to provide information to a model with a new output. This is achieved by training one model, freezing the initial layers, and adding an aditional output to them. TL aims to reduce the training data requirement of NNs (valuable in nuclear systems), although most reserached applications focus on image processing or phenomena with more data than nuclear systems. This work aims to use similar or constituent phenomena for TL applications. Specifically, thermal expansion of SiC/SiC cladding is used to inform irradiation swelling at high temperatures. Results show that the trained TL model generalizes predictions better than traditional feedforward models, and achieves higher accuracy. Additionally, TL models were able to extrapolate information outside of the original model’s bounds more easily than feedforward modules. This type of work is expected to be used in nuclear systems due to the low data availability and benefit of generalizing outputs.