Machine-learning accelerated identification of exfoliable two-dimensional materials
About this tool
This tool allows users to upload the bulk crystal structure in several standard formats (or to choose from a few examples), and then layered structures are identified based on geometrical criteria. Finally, after generating features vectors representing the crystal structure, the tool uses a machine learning model to see if the crystal structure can be exfoliated or have high binding energy.
The output page includes relevant information on the structure (interactive visualizations of the bulk multilayer) and whether the structure is suitable for exfoliation or not based on the geometrical criteria. If yes, the corresponding two-dimensional layers are displayed. In addition, the machine-learning model is run to predict if the structure might actually have a low binding energy, and results are displayed.
- The tool computes representing features vectors of the crystal structure in real-time, and this step might take some time.
- The tool will consider two atoms A and B bonded if their distance is smaller than rA + rB + Δ, where rA and rB are the corresponding van-der-Waals radii from the following paper: S. Alvarez, A cartography of the van der Waals territories, Dalton Trans. 42, 8617-8636 (2013). For the values of Δ, we consider five equally spaced values between 1.1 Å and 1.5 Å to consider possible uncertainties in the values of the van-der-Waals radii.
This tool uses the ASE and pymatgen libraries for structure manipulation, matminer for feature generation, and shap for model explanation.
The tool uses the lowdimfinder code for the initial geometrical analysis, originally developed in N. Mounet et al., Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds, Nature Nanotech. 13, 246-252 (2018). The DFT data for the binding energies used to train our ML model was also extracted from the same paper.
The tool is based upon the tools-barebone framework developed by the Materials Cloud team.
We acknowledge mobility funding from Swiss National Science Foundation (SNSF) (Mobility fellowship number of P1ELP2_195110).
Upload your structure
Otherwise, pick an example
How to cite
If you use this tool, please cite the following work:
- M. T. Vahdat, K. A. Varoon, and G. Pizzi, Machine-learning accelerated identification of exfoliable two-dimensional materials, Machine Learning: Science and Technology, 3, 4 (2022). DOI:10.1088/2632-2153/ac9bca
- You might also want to cite the ASE, pymatgen, matminer and shap libraries that are used internally by the tool, as well as N. Mounet et al., Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds, Nature Nanotech. 13, 246-252 (2018) where the geometrical-screening code was first developed, and from which the DFT data for the binding energies was extracted to train our model.
You can access the source code of this tool on its GitHub project page.