Ai’s Crystal Predictions Now Match Reality with Unprecedented Detail

Researchers are increasingly reliant on machine learned interatomic potentials (MLIPs) to perform accurate and efficient molecular dynamics simulations and large-scale crystal energetics calculations. Abhijith S. Parackal, Rickard Armiento, and Florian Trybel, from the Department of Physics, Chemistry and Biology at Linköping University, present a novel evaluation of these potentials by systematically examining their predicted energy landscapes. This work addresses a critical gap in current benchmarking methods, which often focus on overall validation errors rather than the detailed fidelity of potential energy surfaces. By constructing two-dimensional slices of these surfaces along specific symmetry directions, the authors provide a direct visual comparison with density functional theory calculations, revealing potential artifacts and offering valuable insights into the physical accuracy of pre-trained MLIPs and establishing benchmarks for future advancements in the field.

Detailed potential energy surface reconstruction reveals machine-learned interatomic potential fidelity for accurate simulations

Researchers have developed a novel method for evaluating the accuracy of machine learned interatomic potentials (MLIPs), increasingly used for simulating material behaviour at the atomic level. These potentials, often termed foundation models, offer a computationally efficient alternative to density functional theory (DFT) for large-scale simulations and energetic studies of crystals.
While benchmarks typically assess MLIP performance using validation errors and materials discovery tasks, a comprehensive understanding of their ability to reproduce detailed features of potential energy surfaces (PES) has remained incomplete. This work addresses this gap by systematically probing the predicted energy landscapes of several prominent MLIP architectures, including MACE, CHGNet, M3GNet, ORB, and SevenNet.

The study introduces a technique for constructing two-dimensional slices of the PES, achieved by varying atomic positions along specific symmetry-defined degrees of freedom within a fixed crystal structure. This approach facilitates a direct, visual comparison between the MLIP-predicted surfaces and those calculated using DFT, revealing potential artifacts arising from unique local atomic environments.

By analysing local minima, saddle points, and overall PES topology, the research highlights the strengths and limitations of each potential in capturing the underlying physics of material systems. The findings offer crucial insights into the physical accuracy of current pre-trained MLIPs and establish benchmarks for future model development.

This detailed analysis of PES features is particularly important because inaccuracies can manifest as unphysical structures during geometry optimisations or mispredictions of material properties sensitive to PES details. Previous studies have indicated challenges in the generalisation capabilities of these models, including a tendency to underestimate PES curvature and produce inaccurate results for configurations far from equilibrium.

The presented methodology provides a means to visualise and understand these inaccuracies, offering a more informative assessment than traditional validation metrics. Ultimately, this work contributes to the reliable application of MLIPs in computational materials science, paving the way for more accurate modelling of complex material behaviours and accelerating the discovery of novel materials.

Evaluating machine learned potential accuracy via two-dimensional potential energy surface slices along Wyckoff positions reveals significant improvements over prior generations

Two-dimensional slices of the potential energy surface were constructed to evaluate the accuracy of machine learned interatomic potentials. Atomic positions were systematically varied along selected Wyckoff degrees of freedom within a fixed crystal symmetry to generate these surfaces. This methodology allowed for a direct visual comparison between the predictions of interatomic potentials and density functional theory calculations.

The research focused on probing predicted energy landscapes, enabling the identification of potential artifacts arising from unique local environments within the crystal structure. Researchers employed universal pre-trained potentials, including MACE, CHGNet, M3GNet, ORB, and SevenNet architectures, to model the potential energy surfaces.

These machine learned interatomic potentials were assessed for their ability to accurately reproduce detailed features of the energy landscape, such as local minima and saddle points. The study moved beyond traditional benchmarks based on validation errors and materials discovery tasks, instead focusing on the fidelity of the potentials in capturing the overall potential energy surface topology.

The methodology involved calculating energies using both the machine learned potentials and density functional theory, providing a benchmark for comparison. By restricting the analysis to symmetry-defined planes within the potential energy landscape, the team facilitated a clear visualisation of discrepancies between the different methods.

