Predicting Crystal Stability at High Temperatures

Author: Denis Avetisyan


A new computational approach combines machine learning with quantum mechanics to accurately forecast crystal structures under realistic thermal conditions.

Structure optimization within the SSCHA framework leverages active-learning machine learning interatomic potentials to efficiently refine system design.
Structure optimization within the SSCHA framework leverages active-learning machine learning interatomic potentials to efficiently refine system design.

This work integrates machine learning interatomic potentials and the stochastic self-consistent harmonic approximation within the USPEX framework for finite-temperature crystal structure prediction, successfully demonstrated with LaH10.

Accurate prediction of crystal structures at finite temperatures remains a significant challenge, particularly for lightweight materials exhibiting strong quantum effects. This work, presented in ‘SSCHA-based evolutionary crystal structure prediction at finite temperatures with account for quantum nuclear motion’, introduces a novel approach integrating machine-learned interatomic potentials with the stochastic self-consistent harmonic approximation within an evolutionary algorithm to map the anharmonic free-energy landscape. Demonstrating this method with LaH$_{10}$ at high pressure, we show that incorporating quantum nuclear motion simplifies structure prediction and yields correct stability rankings, even with transferable machine-learned potentials. Could this approach unlock accurate predictions for previously intractable high-temperature phases and accelerate the discovery of novel materials?


Navigating the Complexity of Materials Prediction

The quest for novel materials hinges on the ability to reliably predict stable crystal structures, yet current computational methods face significant hurdles when dealing with complex systems. Traditional approaches, often relying on simplified models or exhaustive searches, struggle to navigate the vast energy landscape and identify the globally stable configuration amongst numerous local minima. This difficulty arises because even minor variations in atomic arrangement can dramatically alter a material’s properties, and accurately modeling these subtle energy differences requires immense computational power. Consequently, predicting the structure of materials with multiple elements, defects, or under extreme conditions remains a considerable challenge, limiting the efficiency of materials discovery and hindering the design of materials with tailored functionalities.

The relentless pursuit of novel materials is fundamentally bottlenecked by the immense computational demands of accurately predicting their stability and properties. Exploring the full spectrum of potential chemical combinations – a space often described as a ‘chemical universe’ – is currently impractical given the exponential increase in computational cost with each added element or structural complexity. Current methods, even with advances in high-throughput computing, struggle to reliably discern between subtly differing energy states of potential crystal structures, leading to a high rate of false positives and wasted experimental validation. This limitation isn’t simply a matter of needing faster computers; it reflects inherent difficulties in modeling the intricate interplay of quantum mechanical effects and atomic interactions that govern material behavior, demanding increasingly sophisticated – and resource-intensive – algorithms to achieve acceptable levels of predictive power.

The quest for novel materials hinges on the ability to accurately forecast their stability, yet this endeavor is profoundly complicated by the delicate balance of energetic factors at play. Minute variations in energy – often fractions of an electronvolt – dictate whether a particular atomic arrangement will endure or readily transform into another configuration. Consequently, computational models must move beyond simple approximations and meticulously capture the intricacies of atomic interactions, including van \, der \, Waals forces, charge transfer, and many-body effects. Failing to account for these subtle nuances can lead to predictions of materials that are thermodynamically unstable or possess properties drastically different from reality, underscoring the need for increasingly sophisticated theoretical frameworks and computational power to navigate the vast landscape of potential materials.

Anharmonic free energy calculations at 300 K and 150 GPa predict three stable <span class="katex-eq" data-katex-display="false">\mathrm{LaH}_{10}</span> polymorphs-<span class="katex-eq" data-katex-display="false">\mathrm{Fm}\bar{3}\mathrm{m}</span>, <span class="katex-eq" data-katex-display="false">\mathrm{Cmmm}</span>, and <span class="katex-eq" data-katex-display="false">\mathrm{P6/mmm}</span>-relative to the ground state <span class="katex-eq" data-katex-display="false">\mathrm{Fm}\bar{3}\mathrm{m}</span> structure, as visualized using STMng.
Anharmonic free energy calculations at 300 K and 150 GPa predict three stable \mathrm{LaH}_{10} polymorphs-\mathrm{Fm}\bar{3}\mathrm{m}, \mathrm{Cmmm}, and \mathrm{P6/mmm}-relative to the ground state \mathrm{Fm}\bar{3}\mathrm{m} structure, as visualized using STMng.

