Abstract
Machine Learning (ML) has the potential to accelerate discovery of new materials and shed light on useful properties of existing materials. A key difficulty when applying ML in Materials Science is that experimental datasets of material properties tend to be small. In this work we show how material descriptors can be learned from the structures present in large scale datasets of material simulations; and how these descriptors can be used to improve the prediction of an experimental property, the energy of formation of a solid. The material descriptors are learned by training a Graph Neural Network to regress simulated formation energies from a material's atomistic structure. Using these learned features for experimental property predictions outperforms existing methods that are based solely on chemical composition. Moreover, we find that the advantage of our approach increases as the generalization requirements of the task are made more stringent, for example when limiting the amount of training data or when generalizing to unseen chemical spaces.
Introduction
Predicting the properties of novel materials before laboratory synthesis remains a central challenge in materials science. While Machine Learning (ML) offers a promising path forward, it is frequently hampered by the "small data" problem. High-fidelity experimental datasets are often limited to a few thousand compounds due to the high cost and time required for laboratory measurements.
Existing computational approaches often rely on large databases of quantum mechanical simulations, such as Density Functional Theory (DFT). However, these simulations contain systematic errors and typically model idealized conditions. Furthermore, many current ML models rely solely on chemical compositions (e.g., $Fe_2O_3$). This approach fails to account for structural polymorphs—materials with identical formulas but different atomic arrangements—which can exhibit vastly different physical properties. This work introduces Atomistic Graph Networks (AGN) to bridge the gap between large-scale simulation data and sparse experimental measurements by incorporating structural information.
Methodology
The core of the proposed approach is the Atomistic Graph Network (AGN). Unlike composition-based models, AGN represents materials as graphs where atoms serve as nodes and interatomic distances/bonds define the edges. This allows the model to capture the 3D geometric relationships essential for distinguishing polymorphs.
The architecture utilizes a Transfer Learning Trunk, where a shared Graph Neural Network (GNN) learns general material embeddings from over 120,000 simulation data points from the Materials Project. The model employs a multi-task learning strategy, using separate "heads" to predict simulation and experimental values simultaneously. To address the challenge of unknown experimental structures, the authors developed a Phase Diagram Decomposition scheme. This method synthesizes an embedding for an experimental compound by performing a weighted interpolation of its neighbors' embeddings within a simulated phase diagram.
Results and Discussion
The AGN model was evaluated against several baselines, including DFT calculations, ElemNet (composition-only), and Automatminer. The results demonstrate that incorporating structural information provides a powerful inductive bias, enabling the model to learn physically plausible rules that composition-only models overlook.
State-of-the-Art Accuracy
The AGN+ model achieved a Mean Absolute Error (MAE) of 0.059 eV/atom on experimental formation energy, a significant improvement over previous composition-based benchmarks (~0.07 eV/atom). This suggests that high-dimensional learned embeddings are far more effective than simple scalar corrections to DFT values.
Data Efficiency and Generalization
A key finding is the model's remarkable data efficiency. When the experimental training set was reduced to just 157 points, the AGN approach maintained a 17% accuracy advantage over the next best method. Furthermore, in "hold-out" tests designed to measure generalization to unseen chemical spaces (e.g., removing all copper-containing compounds), the AGN model exhibited a 31% lower error than unstructured models. This indicates that the model learns underlying chemical trends rather than merely memorizing formulas.
Conclusion
Atomistic Graph Networks represent a significant step forward in leveraging large-scale simulation data to enhance experimental property prediction. By moving beyond chemical composition to include 3D structural information, the AGN model achieves superior accuracy and robustness, particularly in data-scarce regimes. While this study focused on formation energy, future work will explore extending this framework to more complex properties like band gaps and thermal conductivity, as well as refining interpolation strategies for cases where simulated structures may deviate from experimental reality.
Summary updated