Data and Trained Models for: Reliability Assessment of Machine Learning Models for Single-Atom Alloy Catalyst Screening

This repository contains the supplementary data, trained machine learning models, and preprocessed datasets associated with the manuscript: "Reliability Assessment of Machine Learning Models for Single-Atom Alloy Catalyst Screening: A Chemistry-Aware Validation Framework for Methane C–H Activation Barrier Prediction." Dataset and Contents: models/: Serialized trained model weights (.joblib files) for the Random Forest, Deep Forest (gcForest), and Multi-Layer Perceptron (MLP) architectures. This also includes the custom LeakageSafePreprocessor object. All models were fitted strictly on the leakage-safe deployed feature matrix (141 descriptors). Supplementary_Data/: Detailed out-of-fold cross-validation summaries (LOHO, LODO, LOSO), descriptor group ablation results, the 141 deployed descriptors list, and the comprehensive 496-row descriptor provenance audit (Tables S1–S6). Database.csv: The 10,950-entry single-atom alloy (SAA) source database containing C–H activation barriers, originally sourced from the Sun et al. and Wang et al. ML screening workflows. Code Availability: All Python code required to reproduce the in-fold leakage-safe preprocessing, matminer/CALPHAD feature enrichment, model training, cross-validation protocols, and active learning analyses is publicly available on GitHub: https://github.com/shamimalrazi393/SAA-ML-model/tree/main Usage Note: The .joblib model files are binary serialized Python objects and cannot be opened with standard text editors. Please refer to the load_models_example.py script provided in the repository to load and evaluate the models using Python (requires scikit-learn==1.6.1 and deep-forest).