PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
February 21, 2026Next Materials2 citationsOpen Access

Machine-learning-accelerated band gap prediction from chemical composition with near-experimental accuracy

View Full Paper
CGCésar Gabriel Vera de la GarzaSFSerguei Fomine

Key Points

  • The research aims to predict electronic band gaps using only chemical composition through a machine learning approach.
  • Developed a machine learning model using XGBoost to forecast band gaps from compositional data.
  • Employed Bayesian optimization and zero-overlap external validation for enhanced accuracy.
  • Analyzed learning curves to assess data efficiency and required training samples.
  • Conducted SHAP interpretability analysis to understand influential predictive features.
  • Achieved a mean absolute error of 0.424 eV for predicting band gaps on unseen compounds.
  • Demonstrated robust generalization capability with external validation.
  • Reached useful performance (MAE < 0.5 eV) using only 1381 training samples.
  • Confirmed that valence electron configuration is the primary predictor for band gap.

Abstract

Accurate prediction of electronic band gaps from chemical composition alone remains a formidable challenge in materials informatics, with existing approaches often limited by reliance on internal validation or computational targets that inherit density functional theory's systematic errors. Here, we present a machine learning framework that achieves high accuracy in predicting experimental band gaps using only compositional features. Through rigorous Bayesian optimization and zero-overlap external validation, our XGBoost model attains a mean absolute error of 0.424 eV on 1885 completely unseen compounds, demonstrating robust generalization beyond its training distribution. Comprehensive learning curve analysis reveals remarkable data efficiency, with the model achieving useful performance (MAE < 0.5 eV) with only 1381 training samples and showing near-linear computational scaling. SHAP interpretability analysis confirms the model captures physically meaningful relationships, with valence electron configuration emerging as the dominant predictor, aligning with established band structure principles. This work establishes composition-based machine learning as a powerful tool for high-throughput materials screening, accelerating the discovery of functional materials for energy conversion and electronic applications. • Achieves 0.42 eV MAE for experimental band gap prediction from composition alone. • Demonstrates robust generalization via rigorous zero-overlap external validation. • Model shows high data efficiency, reaching useful accuracy with only ∼1400 samples. • SHAP analysis confirms physically meaningful, interpretable model decisions. • Enables millisecond, experimental-quality screening for novel materials discovery.

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Garza et al. (2026) studied this question.

synapsesocial.com/papers/69994b01873532290d01f54dhttps://doi.org/10.1016/j.nxmate.2026.101728
Ask AI
Helpful
Bookmark
Share
View Full Paper