Deep learning guided estimation of laser sintered ceramic’s microstructure based on in-situ surface thermal emission signals

In laser processing, online estimation of microstructures is useful to ensure that desired properties are obtained. Currently, microstructures are usually characterized offline using scanning electron microscopes (SEM), which are labor, time, and cost-consuming. Here, we demonstrate a deep learning-based computational method to simulate micrographs, which resemble real SEM images containing important microstructure features, from the in-situ captured thermal emission strength. Experimental results reveal a strong correlation between the thermal emission brightness and the corresponding microstructure. A conditional generative adversarial network (CGAN) is trained to model this correlation. To obtain the best accuracy, a toy dataset consisting of artificial microstructures is established to investigate the effects of the network’s hyperparameters on performance. Then, the optimized network is trained on a real SEM image dataset and its accuracy is quantitatively evaluated using average grain size as the metric. The CGAN-generated microstructure images were found to be in good agreement, less than 5% in difference in average grain size, with the real SEM images. And the inference time is less than 1 second per image. This fast CGAN-based microstructure estimation method can potentially be used for process control and quality assurance in the laser sintering of ceramics to accelerate material development and improve productivity.