Applying Large Graph Neural Networks to Predict Transition Metal Complex Energies Using the tmQM_rev Dataset

The discovery of novel, high-performing catalysts is essential to the economical development and deployment of many promising materials and fuels. Machine learning (ML) methods have proved useful in the acceleration of the catalyst discovery process, but often yield models restricted to specific chemical domains or types of structures. To obtain generalizable ML models, large and diverse datasets are needed, which tend to exist mostly for heterogeneous catalysis. The tmQM dataset, which contains 86,665 transition metal complexes and their properties calculated at the TPSSh/def2-SVP level of density functional theory, provided a promising dataset to train a generalizable ML model on homogeneous catalyst systems. We trained several ML models on tmQM, and found that these models consistently underpredicted the energies of a chemically distinct subset of the data. To address this, we present the tmQM_rev dataset, which filters out several structures in tmQM found to be missing hydrogens in their molecular geometries and recomputes the energies of all other structures in tmQM at the ωB97M-V/def2-SVPD level of density functional theory. ML models trained on tmQM_rev show no pattern of consistently incorrect predictions and much lower errors than those trained on tmQM. With respect to test set MAE, EwT, and parity, the ML models tested on tmQM_rev were, from best to worst, GemNet-T > PaiNN ~ SpinConv > SchNet. For all models, performance improved when using only neutral structures instead of the entire dataset, which also has charged species. However, models trained on only neutral structures appeared to saturate, while those trained on the entire dataset did not. Learning curves indicate that the models capture the chemical diversity of the neutral species in tmQM_rev. More data improves the model only when including charged species, indicating the importance of accurately capturing a range of oxidation states in future data generation and model development. Furthermore, a fine-tuning approach where model weights were initialized from models trained on OC20 led to drastic improvements in model performance. These results indicate transferability between ML strategies of heterogeneous and homogeneous systems.