Benchmarking physics-inspired machine learning models for transition metal complexes with diverse charge and spin...

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Physics-inspired machine learning (ML) models are crucial for predicting quantum-chemical properties.
• These models can be structure-based or incorporate electronic information.
• Transition metal complexes present unique challenges due to diverse charge and spin states.

Purpose of the Study:

• To benchmark structure-only vs. structure-and-electronic ML models for transition metal complexes.
• To evaluate performance across diverse quantum-chemical properties and datasets.
• To provide guidance on selecting appropriate physics-based ML models.

Main Methods:

• Benchmarking kernel ridge regression with molecular representations (SLATM, FCHL, SOAP, SPA^HM).
• Evaluating geometric deep learning models (MACE, 3DMol).
• Assessing models with and without electronic information (charge and spin embeddings).

Main Results:

• Models with electronic information outperformed structure-only models for spin-splitting and frontier orbital energies.
• Structure-only models performed well for HOMO-LUMO gap and dipole moment magnitude.
• Geometric deep learning models (MACE-QS, 3DMol-QS) achieved highest accuracy, with 3DMol offering best efficiency.

Conclusions:

• Electronic information is essential for properties governed by electronic characters.
• Geometric information alone suffices for certain properties (e.g., HOMO-LUMO gap).
• Geometric deep learning models with charge/spin embeddings offer a powerful approach for transition metal complex property prediction.