Δ -Machine Learning for the Prediction of Metal Complex Properties

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• High-throughput screening is crucial for discovering novel transition metal complexes.
• Traditional methods like density functional theory (DFT) are computationally expensive.
• Machine learning (ML) offers efficiency but faces accuracy and data challenges.

Purpose of the Study:

• To adapt the Δ-ML strategy for accurate quantum property prediction of transition metal complexes.
• To improve computational efficiency and data efficiency in materials discovery.
• To address the challenges of accuracy and data requirements in ML for chemistry.

Main Methods:

• Combined GFN2-xTB geometry optimizations and DFT single-point calculations for low-fidelity approximations.
• Generated featurized graph representations as input for a graph neural network (GNN).
• Utilized the tmQMg dataset for high-fidelity targets including electronic/dispersion energies, HOMO-LUMO gap, dipole moment, and polarizability.

Main Results:

• The proposed Δ-ML method achieved higher accuracy than conventional benchmarks for high-fidelity targets.
• Demonstrated improved data efficiency and out-of-domain transferability.
• Showcased significant computational cost reduction with minor performance loss using low-fidelity methods.

Conclusions:

• The Δ-ML strategy shows great potential for accelerating materials discovery in transition metal chemistry.
• This approach is effective even with limited training data, a common issue in the field.
• Highlights the viability of using cost-effective low-fidelity methods in ML-driven chemical research.