Accelerating Discovery with Machine Learning Potentials

Machine learning interatomic potentials (MLIPs) represent a computational approach to predicting the energy of atomic configurations, thereby accelerating structure prediction calculations. Traditional methods rely heavily on computationally expensive ab initio techniques like Density Functional Theory (DFT) to determine these energies for every atomic arrangement evaluated during optimization or molecular dynamics simulations. MLIPs, however, are trained on a limited set of DFT calculations and then used to rapidly predict energies for a much larger set of configurations. This substitution significantly reduces the computational cost, allowing for the exploration of larger systems and longer timescales than would be feasible with DFT alone. The accuracy of MLIPs depends on the quality and quantity of the training data, as well as the chosen machine learning model, but when properly implemented, they can achieve near-DFT accuracy at a substantially reduced computational expense.

Universal machine learning interatomic potentials (MLIPs) are developed through training on large, diverse datasets encompassing multiple chemical elements and material structures. This extensive training allows the resulting potential to generalize beyond the specific compounds present in the training data, enabling reasonably accurate predictions of atomic interactions in novel materials. Unlike potentials tailored to specific systems, universal MLIPs provide a robust foundation for predicting the behavior of a wide range of materials, including those with limited or no prior data. The accuracy of these potentials is directly correlated to the size and quality of the training dataset, with larger, more representative datasets yielding improved predictive power across diverse chemical spaces.

Active learning (AL) strategies for machine learning interatomic potential (MLIP) development operate by iteratively refining the potential based on the uncertainty in its predictions. Rather than uniformly sampling configurations for density functional theory (DFT) calculations to generate training data, AL algorithms intelligently select configurations where the MLIP is most uncertain or expected to yield the largest improvement in accuracy. This is achieved by quantifying the potential’s error-typically through metrics like the root mean squared error-and prioritizing calculations on configurations that maximize information gain. Subsequent DFT calculations on these selected configurations are then used to retrain the MLIP, reducing uncertainty and improving predictive power with each iteration, ultimately leading to a more accurate potential with fewer total DFT calculations.

The implementation of Active Learning Machine Learning Interatomic Potentials (AL-MLIPs) has resulted in a demonstrated reduction of Density Functional Theory (DFT) calculations exceeding a factor of 3x. This improvement in computational efficiency stems from the ability to reuse previously trained potentials across diverse materials and configurations. By strategically selecting data points for retraining via active learning, the number of computationally expensive DFT calculations required to achieve a target accuracy is minimized. This approach allows researchers to explore a larger chemical space and accelerate materials discovery workflows without a proportional increase in computational resources.

The MaxVol Selection method enhances the efficiency of active learning (AL) by prioritizing retraining configurations based on the volume of the convex hull formed by the potential energy surface. This approach identifies data points that, when added to the training set, maximize the reduction in uncertainty regarding the potential energy surface. Specifically, configurations are selected to minimize the largest volume tetrahedron formed by the potential energy values at those points and a subset of existing training data. This strategy contrasts with purely random or energy-based selection, resulting in faster convergence of the MLIP and a reduction in the number of computationally expensive ab initio calculations required to achieve a target level of accuracy. By focusing on configurations that most effectively constrain the potential energy space, MaxVol Selection optimizes the retraining process and improves the overall efficiency of AL-MLIP development.

Refining Predictions with Thermodynamic Rigor

The USPEX code leverages the combination of thermodynamic perturbation theory and machine learning interatomic potentials (MLIPs) to achieve reliable structure prediction. Thermodynamic perturbation theory provides a framework for accurately calculating free energies and phase stabilities, while MLIPs accelerate these calculations by approximating the potential energy surface. This integration allows USPEX to efficiently explore a vast chemical space and identify globally stable or low-energy structures that may be inaccessible to traditional methods. By combining the theoretical rigor of thermodynamic calculations with the speed of machine learning, USPEX provides both accuracy and computational efficiency in materials discovery.

Density Functional Theory (DFT) calculations are integral to the prediction workflow, serving a dual purpose in conjunction with Machine Learning Interatomic Potentials (MLIPs). DFT provides the high-accuracy data necessary to train the MLIPs, establishing a reliable potential energy surface for larger-scale simulations. Subsequently, DFT calculations are employed to verify the stability and energy of structures predicted using the trained MLIPs, confirming their thermodynamic feasibility and ensuring the accuracy of the overall prediction process. This iterative approach, leveraging the strengths of both DFT and MLIPs, enhances the robustness and reliability of structural predictions.

