CMML: Contrastive Materials Metric Learning

This is the offitial imprementation of Self-supervised learning of materials concepts from crystal structures via deep neural networks.
We named this approach as CMML: Contrastive Materials Metric Learning

Table of Contents

1. PyTorch implementation of proposed method

• Please see src/ directory.

2. Interactive materials map visualisation
   
3. Resulting embeddings with a Jupyter notebook for local neighbourhood analysis
4. Local neighbourhood search results
5. List of the target materials from the Materials Project
6. Citation

---

1. Interactive materials map visualisation

• Corresponding to the results of the "Global distribution analysis" section in the main text.
• You can interactively explore our materials map on your web browser.

Click to open:
➡️ Full map (pretty heavy: ~20MB)

➡️ Reduced map (randomly sampled to 20%)

2. Jupyter notebook examples

Mapping your CIF data with pre-trained model

• Example of exporting embeddings with your data (.CIF files) and ploting in our materials map.

• examples/export_embedding_from_CIF.ipynb:
  ➡️ Open in Google Colab

local neighbourhood analysis

• Corresponding to the results of the "Local neighbourhood analysis" section in the main text and Appendix A in the Supplementary Information (SI).

• You can analyse our embeddings through the local neighbour search on your web browser.

• examples/neighbour_analysis_of_MP.ipynb:
  ➡️ Open in Google Colab

3. Local neighbourhood search results

• Corresponding to the results of the "Local neighbourhood analysis" section in the main text and Appendix A in the SI.
• You can see the extended lists of the top-1000 neighbourhoods for Tables and SI Tables shown below.

Table	Method
Table 1 (Hg-1223)	Ours, ESM, SCM
Table 2 (LiCoO2)	Ours, ESM, SCM
SI Table S1 (Cr2Ge2Te6)	Ours, ESM, SCM
SI Table S2 (Sm2Co17)	Ours, ESM, SCM
SI Table S3 (Hg-1223)	Ours, Crystal structure encoder, XRD encoder
SI Table S4 (LiCoO2)	Ours, Crystal structure encoder, XRD encoder

4. List of the target materials from the Materials Project

• The full list of the 122,543 target materials used in this study is available: metadata.csv
• You can collect the materials data using the shown Materials Project IDs via the Materials Project APIs.

5. Citation

@article{Suzuki_materials_concepts_learning_2022,
doi = {10.1088/2632-2153/aca23d},
url = {https://dx.doi.org/10.1088/2632-2153/aca23d},
year = {2022},
month = {dec},
publisher = {IOP Publishing},
volume = {3},
number = {4},
pages = {045034},
author = {Yuta Suzuki and Tatsunori Taniai and Kotaro Saito and Yoshitaka Ushiku and Kanta Ono},
title = {Self-supervised learning of materials concepts from crystal structures via deep neural networks},
journal = {Machine Learning: Science and Technology},
}