Invertible crystallographic representation for inorganic crystals via new API

Abstract

This project aims to improve thermoelectric material identification and evaluation by utilizing the advanced machine-learning framework. We utilize the vast dataset of the Materials Project database, focusing on material properties including formation energy, band gap, and crystal structure, by combining Variational Autoencoders (VAEs) with semi-supervised learning approaches. Specifically, the approach develops a Fourier Transformed Crystal Properties (FTCP) representation that is carefully designed for deep learning applications. This representation enables our model to encode complex, high-dimensional data into a concise latent space. From this latent space, we develop new materials by the manipulation of latent variables and their subsequent decoding, revealing materials with improved thermoelectric characteristics, such as optimized Seebeck coefficients. This method improves the model's learning process and prediction ability by utilizing both labelled and unlabeled data, going beyond traditional supervised learning. With the help of advanced training strategies like dynamic learning rate adjustments and thorough preparation processes like data normalization and augmentation, the model demonstrates a remarkable ability to predict material attributes. Furthermore, the interaction between different material properties and how that interaction affects thermoelectric performance may be better understood using graphical analysis. This project not only demonstrates the revolutionary potential of machine learning in material science, but it had also established a new standard for the effective and scalable search for advanced thermoelectric materials. Using the model's predictive capability, we set out to investigate the wide range of possible materials to find and develop materials that have the potential to revolutionize thermoelectric technology in the future. Hence, this effort demonstrates the complementary nature of material science and computational science and opens a new path for creative approaches to energy conversion and management.