Poster: Machine Learning Prediction and Experimental Synthesis of New High-Temperature Superconductors

Data-driven machine learning, informed by structural crystallographic data, offers a powerful tool for designing improved materials. In this study, we leverage this approach to synthesize and experimentally evaluate new high temperature superconductors. We test and employ models such as the Crystal Diffusion Variational Autoencoder (CDVAE), AtomGPT, and custom active learning models to train and generate stable superconductors optimized for high critical temperatures. Candidate materials are screened using the Atomistic Line Graph Neural Network (ALIGNN) model to estimate key properties, including formation energy, critical temperature, and band gap. We then experimentally test the most promising, high-temperature, stable candidates. This research demonstrates a successful workflow for materials discovery, drawing on established methodologies and available tools to gather data, curate training datasets, and train graph neural networks to learn patterns between structural and material properties. Our approach enables the generation of novel materials through a physics-informed variational autoencoder, followed by screening for stability and other desired properties, and ultimately, experimental validation of these materials.