Researchers at the University of Florida’s Department of Materials Science & Engineering have developed a machine-learning framework that could revolutionize the way high-temperature superconductors are discovered.
Superconductors are materials capable of conducting electricity without resistance, a property that could transform energy transmission, transportation, and medical technologies. However, finding new superconductors has been a slow, expensive process due to the complexities of analyzing their properties.
To address this challenge, Richard Hennig, Ph.D., a professor in the UF Department of Materials Science & Engineering, and two of his doctoral students, Jason Gibson and Ajinkya Hire, developed BETE-NET (Bootstrapped Ensemble of Tempered Equivariant Graph Neural Networks). This AI-driven model integrates physical principles with machine learning to predict superconducting properties, including the critical temperature (Tc) at which a material becomes superconducting.
In their study, “Accelerating Superconductor Discovery through Tempered Deep Learning of the Electron-Phonon Spectral Function,” published in Nature, the team demonstrated BETE-NET’s ability to outperform conventional methods in predicting key superconducting properties.
“BETE-NET stands out because it incorporates phonon interactions directly into its learning process,” Hennig said. “This approach enables high accuracy without requiring massive datasets, a common limitation in materials science.”
The team trained BETE-NET using data from 818 stable materials, achieving predictions with remarkable precision while significantly reducing computational costs. Using UF’s high-performance computing system, HiPerGator, they applied the model to a large materials database, identifying six known superconductors and proposing 82 new candidates for experimental testing.
The potential impact is broad. Superconducting power grids could eliminate energy loss, reducing costs and improving efficiency. Magnetic levitation trains could become faster and more sustainable. Medical imaging technologies could see both performance improvements and cost reductions. The team also sees potential applications beyond superconductors, including quantum materials and topological insulators.
“We’re combining expertise from physics, computer science, and materials engineering to address a problem that’s been a bottleneck for decades.”
Richard Hennig Ph.D.
Despite the initial promising results, conventional experimental validation remains essential to confirm BETE-NET’s predictions and identify any unforeseen limitations.
Hennig emphasized the collaborative nature of this breakthrough. “We’re combining expertise from physics, computer science, and materials engineering to address a problem that’s been a bottleneck for decades,” Hennig said. “Integrating AI with established physical principles is opening new doors in materials discovery.”
BETE-NET represents a notable step forward in materials science, demonstrating the power of AI to accelerate discoveries. As researchers refine this technology, its applications could drive progress across multiple industries, shaping a more efficient and sustainable future.