Accurate determination of structural stability requires a complete accounting of quantum anharmonicity, as harmonic approximations underestimate the free energy and can lead to incorrect predictions of ground state structures. Anharmonicity arises from deviations of atomic vibrations from simple harmonic motion, necessitating the inclusion of higher-order vibrational terms in the calculation of free energy. Failure to incorporate these anharmonic contributions results in systematically biased free energy differences, potentially misidentifying the true thermodynamic minimum. Techniques such as vibrational perturbation theory or, more rigorously, vibrational self-consistent field methods, are employed to calculate anharmonic contributions to the free energy, thereby improving the reliability of structure predictions and ensuring accurate assessments of structural stability.

Self-consistent second-order perturbation theory (SSCHA) calculations were performed using a 2x2x2 supercell to assess the accuracy of free energy determinations. This supercell size facilitated convergence to an energy resolution of 2 meV/atom, indicating a high level of precision in the calculated thermodynamic properties. The achieved convergence demonstrates the reliability of the method for predicting the stability of crystalline structures and identifying low-energy configurations with minimal systematic error. This level of accuracy is critical for materials discovery and the precise modeling of material behavior under varying conditions.

The Mattersim-5m machine learning interatomic potential (MLIP) demonstrates high convergence rates, achieving successful structure relaxation for 90% of input structures. This performance significantly reduces the computational cost associated with structure prediction workflows. Critically, the universal nature of the uMLIPs minimizes the necessity for system-specific training data; this eliminates the need to generate bespoke potentials tailored to individual material compositions, greatly accelerating the discovery process and broadening the scope of applicable systems without sacrificing accuracy.

Phonon dispersion curves reveal that cubic, orthorhombic, and hexagonal <span class="katex-eq" data-katex-display="false">\mathrm{LaH}_{10}</span> structures exhibit distinct harmonic and anharmonic vibrational modes at 150 GPa, reflecting their differing symmetry groups (<span class="katex-eq" data-katex-display="false">\mathrm{Fm}\bar{3}\mathrm{m}</span>, <span class="katex-eq" data-katex-display="false">\mathrm{Cmmm}</span>, and <span class="katex-eq" data-katex-display="false">\mathrm{P6/mmm}</span>, respectively).
Phonon dispersion curves reveal that cubic, orthorhombic, and hexagonal \mathrm{LaH}_{10} structures exhibit distinct harmonic and anharmonic vibrational modes at 150 GPa, reflecting their differing symmetry groups (\mathrm{Fm}\bar{3}\mathrm{m}, \mathrm{Cmmm}, and \mathrm{P6/mmm}, respectively).

Unveiling Diverse Structures in LaH10

Recent computational work has successfully predicted multiple stable arrangements of lanthanum decahydride (LaH10) by uniting the USPEX evolutionary algorithm with sophisticated machine learning interatomic potentials (MLIPs). This combined approach allows for efficient and accurate exploration of the vast compositional and structural space, identifying several previously unknown, yet energetically viable, configurations of LaH10. The synergy between USPEX’s global optimization capabilities and the MLIPs’ speed and precision overcomes limitations inherent in traditional methods, providing a robust framework for materials discovery and ultimately enhancing the understanding of high-pressure hydride behavior.

The successful prediction of multiple stable LaH10 structures – specifically the Fm-3m, Cmmm, and P6/mmm phases – demonstrates a significant advancement in computational materials science. These structures, while all representing LaH10, exhibit distinct atomic arrangements and, crucially, possess remarkably similar energies. The ability to accurately discern these subtle energy differences-often on the scale of a few milli-electronvolts-is a testament to the combined power of the USPEX evolutionary algorithm and the advanced machine learning interatomic potentials (MLIPs) employed in the study. This sensitivity is critical for identifying truly stable phases that might otherwise be overlooked, and it suggests the methodology isn’t simply converging on the lowest energy structure, but rather comprehensively mapping the energy landscape to reveal a diversity of viable configurations.

