Recent advances in machine learning (ML) have demonstrated significant strides in predicting the band gaps in complex materials known as perovskites. A study investigated the use of ML models to predict direct and indirect band gaps for both double and layered perovskites, highlighting the promising accuracy achievable with modern predictive techniques.
Utilising datasets comprising 1,493 double perovskites and 491 layered perovskites, the research conducted a comprehensive examination employing four ML models: Support Vector Regression (SVR), Random Forest Regression (RFR), Gradient Boosting Regression (GBR), and XGBoost. Feature selection played a vital role in enhancing the models' performance, with initial selections drawn from elemental data, encompassing a range of 41 features derived from the periodic table and material project databases.
The SVR model emerged as the top performer in predicting both direct and indirect band gaps, particularly for double perovskites, where a larger dataset reinforced the model's predictive accuracy. The prediction error metrics for the SVR model, such as Mean Absolute Error (MAE) and Mean Squared Error (MSE), recorded some of the lowest values, indicating high precision in its outputs. For layered perovskites, although data was less abundant, SVR still provided comparatively accurate results.
A noteworthy aspect of the study was the exploration of cross-validation (CV) techniques, specifically k-fold CV, which provided a robust means of evaluating model performance across unseen data, thereby minimising overfitting. This methodological choice highlighted the ability of SVR to generalise effectively in the context of predicting perovskite properties, yielding favourable R-squared values and minimal error margins.
Feature importance was another focal point of the research. Tools such as Local Interpretable Model-Agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), and Permutation Importance (PI) were employed to discern the key features influencing band gap predictions. The "Valence (std)" feature consistently appeared as a crucial determinant in band gap prediction across various models and datasets. The study also explored the impact of derivative discontinuity, revealing its potential to significantly enhance prediction accuracy when included as a feature.
Further statistical analysis, including Pearson correlation coefficients, helped affirm the role of electronic properties, such as valence and group, as significant predictors. This reinforced the utility of ML techniques in circumventing traditional complexity associated with predicting electronic properties of materials.
In parallel, comparative analysis involving different Artificial Neural Network (ANN) architectures was conducted to refine model performance further. The ANN model integrated with Biogeography-Based Optimization (BBO) outperformed other configurations, including Ant Colony Optimization (ACO), in predicting the compressive strength of Waste Marble Concrete (WMC), showcasing a higher degree of precision and reduced computational error.
This ANNs’ performance was assessed using a variety of statistical measures and visual analyses, including radar plots and Taylor diagrams, evidencing its efficacy in correlation and error minimisation.
Overall, the research underscores the significant potential of machine learning approaches in material science, particularly in enhancing the predictability of complex physical properties such as band gaps in perovskites, paving the way for more efficient material design and discovery.
Source: Noah Wire Services