Geodite Achieves Accurate Equivariant Interatomic Potentials Without Tensor Products

Researchers are tackling a significant bottleneck in materials science: the computational cost of accurately simulating atomic interactions. Thiago Reschützegger, Sarp Aykent, and Gabriel Jacob Perin, from IBM Research Rio de Janeiro and Microsoft Redmond, alongside Bruno Henrique Nunes, Flaviu Cipcigan, and Rodrigo Neumann Barros Ferreira et al., present a new equivariant message-passing architecture called Geodite that overcomes limitations of existing interatomic potentials. Their work replaces computationally expensive tensor products with a physically informed approach, achieving comparable , and in some cases superior , accuracy to state-of-the-art methods for predicting materials stability, thermal conductivity, and more, all while dramatically increasing simulation speed. This breakthrough promises to unlock faster, large-scale atomistic simulations and high-throughput materials screening previously considered computationally prohibitive.

Geodite accelerates atomistic simulations with equivariance

Scientists have unveiled Geodite, a novel equivariant message-passing architecture poised to accelerate atomistic simulations. This breakthrough addresses a critical limitation in current machine-learned interatomic potentials, the computational cost associated with accurately representing atomic interactions. The research team replaced computationally expensive Clebsch-Gordan tensor products with a new approach, incorporating physical priors to ensure the creation of smooth and reliable potential energy surfaces. Trained on a vast dataset of inorganic crystals from the Materials Project, termed MPtrj, Geodite-MP demonstrates accuracy comparable to leading methods in predicting material stability, thermal conductivity, and phonon-derived properties.
Experiments show Geodite-MP achieves competitive performance on benchmarks assessing a model’s ability to identify a material’s ground state and reproduce its vibrational properties, while simultaneously offering significant speed improvements. Specifically, the new architecture runs 3, 5× faster than models achieving similar accuracy, a crucial advancement for large-scale simulations and high-throughput screening. The team meticulously validated Geodite-MP across diverse tasks, confirming its predictive accuracy and computational efficiency when compared to other foundational machine-learned interatomic potentials trained on the same dataset. This enhanced speed is achieved by avoiding the computational bottlenecks inherent in traditional equivariant architectures reliant on tensor products.

The study establishes that Geodite-MP not only predicts material properties accurately but also generates smooth binding curves with correct short-range repulsion, a feature lacking in several existing machine-learning potentials. Analysis of diatomic systems revealed that Geodite-MP avoids unphysical attractions often observed in other models, ensuring more realistic simulations. Furthermore, nanosecond-scale molecular dynamics simulations of 49 solid-state electrolytes demonstrated the long-term stability and accuracy of Geodite-MP, closely mirroring results obtained from ab initio molecular dynamics. By combining predictive power, computational efficiency, and physical realism, Geodite opens new avenues for exploring complex materials and accelerating the discovery of novel compounds.
The work enables faster, large-scale atomistic simulations and high-throughput screening that were previously computationally prohibitive, promising to transform materials science and accelerate scientific discovery. This innovative architecture represents a significant step towards creating truly foundational machine-learned interatomic potentials capable of tackling increasingly complex scientific challenges. . Experiments revealed that Geodite-MP, trained on the Materials Project trajectories dataset of inorganic crystals, delivers this performance while running approximately 5× faster than Allegro-MP-L, 3× faster than Eqnorm MPtrj, and 2.5× faster than NequIP-MP-L. This breakthrough delivers a substantial reduction in computational cost without sacrificing predictive power.

Researchers systematically validated Geodite-MP across diverse tasks, comparing it to other force field machine learning interatomic potentials (fMLIPs) trained on the same dataset. Results demonstrate competitive accuracy on established benchmarks like Matbench Discovery and MDR, which assess a model’s ability to identify a material’s ground state and reproduce its vibrational properties. Analysis of diatomic systems showed Geodite-MP produces smooth binding curves with correct short-range repulsion, a critical feature for stable simulations, unlike several other MLIPs exhibiting unphysical short-range attraction. The work successfully incorporates physical priors to ensure smooth, well-behaved potential energy surfaces.

Tests prove that Geodite-MP maintains stability and accurately reproduces local structure observed in ab initio molecular dynamics (AIMD) simulations during 1ns simulations of 49 solid-state electrolytes (SSEs) at various temperatures. The architecture extends the Geometric Tensor Network (GotenNet) with key physical priors and inductive biases, enhancing its robustness for large-scale simulations. Geodite constructs scalar edge features from node features using element-wise multiplication with distance-dependent radial features, enhancing the representation of interatomic interactions. The Geodite architecture predicts energies from atomic positions and species via an equivariant message-passing neural network, with forces and stresses obtained through backpropagation following a conservative approach. Initial latent features are produced by embedding layers encoding atomic types, which are then aggregated with neighboring atoms weighted by radial basis embeddings and passed through multilayer perceptrons to produce initial node features. Messages are constructed from scalar node features and edge features, decomposed into components that update invariant and equivariant representations, ultimately enabling faster large-scale atomistic simulations and high-throughput screening.

Geodite unlocks efficient, accurate atomistic material predictions

Researchers have developed Geodite, a new equivariant message-passing architecture for atomistic simulations that offers a compelling balance between accuracy and computational efficiency. This innovative model addresses a key limitation of existing methods, the steep scaling of computational cost with angular resolution, by replacing computationally expensive tensor products with a more streamlined approach. Trained on a large dataset of inorganic crystal trajectories from the Materials Project, Geodite-MP demonstrates performance competitive with state-of-the-art methods in predicting materials stability, thermal conductivity, and phonon-derived properties. Geodite’s ability to achieve high accuracy while maintaining faster inference speeds unlocks opportunities for large-scale atomistic simulations and high-throughput materials screening that were previously computationally prohibitive. The model was tested on a diverse set of solid-state electrolytes, encompassing lithium-, sodium-, and copper-based ionic conductors, across a range of temperatures and system sizes, demonstrating its robustness and applicability to various materials systems. While the authors acknowledge limitations related to the specific training dataset and model parameters, they suggest future work could explore extending the training data and refining the architecture to further enhance performance and generalizability.