Uncertainty-aware yield prediction for data-limited polystyrene autoxidation using heteroscedastic ensemble learning

Machine learning-based yield prediction in chemical reaction systems is often limited by small experimental datasets and condition-dependent variability. Here, we develop an ensemble regression framework that combines Natural Gradient Boosting (NGBoost) with a heteroscedastic deep learning with Multi-Layer Perceptron (MLP) to predict reaction yields from limited data while estimating condition-specific uncertainty. The framework was evaluated using 137 experimental samples from polystyrene autoxidation reactions. The ensemble model achieved a mean absolute error (MAE) of 5.4 ± 0.2, up to a 62% reduction relative to linear baseline methods (Linear Regression, Ridge, Lasso), and an approximately 24% reduction relative to the strongest single-model baseline evaluated (NGBoost alone, MAE = 7.1), outperforming the individual NGBoost and MLP models. Shapley Additive Explanations (SHAP)-based interpretation showed that the model captured non-linear response patterns associated with NaBr/Mn acetate loading, and reaction time, which were consistent with experimentally observed trends. These results suggest that uncertainty-aware ensemble learning can support reaction condition screening and prioritization in data-limited chemical systems.