Redox potential prediction of Fe(ii)/Fe(iii) complexes: a density functional theory and graph neural network approach

Abstract

This work presents an integrated computational approach that combines tight-binding density functional theory (DFT) with standard DFT calculations to accurately compute the redox potential of micro-solvated iron-based transition metal complexes. Comparison with experimental and computational reference values confirmed the reliability of the computational approach. A comprehensive redox dataset of 2267 iron complexes was generated using the computational approach for machine learning (ML) applications. Chemical analysis of the ligand space in the dataset provided detailed insights into how different ligand classes and local Fe-coordination environments can systematically influence redox potential. Finally, a graph neural network (GNN) framework featuring automated graph data generation directly from 3D Cartesian coordinates was developed to model and predict redox potentials of metal complexes. Four distinct GNN architectures (GCN, GAT, DimeNet++, and SchNet) were evaluated using the curated redox potential dataset. The best-performing model achieved a root mean squared error of 0.24 ± 0.01 V, representing state-of-the-art performance for redox potential prediction of transition metal complexes. Feature attribution analysis using integrated gradients provided insights into the factors influencing the GNN model predictions. This combined DFT-ML workflow provides both predictive power and chemical insight, offering a scalable pathway to accelerate the discovery and optimization of transition metal complexes for any application.