Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals (1812.05055v2)

Abstract: Graph networks are a new ML paradigm that supports both relational reasoning and combinatorial generalization. Here, we develop universal MatErials Graph Network (MEGNet) models for accurate property prediction in both molecules and crystals. We demonstrate that the MEGNet models outperform prior ML models such as the SchNet in 11 out of 13 properties of the QM9 molecule data set. Similarly, we show that MEGNet models trained on $\sim 60,000$ crystals in the Materials Project substantially outperform prior ML models in the prediction of the formation energies, band gaps and elastic moduli of crystals, achieving better than DFT accuracy over a much larger data set. We present two new strategies to address data limitations common in materials science and chemistry. First, we demonstrate a physically-intuitive approach to unify four separate molecular MEGNet models for the internal energy at 0 K and room temperature, enthalpy and Gibbs free energy into a single free energy MEGNet model by incorporating the temperature, pressure and entropy as global state inputs. Second, we show that the learned element embeddings in MEGNet models encode periodic chemical trends and can be transfer-learned from a property model trained on a larger data set (formation energies) to improve property models with smaller amounts of data (band gaps and elastic moduli).

Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals

The paper "Graph Networks as a Universal Machine Learning Framework for Molecules and Crystals" by Chi Chen et al. explores the application of graph networks within the field of material science as a powerful tool for property prediction. It presents the development of the MatErials Graph Network (MEGNet) models as a sophisticated and versatile machine learning framework capable of handling the structural complexity inherent to both molecular and crystalline materials. The MEGNet models are framed as a universal representation for structured data, highlighting their ability to outperform existing models in the domain.

Model Development and Methods

Graph networks serve as the foundation of the MEGNet models. They allow for efficient relational reasoning and the capturing of combinatorial generalization. By treating atoms as nodes and bonds as edges, MEGNet models formulate molecules and crystals as graphical data that can be processed through sequential updates across graph network modules. This modularity and composability enable MEGNet models to capture both local atomic interactions and longer-range relations.

The development further introduces elemental embedding strategies and the inclusion of global state variables such as temperature, pressure, and entropy. Such innovations expand the applicability of MEGNet models, enabling them to predict related properties within a single framework. For example, a unified model was demonstrated for predicting internal energy, enthalpy, and Gibbs free energy of molecules by including global state attributes, a move away from the traditional single-purpose models.

Numerical Results and Achievements

MEGNet models exhibit significant advancements in predicting properties of both molecules and crystals:

• QM9 Dataset: MEGNet models achieved lower mean absolute errors (MAEs) than existing models in predicting 11 out of 13 properties in the QM9 dataset. Notably, the error in predicting zero-point vibrational energy (ZPVE) and electronic spatial extent ($\left<R^2\right>$) required more graph network blocks, illustrating their adaptability to capture detailed interactions.
• Materials Project Dataset: For crystalline materials, MEGNet models outperformed previous models like CGCNN and SchNet in anticipating formation energies, band gaps, and elastic moduli properties. Transfer learning of elemental embeddings from formation energy models was showcased as an effective strategy to enhance models with smaller datasets, exemplifying the adaptability of graph networks.

Implications and Future Developments

The broader implications of this research are substantial for the fields of chemistry and materials science. The demonstration of graph networks' capabilities in unifying predictions for thermodynamically related properties paves the way for more integrated predictive models. This composability and utilitarian nature address significant challenges in ML implementations by maximizing training efficiency and adaptability across diverse property targets.

Furthermore, elemental embeddings derived from MEGNet models provide interpretable trends that align well with known chemical groupings in the periodic table, supporting MEGNet's role as both a predictive and explanatory tool. Future research could extend this model to less understood or highly specialized material types, further broadening the scope of graph network applications.

Conclusion

This work has successfully argued for graph networks as a universal and robust framework for material property prediction. By capturing complex interatomic relations across molecules and crystals, MEGNet offers a significant advancement in bridging discrete domains within materials science and promises new horizons in computational discovery and material innovation. Future explorations and methodological refinements could push the envelope further, enhancing the refinement of predictive capabilities in material structure-property relationships.