Applying Large Graph Neural Networks to Predict Transition Metal Complex Energies Using the tmQM_wB97MV Data Set.

Area of Science:

• Computational chemistry
• Materials science
• Catalysis

Background:

• Machine learning (ML) models show potential for catalyst discovery but often lack generalizability.
• Existing large datasets are primarily for heterogeneous catalysis, limiting applications in homogeneous catalysis.
• The tmQM dataset, while large, contains inaccuracies affecting ML model performance.

Purpose of the Study:

• To create a more accurate and generalizable dataset for training ML models in homogeneous catalysis.
• To evaluate the performance of different ML models on the improved dataset.
• To investigate the impact of data quality and chemical diversity on model accuracy.

Main Methods:

• Developed the tmQM_wB97MV dataset by correcting structural errors and recomputing energies using higher-level DFT (ωB97M-V/def2-SVPD).
• Trained and evaluated various ML models (GemNet-T, PaiNN, SpinConv, SchNet) on both tmQM and tmQM_wB97MV datasets.
• Assessed the impact of including neutral versus charged species and employed a fine-tuning strategy using pre-trained models.

Main Results:

• ML models trained on tmQM_wB97MV demonstrated significantly lower prediction errors compared to those trained on tmQM.
• GemNet-T outperformed other models, followed by PaiNN and SpinConv, then SchNet.
• Including charged species improved model performance, though models saturated with neutral structures alone.
• Fine-tuning models pre-trained on heterogeneous catalysis data (OC20) led to substantial performance gains.

Conclusions:

• The tmQM_wB97MV dataset significantly enhances the accuracy and generalizability of ML models for homogeneous catalysis.
• Data quality, chemical diversity (including oxidation states), and transfer learning are crucial for developing robust ML catalysts.
• The findings suggest a promising path towards more reliable in silico discovery of homogeneous catalysts.