ReactionForge: Temporal Graph Networks with Cross-Attention and Evidential Learning Surpass State-of-the-Art in Suzuki-Miyaura Yield Prediction

Accurate prediction of chemical reaction yields remains essential for accelerating synthesis optimization, yet current machine learning models face critical limitations in capturing temporal dynamics, providing calibrated uncertainty estimates, and explicitly modeling reactant-to-product transformations. Here we introduce ReactionForge, a novel Temporal Graph Network architecture specifically designed for Suzuki-Miyaura cross-coupling yield prediction that addresses these challenges through five key innovations. First, we implement persistent temporal memory mechanisms using Gated Recurrent Units to track catalyst evolution and reagent dynamics across reaction sequences. Second, we develop cross-attention layers that explicitly compare reactant and product molecular graphs, learning which structural changes most influence reaction outcomes. Third, we incorporate hierarchical graph pooling via Self-Attention Graph Pooling to automatically discover functional group patterns. Fourth, we employ evidential deep learning to provide calibrated epistemic and aleatoric uncertainty in a single forward pass. Fifth, we use multi-task learning with yield, selectivity, and reaction time as joint objectives to improve generalization. Evaluated on 5,760 Suzuki-Miyaura reactions spanning five metal catalysts and diverse substrates, ReactionForge achieves R² = 0.968 ± 0.004 (RMSE = 5.12 ± 0.18%, MAE = 3.89 ± 0.12%), representing statistically significant improvements over YieldGNN (R² = 0.957 ± 0.005, paired t-test p = 0.002) and YieldBERT (R² = 0.810 ± 0.010, p < 0.001). The model provides well-calibrated uncertainty estimates (Expected Calibration Error = 0.031) that enable uncertainty-guided active learning, achieving 37% improved data efficiency over random sampling. Systematic ablation studies reveal that each architectural component contributes measurably to performance, with cross-attention and temporal memory each adding approximately ΔR² = 0.005. Interpretability analysis shows that learned attention weights successfully recover known structure-reactivity relationships, including chloride activation challenges and ligand-substrate compatibility patterns. Despite architectural complexity, ReactionForge trains 28% faster than YieldGNN. This work demonstrates that chemically motivated architectural innovations in graph neural networks can meaningfully advance reaction prediction when properly grounded in mechanistic understanding.