Advances in Crystal Structure Prediction Unlock Superconducting Hydride Stability at 150 GPa

Predicting the stable arrangement of atoms within a material, known as crystal structure prediction, becomes increasingly difficult when considering the effects of temperature and atomic vibrations, particularly in materials containing light elements. Daniil Poletaev and Artem Oganov, both from the Skolkovo Institute of Science and Technology, along with their colleagues, address this challenge by combining advanced machine learning techniques with a method for accurately modelling atomic motion. Their work demonstrates a new approach to predicting crystal structures at realistic temperatures, accounting for the quantum behaviour of atoms within the material. This breakthrough simplifies the complex energy landscape of potential structures and allows for the correct identification of stable phases, which is crucial for discovering materials with desirable properties, such as superconductivity, that might otherwise be missed by conventional methods.

DFT and Machine Learning for Materials Discovery

Computational materials science is rapidly evolving, progressing from traditional density functional theory (DFT) methods to increasingly sophisticated techniques incorporating machine learning. This advancement aims to improve the accuracy, efficiency, and scalability of materials modeling, enabling the prediction of material properties at the atomic level. Key areas of investigation include high-pressure hydrogen phases, potentially exhibiting superconductivity, and the accurate modeling of atomic vibrations through anharmonic lattice dynamics. A central component of this progress is the development of machine learning potentials (MLPs), which are trained on DFT data to predict material energies and forces, significantly accelerating simulations.

Various MLP approaches are being explored, including charge-informed neural networks (CHGNet), graph neural networks (GNNs), equivariant neural networks, and graph atomic cluster expansions (GACE). Recent advancements include foundation models, like MatterSim and Orb, designed for speed and accuracy, alongside data-efficient learning techniques to minimize the need for extensive DFT data. Researchers are also employing multifidelity learning, combining data from different theoretical levels to optimize computational cost and accuracy. These calculations utilize software packages such as VASP for DFT calculations and ASE, a Python library for atomic simulations. This integration of DFT and machine learning represents a state-of-the-art approach to materials discovery and design, promising to accelerate the identification of novel materials with tailored properties.

Machine Learning Accelerates Anharmonic Crystal Structure Prediction

Scientists have developed a new method for crystal structure prediction, integrating machine learning interatomic potentials (MLIPs) with the stochastic self-consistent harmonic approximation (SSCHA) to explore anharmonic free-energy landscapes. This is particularly important for materials with lightweight atoms, such as superconducting hydrides, where atomic vibrations are complex. The team pioneered the use of evolutionary CSP, overcoming the challenge of accurately predicting crystal structures at finite temperatures. Researchers employed USPEX, a structure prediction algorithm, and combined it with actively learned MLIPs (AL-MLIPs) and universal MLIPs (uMLIPs) to accelerate computationally intensive SSCHA calculations.

Using lanthanum hydride (LaH10) at 150 GPa and 300 K as a benchmark, they demonstrated two strategies for SSCHA-based CSP. One approach involved training AL-MLIPs for each structure, while the other harnessed pre-trained universal MLIPs, MatterSim-5m, from the Matbench project, enabling structure prediction without per-structure training. Experiments involving hundreds of SSCHA optimizations, traditionally limited by computational cost, were significantly accelerated by MLIP integration. Results demonstrate that incorporating anharmonicity simplifies the free-energy landscape, crucial for correctly identifying stable phases. The team successfully predicted the experimentally verified cubic Fm m phase of LaH10, confirming the effectiveness of the methodology and extending the reach of CSP to systems dominated by nuclear motion and anharmonicity.

Machine Learning Predicts Lanthanum Hydride Structure

Scientists achieved a breakthrough in crystal structure prediction by integrating machine learning interatomic potentials with the stochastic self-consistent harmonic approximation to accurately model materials at finite temperatures and under pressure. This work addresses the challenge of predicting the behavior of systems containing lightweight atoms, such as superconducting hydrides, where quantum effects are significant. The team focused on lanthanum hydride, LaH10, at 150 GPa and 300 K, employing two approaches to refine the SSCHA-based crystal structure prediction process. Utilizing active-learning machine learning interatomic potentials (AL-MLIPs), trained independently for each structure, correctly identified the experimentally confirmed cubic Fm-3m phase as the most stable form of LaH10.

Consistent results with AL-MLIPs required corrections using thermodynamic perturbation theory. Alternatively, the team explored universal MLIPs, Mattersim-5m, from a pre-existing database, which conducted SSCHA-based predictions without individual training for each structure. Results demonstrate that incorporating quantum anharmonicity simplifies the free-energy landscape and is crucial for determining the correct stability ranking of crystal structures. This is particularly important for predicting high-temperature phases, often overlooked by traditional methods. The breakthrough delivers a powerful new tool for materials discovery and design, enabling scientists to explore a wider range of materials with enhanced accuracy and efficiency.

Anharmonicity Predicts Stable Lanthanum Hydride Structure

This research presents a new method for predicting crystal structures at realistic temperatures, accounting for the effects of atomic motion and anharmonicity. Scientists successfully integrated machine learning interatomic potentials with a stochastic self-consistent harmonic approximation to explore the free-energy landscape of materials, a crucial step for accurate prediction. Applying this approach to lanthanum hydride at high pressure and temperature, the team demonstrated the ability to correctly identify the experimentally observed cubic phase as the most stable form. The study highlights the importance of including anharmonicity in crystal structure prediction, as it simplifies the complexity of the energy landscape and improves the accuracy of stability rankings, particularly for high-temperature phases.

Researchers compared two machine learning strategies, training new models specifically for lanthanum hydride or utilizing pre-trained, universal models, finding both approaches viable with appropriate refinements. While locally trained models require careful error correction, the universal models offer a promising path toward structure prediction without extensive per-structure training. The authors acknowledge that their method may miss certain high-energy structures stable at elevated temperatures, necessitating a comprehensive search of the full anharmonic landscape. Future work could focus on further refining the machine learning potentials and extending the method to a wider range of materials, potentially accelerating the discovery of novel compounds with desirable properties, such as high-temperature superconductors. This simplification of the structural search space offers a significant advantage by reducing the computational cost of property calculations for candidate materials.