Researchers are grappling with a critical 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, along with Bruno Henrique Nunes, Flavu Cipcigan, Rodrigo Neumann Barros Ferreira, and others, present a new equivariant message passing architecture called Geodite that overcomes the limitations of existing interatomic potentials. Their work replaces computationally expensive tensor products with a physically-informed approach, dramatically speeding up simulations while achieving accuracy comparable to, and in some cases exceeding, state-of-the-art methods for predicting things like material stability and thermal conductivity. This breakthrough is expected to enable faster, larger-scale atomic simulations and high-throughput materials screening, which were previously considered computationally impossible.
Geodite speeds up atomic simulations with homoscedasticity
Scientists have announced Geodite, a new equivariant message transfer architecture for accelerating atomic simulations. This breakthrough addresses significant limitations in the computational costs associated with current machine-learned interatomic potentials, i.e., accurate representations of atomic interactions. The researchers replaced the computationally expensive Klebsch-Gordan tensor product with a new approach that incorporates physical priors to ensure the creation of a smooth and reliable potential energy surface. Trained on a large dataset of inorganic crystals from the Materials Project, called MPtrj, Geodite-MP shows comparable accuracy to leading methods in predicting material stability, thermal conductivity, and phonon-derived properties.
Experiments show that Geodite-MP achieves competitive performance on benchmarks that assess the model’s ability to identify the material’s ground state and reproduce its vibrational properties, while providing significant speed improvements. Specifically, the new architecture runs three to five times faster than models achieving similar accuracy, an important advance for large-scale simulations and high-throughput screening. The team carefully validated Geodite-MP across a variety of tasks, confirming its predictive accuracy and computational efficiency when compared to other basic machine-learned interatomic potentials trained on the same dataset. This speedup is achieved by avoiding the computational bottlenecks inherent in traditional equivariant architectures that rely on tensor products.
This study proves that Geodite-MP not only accurately predicts material properties but also produces smooth bonding curves with accurate short-range repulsion, a feature lacking in some of the existing machine learning possibilities. Analysis of diatomic systems reveals that Geodite-MP avoids the nonphysical attraction forces often observed in other models, ensuring more realistic simulations. Furthermore, nanosecond-scale molecular dynamics simulations of 49 solid electrolytes demonstrated the long-term stability and accuracy of Geodite-MP, which closely mirrored the 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.
This research enables faster, larger-scale atomistic simulations and high-throughput screening for which computational power was previously prohibitive, and is expected to transform materials science and accelerate scientific discovery. This innovative architecture represents an important step toward creating truly fundamental machine learning interatomic potentials that can tackle increasingly complex scientific challenges. . Experiments reveal that Geodite-MP, trained on the Materials Project trajectory dataset of inorganic crystals, achieves this performance while running approximately 5 times faster than Allegro-MP-L, 3 times faster than Eqnorm MPtrj, and 2.5 times faster than NequIP-MP-L. This breakthrough significantly reduces computational costs without sacrificing predictive power.
The researchers systematically validated Geodite-MP across a variety of tasks compared to other force-field machine learning interatomic potentials (fMLIPs) trained on the same dataset. The results demonstrate excellent accuracy on established benchmarks such as Matbench Discovery and MDR, which assess a model’s ability to identify a material’s ground state and reproduce its vibrational properties. Analysis of the diatomic system showed that Geodite-MP produces smooth binding curves with correct short-range repulsion, an important feature for stable simulations, unlike some other MLIPs that exhibit unphysical short-range attraction. This study successfully incorporates a physical prior distribution to ensure a smooth and well-behaved potential energy surface.
Tests demonstrated that Geodite-MP maintains stability and accurately reproduces the local structure observed in ab initio molecular dynamics (AIMD) simulations during 1ns simulations of 49 solid electrolytes (SSEs) at various temperatures. This architecture extends Geometric Tensor Networks (GotenNet) with key physical priors and inductive biases to enhance its robustness to large-scale simulations. Geodite constructs scalar edge features from node features using element-wise multiplication with distance-dependent radial features to enhance the representation of atomic interactions. The Geodite architecture uses forces and stresses obtained through backpropagation following a conservative approach to predict atomic positions and energies from species via an equivariant message-passing neural network. Initial latent features are generated by embedding a layer encoding the atom type, then aggregated with neighboring atoms weighted by a radial basis embedding and passed through a multilayer perceptron to generate initial node features. Messages are constructed from scalar node and edge features and decomposed into components that update invariant and equivariant representations, ultimately enabling faster large-scale atomic simulations and high-throughput screening.
Geodite enables efficient and accurate atomic material prediction
Researchers have developed Geodite, a new equivariant message passing architecture for atomic simulations that offers an attractive balance between accuracy and computational efficiency. This innovative model addresses a major limitation of existing methods, the rapid scaling of computational cost with angular resolution, by replacing computationally expensive tensor products with a more streamlined approach. Geodite-MP, trained on a large dataset of inorganic crystal orbitals from the Materials Project, demonstrates performance comparable to state-of-the-art methods in predicting material stability, thermal conductivity, and phonon-derived properties. Geodite can achieve high accuracy while maintaining faster inference speeds, unlocking opportunities for large-scale atomistic simulations and high-throughput materials screening that were previously computationally infeasible. The model was tested in various solid electrolytes, including lithium, sodium, and copper-based ionic conductors, over various temperatures and system sizes, demonstrating its robustness and applicability to various material systems. The authors acknowledge that there are limitations associated with the specific training dataset and model parameters, but suggest that future work may consider extending the training data and improving the architecture to further improve performance and generalizability.
