Proposal of Bayesian Framework for Inverse Inference to Predict Material Properties from Microstructures

This presentation introduces a Bayesian framework designed for inverse inference to predict material properties or process parameters from microstructure images. The basic strategy of the proposed framework is the integration of Bayesian inference methodologies with machine learning techniques. This enables us to not only predict target properties but also quantify the uncertainty associated with these predictions. Specifically, our focus lies on the development of a Bayesian framework based on generative networks which recently have received much attention in the field of computational material science. The integration could contribute to the examination of the prediction uncertainty.<br/><br/>In this presentation, we will explain the fundamental concepts of our framework and the outcomes of its application of this to a specific case study. Our chosen case is the prediction of material properties from artificial dual-phase steel microstructures. Also, we will discuss the comparison between our proposed Bayesian framework and a conventional inference method based on convolutional neural networks. By the end of this presentation, you will gain insights into the novel Bayesian approach and its potential advantages over conventional methods, providing a powerful tool for more robust predictions in the field of computational material science.