The discovery of multiple stable structures within lanthanum decahydride (LaH10) reveals a surprisingly nuanced relationship between its atomic arrangement, chemical bonding, and overall stability. High-pressure hydrides, like LaH10, don’t conform to a single, predictable structure; instead, subtle shifts in how atoms bond dramatically alter the material’s properties. Researchers found that even small changes to the crystal lattice – transitioning between Fm-3m, Cmmm, and P6/mmm arrangements – impact the way hydrogen atoms interact with lanthanum, influencing the strength of the chemical bonds and, consequently, the material’s ability to maintain its form under extreme pressure. This complex interplay suggests that tailoring the structure of LaH10, and similar hydrides, could unlock enhanced performance characteristics, potentially leading to breakthroughs in superconductivity or high-energy density materials.

The computational methodology – combining the USPEX evolutionary algorithm with machine-learned interatomic potentials – transcends the specific study of lanthanum decahydride. It establishes a robust framework for de novo materials discovery, enabling the prediction of stable and potentially groundbreaking compounds beyond high-pressure hydrides. By efficiently navigating complex chemical spaces and accurately assessing structural stability, this approach allows researchers to proactively design materials with targeted characteristics – such as superconductivity, enhanced catalytic activity, or improved energy storage capabilities – before physical synthesis. The ability to computationally screen vast numbers of compositions and crystal structures significantly accelerates the materials innovation cycle, promising a future where materials are engineered with unprecedented precision and efficiency.

Phonon dispersion curves reveal that cubic, orthorhombic, and hexagonal <span class="katex-eq" data-katex-display="false">\mathrm{LaH}_{10}</span> structures exhibit distinct harmonic and anharmonic vibrational modes at 150 GPa, reflecting their differing symmetry groups (<span class="katex-eq" data-katex-display="false">\mathrm{Fm}\bar{3}\mathrm{m}</span>, <span class="katex-eq" data-katex-display="false">\mathrm{Cmmm}</span>, and <span class="katex-eq" data-katex-display="false">\mathrm{P6/mmm}</span>, respectively).
Phonon dispersion curves reveal that cubic, orthorhombic, and hexagonal \mathrm{LaH}_{10} structures exhibit distinct harmonic and anharmonic vibrational modes at 150 GPa, reflecting their differing symmetry groups (\mathrm{Fm}\bar{3}\mathrm{m}, \mathrm{Cmmm}, and \mathrm{P6/mmm}, respectively).

The pursuit of accurate crystal structure prediction, as demonstrated with LaH10, necessitates a holistic understanding of system behavior, not merely isolated calculations. This work elegantly addresses the limitations of static harmonic approximations by incorporating quantum anharmonicity via SSCHA and MLIPs. It mirrors the sentiment expressed by David Hume: “The mind is not a passive recipient of experience, but actively constructs its own reality.” Just as Hume suggests the mind actively shapes perception, this methodology actively constructs a more realistic thermodynamic landscape, acknowledging that structural stability isn’t a fixed point but emerges from the dynamic interplay of quantum effects and thermal fluctuations. The integration of machine learning further refines this process, enabling a more nuanced and predictive understanding of material behavior.

Future Directions

The pursuit of stable materials at elevated temperatures invariably exposes the limitations of harmonic approximations. This work, integrating machine learning potentials with the stochastic self-consistent harmonic approximation, represents a necessary, if incremental, step towards addressing that reality. The demonstrated success with LaH10, while encouraging, subtly underscores a broader truth: predictive power rests not merely on computational refinement, but on the quality of the underlying potential energy surface. Further gains will likely demand a more holistic approach to potential development, one that systematically incorporates known anharmonicity and extends beyond pairwise interactions.

A crucial, often overlooked, aspect lies in the inherent tension between computational tractability and physical realism. The current framework, like many before it, trades complexity for speed. Future iterations should explore methods to efficiently account for higher-order anharmonic effects without sacrificing the ability to screen large chemical spaces. Perhaps a more nuanced understanding of the relationship between structural simplicity and thermodynamic stability will emerge – a principle suggesting that nature often favors elegance even amidst thermal fluctuations.

Ultimately, the field must move beyond simply predicting a structure, and begin to map the full, dynamic landscape of possible structures as a function of temperature. This demands not just more accurate potentials and algorithms, but a fundamental shift in perspective – viewing materials not as static entities, but as evolving systems, constantly adapting to their environment. The quest for novel materials, after all, is a study in resilience, and resilience is born from adaptability.


Original article: https://arxiv.org/pdf/2512.24849.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-01-02 18:44