Genome-inspired AI for extrapolative discovery of high-temperature superconductors

Abstract

Inspired by nature, this study employs the Materials Genome Initiative to identify key components of high-temperature superconductors. AI with high-throughput screening (HTS), we uncover crucial superconducting "genes". Through HTS techniques and ML models, we demonstrate that Functional Convolutional Neural Networks (CNNs) ensure accurate extrapolation of potential compounds. Leveraging extensive datasets from the ICSD, the COD, our implemented HTS pipeline classifies likely superconductors, with CNN and long short-term memory (LSTM) models predicting Tc and their foundational elements. We address the scarcity of non-superconducting material data by compiling a dataset of 53,196 non-superconducting materials (DataG Non-Sc) and introduce a novel neural network architecture using Functional API for improved prediction, offering a powerful tool for future superconductor discovery. Our findings underscore the transformative potential of combining HTS with AI-driven models in advancing HTSC materials, highlighting Pu and H elements (with Tc ~ 100 K) as significant predictors of high-temperature superconductivity, suggesting their role as a crucial “gene” in these materials. Our HTS pipeline predicts 24 new binary Pu-based superconductors with Tc > 50 K at ambient pressure, including 16 compounds exceeding 70 K, in three certain categories, alongside over 40 hydrides with Tc values up to 100 K (21 systems surpassing 64 K).

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