Prediction of crater formation in a large pulsed electron beam (LPEB) irradiation process using deep learning

In a large pulsed electron beam (LPEB) process, it is crucial to optimize processing parameters to minimize crater formation on a metal surface. Traditional approaches have relied on physics-based models of predicting temperature distribution and melting depth. In this study, a novel data-driven deep learning model is presented to predict crater formation in the LPEB process, from an input vector consisting of material properties (non-metallic chemical composition and heat diffusivity) and processing parameters (energy density and the number of electron pulses). The model was a spectral-norm-based conditional residual generative adversarial network (GAN), which ensured a stable translation from the input vector to the LPEB surface image including the craters. LPEB-processed optical microscopic images were provided as ground truths for four different steel alloys (SKD11, SKD61, NAK80, and KP1). Subsequently, for a more accurate quantitative analysis of the craters, an unsupervised deep learning model was proposed coupled with a noise filtering technique. The deep learning model successfully predicted the crater formation with accuracies of 84.5 % for crater size (mean absolute error of 3.70 μm), 93.8 % for crater number, and 88.9 % for crater distribution. Additionally, an experiment involving 'walking in the condition space' was conducted, revealing a sound level of understanding by the deep learning model. The prediction time was less than a second. • Crater-like defects on LPEB-processed steel alloys were predicted using a novel deep learning model. • Optimized model architecture for generating LPEB surface images with craters was presented. • Quantitative analysis was performed in terms of crater size, number, and distribution using an unsupervised model. • Prediction results showed good physical consistency including wave patterns at the edges. • Physical effectiveness of the model was further assessed through walking in the condition space.