This approach enabled a detailed assessment of each potential’s strengths and limitations in capturing the complex features of the energy landscape, offering insights into their physical accuracy and providing valuable data for future model development. The work systematically examined the predicted energy landscapes to understand how well these potentials capture the underlying physics of material systems.

Potential energy surface accuracy of machine-learned interatomic potentials for tungsten dinitride is crucial for reliable simulations

Researchers evaluated the accuracy of machine learned interatomic potentials (MLIPs) by probing predicted energy landscapes using two-dimensional potential energy surface slices. These slices were constructed by varying atomic positions along selected Wyckoff degrees of freedom within fixed crystal symmetry, enabling direct comparison with density functional theory (DFT) calculations.

Analysis focused on capturing local minima, saddle points, and the overall potential energy surface topology, offering benchmarks for future model development. For tungsten dinitride (W2N3), a material not represented in the training corpus of the tested MLIPs, most models accurately identified the local energy minimum at the equilibrium position and captured the general curvature of the potential energy surface.

However, SevenNet0 and CHGNet exhibited unphysical drops in predicted energy in regions of significant atomic overlap, a common issue for models lacking auxiliary repulsion terms. MACE_medium also displayed artifacts, while ORB v2 predicted a narrower energy range, differing by approximately 8 eV/atom between the highest and lowest energies.

Investigation of AlTiN3, a structure with six Wyckoff degrees of freedom, revealed pronounced discrepancies between MACE_MPA-0 and other MACE models, as well as DFT calculations. MACE_MPA-0 predicted a distinct local minimum where DFT and other MACE models showed no stable configurations, demonstrating a model-specific artifact capable of distorting geometry optimization.

Conversely, MACE_OMAT-0 yielded an energy landscape closely matching the DFT reference, highlighting the benefits of improved training data and architectural refinements. Analysis of Cu2O8S4 further demonstrated discrepancies, particularly between MACE_MATPES-PBE and other models during symmetry-constrained relaxation, showcasing the importance of diverse training datasets and model architectures for applications like crystal structure search. The generated two-dimensional potential energy surface plots facilitate visual interpretation and efficient analysis of model behavior, enabling tracking of modifications to training data or model architecture.

Potential energy surface fidelity defines machine-learned interatomic potential accuracy, and its improvement is crucial for reliable simulations

Researchers have developed a systematic method for evaluating the accuracy of machine learned interatomic potentials (MLIPs) by probing the detailed features of potential energy surfaces. This approach constructs two-dimensional slices of the potential energy surface, varying atomic positions along specific degrees of freedom within a fixed crystal symmetry, allowing for direct comparison with density functional theory (DFT) calculations.

The analysis reveals that while current pre-trained MLIPs generally capture local energy minima, discrepancies exist in their ability to accurately reproduce the overall shape of the potential energy surface. The investigation of several crystal structures, including AlTiN3 and Cu2O8S4, highlighted variations in performance between different MLIP architectures and training datasets.

Specifically, the study identified spurious local minima predicted by certain models, such as MACE_MPA-0 for AlTiN3, which could lead to incorrect results during geometry optimization. Conversely, newer models like MACE_OMAT-0 demonstrated improved accuracy, closely matching DFT reference landscapes, and indicating the benefits of enhanced training data and architectural improvements.

The findings also showed that ORB v2 predicts a narrower energy range compared to other models, and exhibits less dramatic energy drops in regions of close atomic proximity. The authors acknowledge that the analysis is limited to specific two-dimensional slices of the potential energy surface and may not fully capture the complexity of higher-dimensional landscapes.

Future research could extend this methodology to explore more degrees of freedom and a wider range of materials. These results underscore the importance of carefully evaluating MLIPs beyond simple validation errors, and suggest that employing models trained on diverse datasets and utilising different architectures can improve the reliability of large-scale simulations and crystal structure searches.