Advances in crystal structure prediction unlock stability of superconducting hydrides at 150 GPa

Predicting the stable arrangement of atoms in a material, known as crystal structure prediction, becomes increasingly difficult when considering the effects of temperature and atomic vibrations, especially in materials containing light elements. Daniil Poletaev and Artem Oganov from Skolkovo University of Science and Technology, together with their colleagues, are tackling this challenge by combining advanced machine learning techniques with methods for accurately modeling the motion of atoms. Their work demonstrates a new approach to predicting crystal structure at realistic temperatures by considering the quantum behavior of atoms within materials. This breakthrough simplifies the complex energy landscape of the potential structure and allows precise identification of stable phases. This is crucial for discovering materials with desirable properties, such as superconductivity, that may be missed by traditional methods.

DFT and machine learning for materials discovery

Computational materials science is rapidly evolving, moving from traditional density functional theory (DFT) techniques to increasingly sophisticated techniques that incorporate machine learning. This advancement aims to improve the accuracy, efficiency, and scalability of materials modeling and enable the prediction of material properties at the atomic level. Key areas of research include high-pressure hydrogen phases that may exhibit superconductivity and accurate modeling of atomic vibrations through anharmonic lattice dynamics. A central element of this progress is the development of machine learning potentials (MLPs). MLP is trained on DFT data to predict material energies and forces, greatly accelerating simulations.

Various MLP approaches have been investigated, including charge-informed neural networks (CHGNet), graph neural networks (GNN), equivariant neural networks, and graph atomic cluster expansion (GACE). Recent advances include foundational models such as MatterSim and Orb designed for speed and accuracy, and data-efficient learning techniques that minimize the need for large amounts of DFT data. Researchers also employ multi-fidelity learning, which combines data from different levels of theory, to optimize computational cost and accuracy. These calculations utilize software packages such as VASP for DFT calculations and ASE, a Python library for atomistic simulations. This integration of DFT and machine learning represents a cutting-edge approach to materials discovery and design and is expected to accelerate the identification of novel materials with customized properties.

Machine learning accelerates prediction of anharmonic crystal structures

Scientists have developed a new crystal structure prediction method that integrates machine learning interatomic potentials (MLIP) and stochastic self-consistent harmonic approximation (SSCHA) to explore the anharmonic free energy landscape. This is especially important for materials containing lightweight atoms, such as superconducting hydrides, where the atomic vibrations are complex. The research team pioneered the use of evolutionary CSP to overcome the challenge of accurately predicting crystal structure at finite temperatures. The researchers adopted a structure prediction algorithm, USPEX, and combined it with active learning MLIP (AL-MLIP) and universal MLIP (uMLIP) to speed up the computationally intensive SSCHA computation.

They demonstrated two strategies for SSCHA-based CSP using lanthanum hydride (LaH10) at 150 GPa and 300 K as a benchmark. One approach required training AL-MLIP for each structure, while the other approach leveraged the Matbench project's pre-trained universal MLIP, MatterSim-5m, to enable structure prediction without training for each structure. Experiments involving hundreds of SSCHA optimizations, which were previously limited by computational cost, were significantly accelerated by the MLIP integration. The results demonstrate that incorporating anharmonicity simplifies the free energy landscape and is important for accurately identifying stable phases. The research team succeeded in predicting the experimentally verified cubic Fm m phase of LaH10, confirming the validity of the methodology and extending the scope of CSP to systems dominated by nuclear motion and anharmonicity.

Machine learning predicts the structure of lanthanum hydride

Scientists have achieved a breakthrough in crystal structure prediction by integrating machine learning interatomic potentials and stochastic self-consistent harmonic approximations to accurately model materials under finite temperature and pressure. This research addresses the challenge of predicting the behavior of systems containing lightweight atoms, such as superconducting hydrides, where quantum effects are significant. Focusing on lanthanum hydride LaH10 at 150 GPa and 300 K, the research team adopted two approaches to improve the SSCHA-based crystal structure prediction process. By utilizing active learning machine learning interatomic potentials trained individually for each structure (AL-MLIP), the experimentally confirmed cubic Fm-3m phase was correctly identified as the most stable form of LaH10.

To obtain consistent results with AL-MLIP, correction using thermodynamic perturbation theory was necessary. Alternatively, the team explored a universal MLIP, Mattersim-5m, from an existing database. It performed SSCHA-based prediction without separate training for each structure. The results show that incorporating quantum anharmonicity is important to simplify the free energy landscape and determine the correct stability ranking of crystal structures. This is especially important when predicting high-temperature phases that are often overlooked by traditional methods. This breakthrough provides powerful new tools for materials discovery and design, allowing scientists to explore a wider range of materials with increased precision and efficiency.

Anharmonicity predicts a stable lanthanum hydride structure

This study presents a new method for predicting crystal structures at realistic temperatures, taking into account the effects of atomic motion and anharmonicity. Scientists have successfully integrated machine learning interatomic potentials with stochastic self-consistent harmonic approximations to explore the free energy landscape of materials. This is an important step for accurate predictions. By applying this approach to hydrided lanthanum at high pressures and temperatures, the researchers demonstrated that they could accurately identify the experimentally observed cubic phase as the most stable form. This study highlights the importance of including anharmonicity in crystal structure predictions to simplify the complexity of the energy landscape and improve the accuracy of stability rankings, especially for high-temperature phases.

The researchers compared two machine learning strategies: training a new model specifically for lanthanum hydride or leveraging a pre-trained general-purpose model, and found that either approach is viable with appropriate refinements. While locally trained models require careful error correction, universal models offer a promising route to structure prediction without the need for extensive structure-by-structure training. The authors acknowledge that their method may miss certain high-energy structures that are stable at high temperatures and that a comprehensive search for fully anharmonic situations is required. Future research will focus on further refining the potential of machine learning 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 provides significant benefits by reducing the computational cost of calculating properties of candidate materials.