Predicting Gamow–Teller β-Decay Matrix Elements: Reproducible ML Benchmark (R Markdown Run-Book, RData Workspace, and Tab-Delimited Data)

This repository accompanies the manuscript submitted to the Nuclear Physics A and provides a fully scripted, reproducible workflow for predicting Gamow–Teller (M(GT)) β-decay matrix elements from tabulated nuclear data. The package contains:

a cache-aware R Markdown run-book that reproduces all offline analyses, tables, and figures from a clean R session (instant rendering when the workspace file is present),

a complementary non-cached R Markdown that retrains all models under the same protocol,

a pre-trained RData workspace (models, preprocessed test matrices/targets, theoretical baselines), and

the tab-delimited input table used in the study.

The online M(GT) calculator referenced in the manuscript is a demonstration layer built on the same code base provided here; it is not used to compute the reported benchmark metrics.

Methods (concise): We evaluate multiple learners (random forests, Cubist, kernel SVM, linear-XGBoost, BRNN) under a stratified 10-fold cross-validation and held-out testing protocol. Hyperparameters are tuned strictly within training folds; final metrics are reported exclusively on unseen test partitions. Post-hoc explanations (e.g., SHAP) are included for interpretability. A fixed random seed and a clean parallel backend are used for deterministic runs.

Data provenance: The tab-delimited TXT table aggregates observations from the databases cited in the article (as detailed in the manuscript’s Data section). Column names match those used in the paper (e.g., A1, Z1, A2, Z2, J, N, Time, QEN, IBE, LOGft, EXPT) alongside theoretical baselines (USDB, IMSRG, CCEI, CEFT). Please refer to the article for the exact sources and curation details.

Files and Descriptions (File manifest)
reproduce_cached.Rmd — Primary, cache-aware run-book.
Loads pre-trained objects from the accompanying .RData when present and reproduces all offline results (metrics, tables, Taylor diagrams) without retraining; otherwise falls back to training.
reproduce.Rmd — Full retraining run-book.
Regenerates all models and figures from scratch under the same protocol (use when the .RData is not available or to re-train deliberately).
RIMAGE_reduced.rdata — Pre-trained workspace (cache).
Contains model lists (RF, Cubist, SVM-Radial, XGBoost-Linear, BRNN), corresponding preprocessed test matrices/targets, and theoretical baselines for both Full and Reduced feature regimes.
veri.txt — Tab-delimited input table (as used in the manuscript).
Aggregated from the databases cited in the paper (NUBASE2016, NNDC); includes experimental target values and the theoretical baselines (USDB, IMSRG, CCEI, CEFT). Column names match those referenced in Methods.
Related Article
Manuscript submitted to the Journal of Computational Science (JoCS) for evaluation. The repository mirrors the exact workflow used for the offline results reported in the paper. The online M(GT) calculator derives from the same scripts but is not used to produce benchmark metrics.