研究目的

To discover the chemical composition of materials likely to exhibit layered and two-dimensional phases but have yet to be synthesized, and to provide a roadmap for the synthesis community by combining physics with machine learning on experimentally obtained data.

研究成果

The study demonstrates that physics-based machine learning can enable the prediction of the existence of layered phases of binary and ternary materials using the chemical formula, providing the first complete mapping of layered materials. The DFT calculations on a subset of predicted materials suggest most are indeed mechanically stable and potentially synthesizable. The precision of the model is higher than most human experts in the field and orders of magnitude faster. The application of semi-supervised learning shows how non-labeled data can be integrated into physics-based machine learning, offering improvements over supervised learning in the present case.

研究不足

The accuracy of the assumption that all chemical compositions lacking a layered phase in the databases definitely exhibit no stable layered phase is dependent on the completeness of the databases and is difficult to quantify. The study also notes that the development of a regression model to predict DFT band gap values is expected to result in relatively large errors when compared with experimental measurements.