GLASS: A Geometric Local Attention-Based Model to Design Superconducting Structures

Overview

Superconductors, materials that achieve zero resistivity when cooled below their critical temperature, show promising applications in quantum computing, nuclear fusion, and electrical transmission. However, the development of superconductors currently relies on a slow and expensive trial-and-error process. To rapidly accelerate the discovery of new superconductors, this research proposes GLASS, a novel, deep-learning approach to generate high-temperature superconductors. GLASS is the first-ever end-to-end pipeline to utilize a molecule’s 3D crystal structure for superconductor design. GLASS consists of two parts: a novel Conditional Deep-Feature-Consistent Variational Autoencoder (Cond-DFC-VAE) that generates new potential superconductors, and a novel Geometric-Information-Enhanced Crystal Graph Neural Network (GeoCGNN) that models electron-phonon interactions to predict a material’s critical temperature. GLASS designs superconductors efficiently and effectively. GLASS's generative approach allows it to freely explore the entirety of crystal space, while its high physical fidelity, with 87% of GLASS-designed superconductors containing an Ef < 0, ensures that these molecules are stable and synthesizable. Generating 10,000 new superconductors, GLASS discovered 7 new high-temperature superconductors with a critical temperature over 100K. The validity of these results is reinforced by GLASS's critical temperature prediction accuracy, which outperformed all state-of-the-art methods by over 50%. Not only does GLASS design targeted, high-temperature superconductors, but it does so 1,400,000x faster than the traditional trial-and-error process. Through the creation of a novel pipeline for superconductor discovery, GLASS pushes us closer to a quantum, carbon-free future.

Competition history

ISEF 2023 Materials Science · Entry MATS045 · Dallas, Texas, United States

Resources

ISEF abstract page

Related projects

ISEF · 2022

ScGAN: A Generative Adversarial Network To Predict Hypothetical Superconductors

ISEF · 2025

Symmetry-Preserving Variational Autoencoder and Latent Space Rectified Flow Diffusion for Accelerated Materials Discovery With Geometric Graph Neural Networks and Integrated Property Network Multi-Layer Perceptron

ISEF · 2019

Disentangling Spatial Correlations from Inhomogeneous Materials with Shift-Invariant Artificial Neural Networks: A Novel Approach to Study Superconductivity

ISEF · 2018

Procedural Determination of Novel Stoichiometric Topological Superconductors through Surface and Pressure Effects

Closest projects by meaning, across every fair and year in the corpus.