Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leakage-Aware Bandgap Prediction on the JARVIS-DFT Dataset: A Phase-Wise Feature Analysis

Gaurav Kumar Sharma

TL;DR

This work addresses the problem of feature leakage in ML bandgap prediction by constructing a leakage-free subset of the JARVIS-DFT data and evaluating a three-phase feature framework. It combines careful data curation, a hierarchical feature expansion, and SHAP-based interpretability to assess how descriptor complexity affects accuracy, finding that $R^2$ saturates around $0.88$–$0.90$ despite richer feature spaces. The principal contribution is the leakage-controlled dataset and the baseline performance metrics, plus a robust finding that dielectric tensor features dominate predictions across phases. The practical impact is a more reliable framework for leakage-aware bandgap regression, guiding future work toward larger, diverse datasets and potentially deep learning approaches for further gains while maintaining interpretability.

Abstract

In this study, we perform a systematic analysis of the JARVIS-DFT bandgap dataset and identify and remove descriptors that may inadvertently encode band-structure information, such as effective masses. This process yields a curated, leakage-controlled subset of 2280 materials. Using this dataset, a three-phase modeling framework is implemented that incrementally incorporates basic physical descriptors, engineered features, and compositional attributes. The results show that tree-based models achieve R2 values of approximately 0.88 to 0.90 across all phases, indicating that expanding the descriptor space does not substantially improve predictive accuracy when leakage is controlled. SHAP analysis consistently identifies the dielectric tensor components as the dominant contributors. This work provides a curated dataset and baseline performance metrics for future leakage-aware bandgap prediction studies.

Leakage-Aware Bandgap Prediction on the JARVIS-DFT Dataset: A Phase-Wise Feature Analysis

TL;DR

This work addresses the problem of feature leakage in ML bandgap prediction by constructing a leakage-free subset of the JARVIS-DFT data and evaluating a three-phase feature framework. It combines careful data curation, a hierarchical feature expansion, and SHAP-based interpretability to assess how descriptor complexity affects accuracy, finding that saturates around – despite richer feature spaces. The principal contribution is the leakage-controlled dataset and the baseline performance metrics, plus a robust finding that dielectric tensor features dominate predictions across phases. The practical impact is a more reliable framework for leakage-aware bandgap regression, guiding future work toward larger, diverse datasets and potentially deep learning approaches for further gains while maintaining interpretability.

Abstract

In this study, we perform a systematic analysis of the JARVIS-DFT bandgap dataset and identify and remove descriptors that may inadvertently encode band-structure information, such as effective masses. This process yields a curated, leakage-controlled subset of 2280 materials. Using this dataset, a three-phase modeling framework is implemented that incrementally incorporates basic physical descriptors, engineered features, and compositional attributes. The results show that tree-based models achieve R2 values of approximately 0.88 to 0.90 across all phases, indicating that expanding the descriptor space does not substantially improve predictive accuracy when leakage is controlled. SHAP analysis consistently identifies the dielectric tensor components as the dominant contributors. This work provides a curated dataset and baseline performance metrics for future leakage-aware bandgap prediction studies.
Paper Structure (31 sections, 1 equation, 7 figures, 4 tables)

This paper contains 31 sections, 1 equation, 7 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Data Preprocessing and Curation
  4. Statistical Analysis for Checking Data Integrity
  5. Distribution Consistency and Chemical-Space Coverage
  6. Physical-Property Range Preservation
  7. Summary and Interpretation
  8. Data Leakage Detection and Mitigation Framework
  9. Feature Risk Categorization
  10. Baseline Performance Estimation
  11. Incremental Feature-Impact Analysis
  12. Machine Learning
  13. Hierarchical Feature Engineering Framework
  14. Phase I: Fundamental Physical Descriptors
  15. Phase II: Engineered Physical Descriptors
  16. ...and 16 more sections

Figures (7)

  • Figure 1: Effect of effective-mass descriptors on model performance. (a) $R^2$ score, (b) mean absolute error (MAE), and (c) mean squared error (MSE) for models trained with and without effective-mass features.
  • Figure 2: Model performance as a function of the number of input features used to identify the optimal feature subset size.
  • Figure 3: Error distribution for the CatBoost model trained using the Phase III feature set, showing the distribution of prediction residuals.
  • Figure 4: Phase-wise parity plots for the best-performing models in Phase I, Phase II, and Phase III, comparing predicted and reference bandgap values. The predicted values align closely with the ideal-fit line across all phases, indicating consistent predictive behavior despite increasing feature complexity.
  • Figure :
  • ...and 2 more figures