Interpretable Small-Sample Deep Learning with Approximate Attribution Enhancement for Atomically Precise Gold...

Small experimental datasets make synthesis-condition modelling particularly sensitive to overfitting and data leakage. Here, a graph convolutional neural network (GCNN) was coupled with approximate attribution enhancement (AAE) and evaluated on 54 gold-nanocluster synthesis records using a strict fold-local pipeline. All preprocessing, supervised representation learning, augmentation, model selection, and calibration were confined to each training fold, and metrics were calculated only on untouched original records. The strongest no-augmentation model achieved a Matthews correlation coefficient (MCC) of 0.729. With 20,000 fold-local AAE samples, Logistic Regression reached an accuracy of 0.864, an F1 score of 0.862, and an MCC of 0.758. The corresponding MCC values decreased to 0.579, 0.526, and 0.652 under leave-one-ligand-out, leave-one-publication-out, and leave-region-out evaluation, respectively, indicating that extrapolation beyond represented chemistry remained more difficult than random-fold interpolation. Feature, label-weight, augmentation-baseline, covariance, calibration, and stability analyses consistently identified ligand aromaticity, temperature, and HAuCl₄ concentration as the most influential factors. An interpretable analysis of a 34-record aqueous Au₂₅ subset further produced condition rules and a pH-temperature map, while nine literature formulations provided an external audit within the represented domain. The results position AAE as a local regularisation strategy for data-limited synthesis modelling rather than as a source of new chemical information.