AI accelerates the discovery for high-temperature hydride superconductors

When cooled below a critical temperature, superconductors exhibit zero electrical resistance and completely expel magnetic fields—a phenomenon known as the Meissner effect. These properties hold revolutionary potential for power transmission, maglev transportation, medical imaging, and quantum computing. However, the practical application of conventional superconductors is severely limited by their requirement for extreme cryogenic cooling. Thus, the search for materials that superconduct at higher temperatures remains a central challenge in condensed matter physics and materials science.

In recent years, hydrogen-rich compounds, or superhydrides, have attracted much attention for high-temperature superconductivity research under high pressure. From sulfur hydrides to alkaline-earth and rare-earth hydrides, a series of materials have demonstrated superconducting transition temperatures exceeding 200 Kelvin (approximately -73°C) under pressure, igniting hope for achieving even higher or even room-temperature T_c.

Finding new superconducting materials, however, is challenging task. For ternary hydrides alone, considering just a limited set of elemental combinations can yield tens of thousands of possible chemical compositions and crystal structures. Traditional trial-and-error experiments are prohibitively costly and time-consuming. While highly accurate, quantum mechanical first-principles calculations demand immense computational resources, making large-scale systematic searches impractical.

To overcome this bottleneck, a collaborative team from Institute of Applied Physics and Computational Mathematics (IAPCM), Jilin University (JLU), and Peking University (PKU) has developed a "deep learning-driven" framework to intelligently and efficiently screen vast numbers of candidate materials for high-temperature superconducting hydrides. Their work, titled "Computational discovery of High-Temperature Superconducting Ternary Hydrides via Deep Learning," has been published in National Science Review. The corresponding authors are Researcher Han Wang, Professor Jian Lv, Professor Hanyu Liu, and Academician Yanming Ma.

The core of their framework is a "Large Atomic Model," a deep learning interatomic potential model trained on first-principles data from over 200,000 hydride structures. This model enables rapid and accurate prediction of the energy and thermodynamic stability of ternary hydrides under high pressure. The team integrated this model with the crystal structure prediction software CALYPSO to systematically scan a vast chemical space comprising hydrogen, 19 metals, and 9 non-metals, exploring over 36 million ternary hydride structures—a scale that far exceeds the conventional methods.

Facing this deluge of candidates, the researchers designed a multi-step, physics-informed screening pipeline:

1. Thermodynamic Stability: The Large Atomic Model performed an initial fast filter, retaining only low-energy, potentially stable structures.
2. Crystal Symmetry: High-symmetry crystal structures, often more conducive to superconductivity, were prioritized.
3. Atomic Hydrogen Content: Materials with a high proportion of hydrogen in atomic (rather than molecular) form were retained, a key feature for strong superconductivity.
4. Dynamic Stability: Structures with unstable vibrational modes ("imaginary frequencies") were eliminated.
5. First-Principles Verification: Final candidate materials underwent precise quantum mechanical calculations for confirmation.

This rigorous process narrowed the pool from over 36 million structures to just over 2,400 promising configurations.

Performing precise Tc calculations on thousands of materials remained computationally expensive. The team innovated further: by fine-tuning their Large Atomic Model on a smaller set of Tc data, they created a dedicated model capable of directly predicting Tc with an mean absolute error of only about 20 Kelvin at 200 GPa.

Using this predictor to prioritize candidates and combining it with precise validation calculations, the team successfully identified 144 potential high-temperature superconducting ternary hydrides. Among these, 129 compounds and their corresponding 27 structural prototypes are newly predicted, nearly doubling the number of known high-Tc hydride prototypes. All predicted materials have a Tc above 200 K and show good relative thermodynamic stability. The study highlights several particularly promising elemental systems, such as Y-Th-H, Ca-Y-H, and Y-Lu-H, which contain multiple high-Tc, stable candidates for prioritized experimental synthesis.

The significance of this work extends far beyond the discovery of new candidate materials. It successfully demonstrates a new paradigm of "AI-driven high-throughput materials discovery." This paradigm deeply integrates the efficient screening power of deep learning, the global search capability of crystal structure prediction, and the precise verification of first-principles calculations into a scalable, highly efficient workflow. While all predictions await future experimental validation under high pressure, this research undoubtedly introduces new possibility in the pursuit of high-temperature superconductivity and serves as a compelling example of "AI for Science" in action.