fig2

Figure 2. ML-accelerated prediction of catalytic activity for transition metal Pc DMSCs. (A) Workflow for predicting the limiting potential U_L: DFT- and database-derived features spanning atomic (r₁, r₂), electronic (e_d, N_d, H_f,ox, χ, I_m), and thermodynamic (Q, ΔG_COOH*, ΔG_Max) descriptors are used for model training, model selection, and prediction of target TM₁–TM₂ Pc DMPcs, where TM₁ is the central and TM₂ the side transition-metal atom; (B) Structure of the Pc DMSCs and the transition metals surveyed; (C) Parity plot for the best-performing GBR model: DFT-computed vs. ML-predicted limiting potential (-U_L), with test-set R² = 0.80 and RMSE = 0.12 V. (A-C) Reprinted (adapted) with permission from Ref.^[57]. Copyright © 2021 American Chemical Society. ML: Machine-learning; Pc DMSCs: phthalocyanine dual-metal-site catalysts; DFT: density functional theory; GBR: gradient boosting regression; RMSE: root mean squared error.