Abstract
Accurately predicting the physical and chemical properties of materials remains one of the most challenging tasks in material design. One effective strategy is to construct a reliable data set and use it for training a machine learning model. In this study, we constructed a large-scale material genome database (Matgen) containing 76,463 materials collected from experimentally-observed databases and computed their bandgap properties through Density Functional Theory (DFT) with the Perdew-Burke-Ernzerhof (PBE) functional. We verified the computation method by comparing our results with those from the Materials Project (MP) and Open Quantum Materials Database (OQMD). Based on this comprehensive dataset, we developed a new graph-based deep learning model, CrystalNet, using our Communicative Message Passing Neural Network (CMPNN) framework. CrystalNet outperformed other state-of-the-art prediction models. Further fine-tuning on 1,716 experimental bandgap values (CrystalNet-TL) achieved a superior performance with a mean absolute error (MAE) of 0.77 eV, outperforming pure PBE calculations (1.14–1.45 eV).
Introduction
The search for advanced materials for solar cells, semiconductors, and batteries hinges on understanding the "bandgap"—the energy range where no electron states can exist. Traditionally, researchers rely on expensive experiments or "ab initio" quantum mechanics simulations like Density Functional Theory (DFT). While DFT is a standard tool, it is computationally intensive and suffers from the "bandgap problem," where the common PBE functional systematically underestimates values by more than 1 eV. Existing machine learning (ML) models attempt to bridge this gap but often struggle with datasets dominated by hypothetical materials or architectures that fail to capture complex atomic interactions.
The Matgen Database and Methodology
To address these challenges, we curated the Matgen database, which consists of 76,463 inorganic structures derived from the Inorganic Crystal Structure Database (ICSD) and the Crystallography Open Database (COD). Using the Tianhe-2 supercomputer, we invested approximately 100 million CPU hours to calculate properties such as bandgap, Fermi energy, and formation energy using DFT-PBE.
As shown in Figure 1, Matgen is significantly larger than the experimental subsets of other popular databases like the Materials Project (MP) and OQMD, including over 20,000 unique structures. By focusing on experimentally observed structures, we ensure the model learns from chemically synthesizable materials.
CrystalNet: Communicative Message Passing
We developed CrystalNet, a directed graph neural network based on the Communicative Message Passing Neural Network (CMPNN) framework. Unlike traditional graph models that treat atoms and bonds as static nodes and edges, CrystalNet allows them to "communicate" through a kernel that strengthens interactions during learning. This dynamic approach better mimics the complex ionic and metallic bonding inherent in real crystals, particularly for large unit cells containing more than 80 atoms.
To solve the inherent bias in PBE data, we employed transfer learning (CrystalNet-TL). We pre-trained the model on the massive Matgen PBE dataset and then fine-tuned it on a high-quality set of 1,716 experimental values. This allowed the model to learn the systematic bias of simulations and correct for it using experimental ground truth.
Key Results and Performance
CrystalNet demonstrated superior accuracy, outperforming existing models by more than 29.2% on the Matgen test set. Notably, the fine-tuned CrystalNet-TL model achieved an MAE of 0.77 eV against experimental data, significantly better than the standard PBE simulation error range of 1.14 to 1.45 eV.
As illustrated in Figure 5, CrystalNet-TL achieves the lowest MAE among machine learning approaches and even surpasses the accuracy of standard PBE calculations. Furthermore, when validated against high-accuracy HSE calculations for hypothetical materials, the model maintained its 0.77 eV MAE, proving its ability to generalize beyond its training set. This speed and accuracy allow for material property predictions in milliseconds compared to the thousands of CPU hours required for traditional quantum simulations.
Conclusion
By combining a massive, experimentally-grounded database with a communicative neural network architecture, we have created a tool that provides bandgap predictions with accuracy exceeding traditional physics-based simulations. While the current database relies on PBE functionals, future work will involve expanding the Matgen database with higher-fidelity computations and applying the CrystalNet framework to other critical material properties like thermal conductivity.
The Matgen database and CrystalNet models are publicly available at https://matgen.nscc-gz.cn.
Summary updated