Machine Learning based prediction of noncentrosymmetric crystal materials

Yuqi Song, Joseph Lindsay, Yong Zhao, Alireza Nasiri, Steph-Yves M. Louis, Jie Ling, Ming Hu, Jianjun HuarXiv 2020 · arXiv

TL;DR

This paper presents a machine learning approach to predict whether a crystalline material will be noncentrosymmetric, a crucial property for applications like advanced lasers and quantum computing. By leveraging composition-based features and a Random Forest model, they achieve high accuracy in identifying these materials, paving the way for faster discovery of novel functional materials.

What Problem Does This Paper Address?

Noncentrosymmetric materials are essential for a wide range of advanced technologies, including lasers, quantum computing, and optical data storage. These materials possess a unique crystal structure that lacks a center of symmetry, enabling them to interact with light in specific ways (e.g., frequency modification). However, experimentally discovering and synthesizing new noncentrosymmetric materials is a significant challenge. The vast number of possible chemical compositions and the complexity of crystal structure prediction make traditional trial-and-error methods inefficient. Current computational methods, while powerful, are often too computationally expensive for large-scale screening. This paper aims to bridge this gap by developing a machine learning model that can predict noncentrosymmetry directly from a material's chemical composition, thereby accelerating the discovery process.

Methods

The researchers employed a two-pronged machine learning approach:

  • Feature Engineering: They utilized the matminer library to extract 132 elemental features (Magpie descriptors) from the chemical composition of materials. These features capture various statistics of elemental properties, such as atomic number, melting temperature, and valence electron counts.
  • Machine Learning Models: Two primary models were evaluated:
  • Random Forest (RF) Classifier: This ensemble learning method builds multiple decision trees and aggregates their predictions. It's known for its robustness and good performance in materials science applications.
  • Multi-Layer Perceptron (MLP) Neural Network: A deep learning model with fully connected layers, designed to learn complex patterns from the data.

Hyperparameter tuning for the RF models was performed using Bayesian optimization with the hyperopt package to find optimal settings for parameters like n_estimators, max_depth, and max_features.

Datasets

The study utilized the following datasets:

  • Materials Project (MPF) Dataset: This forms the primary dataset, containing 82,506 material compositions. It was curated by removing entries with ambiguous symmetry assignments, resulting in 60,587 noncentrosymmetric and 21,919 centrosymmetric samples.
  • MP3 Dataset: A subset of the MPF dataset, specifically focusing on materials with only 3 elements. This dataset comprises 39,726 samples (30,762 centrosymmetric, 8,964 noncentrosymmetric).
  • Borates315 Dataset: A hold-out test set consisting of 315 borate compounds, selected because borates are known to often exhibit nonlinear optical properties and typically contain 3 elements.

Key Insights and Ideas

  • Composition is Key: The study demonstrates that predicting noncentrosymmetry can be effectively achieved using only the chemical composition, bypassing the need for complex structural information. This significantly broadens the scope of materials that can be screened.
  • Random Forest Outperforms DNN: For this specific task, the Random Forest classifier consistently outperformed the Deep Neural Network (DNN) model in terms of precision and accuracy, suggesting that for this type of structured, feature-engineered data, RF is a more suitable choice.
  • 3-Element Materials Excel: Training the RF model on a dataset of 3-element materials (MP3) and then testing it on a hold-out set of 3-element borates yielded higher prediction performance than using the broader MPF dataset. This highlights the potential for specialized models for specific material classes.
  • Valence Electrons are Crucial: Feature importance analysis revealed that descriptors related to valence electrons (e.g., number of valence electrons, distribution of valence s and d electrons) are the most significant predictors of noncentrosymmetry. This aligns with physical intuition, as valence electrons are directly involved in chemical bonding and optical properties.

Key Results

  • High Accuracy: The best performing Random Forest model achieved an accuracy of 84.8% on the MPF dataset using 10-fold cross-validation.
  • Improved Performance with 3-Element Data: When trained on the MP3 dataset (3-element materials), the RF model achieved an even higher accuracy of 86.9% on a 3-element test set.
  • Extrapolative Prediction: Hold-out experiments on the borate dataset showed acceptable performance (AUC of 0.71 for MPF-trained and 0.68 for MP3-trained models), indicating the model's ability to generalize to unseen material families, albeit with reduced accuracy compared to cross-validation.
  • Screening Novel Materials: The developed RF model was used to screen 2 million hypothetical materials, identifying top candidate noncentrosymmetric materials, including 20 with 2-4 elements and 20 borate candidates.

Limitations and Future Work

  • Extrapolation Performance: While the model shows promise for extrapolation, the performance on hold-out datasets (like borates) is lower than cross-validation results. This is attributed to potential differences in data distribution and the inherent challenges of predicting properties for entirely new material classes.
  • Hypothetical Materials: The identified novel materials are currently hypothetical. Experimental verification is needed to confirm their noncentrosymmetric nature and potential applications.
  • Future Work: The authors suggest that their model can be applied to discover novel nonlinear optical materials by conducting large-scale screenings. Further research could explore more complex feature sets, different machine learning architectures, and the integration of structural information when available. The paper also notes that more detailed prediction results can be provided upon request.

Summary updated

BibTeX
@misc{song2020machine,
  author = {Yuqi Song and Joseph Lindsay and Yong Zhao and Alireza Nasiri and Steph-Yves M. Louis and Jie Ling and Ming Hu and Jianjun Hu},
  title = {{Machine Learning based prediction of noncentrosymmetric crystal materials}},
  year = {2020},
  eprint = {2002.11295},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2002.11295},
}
Figure