Exchange Spin Coupling from Gaussian Process Regression.

Area of Science:

• Computational chemistry and materials science.
• Quantum mechanics and electronic structure theory.

Background:

• Heisenberg exchange spin coupling is vital for molecular catalysts, metalloenzymes, and magnets.
• Understanding spin coupling is key for applications in information technology and catalysis.
• Machine learning applications in chemistry often require extensive datasets, posing challenges for complex systems.

Purpose of the Study:

• To investigate the machine-learnability of Heisenberg exchange spin coupling using Gaussian process regression.
• To explore the effectiveness of various descriptors and kernels for predicting spin coupling.
• To compare machine learning performance against traditional methods like linear regression.

Main Methods:

• Employed Gaussian process regression (GPR) for its ability to handle small datasets and provide uncertainty estimates.
• Utilized a dataset of 257 small dicopper complexes.
• Compared a simple, chemically intuitive descriptor (Cu-X-Cu angles, Cu-Cu distances) against more sophisticated descriptors.

Main Results:

• A simple descriptor based on chemical intuition significantly outperformed sophisticated descriptors in extrapolating to larger complexes.
• Gaussian process regression demonstrated that exchange spin coupling is as machine-learnable as polarizability, but less so than dipole moments.
• Linear ridge regression performed comparably to kernel-based GPR on the small dicopper dataset due to descriptor-induced linearization.

Conclusions:

• The choice of descriptor is critical for accurate machine learning predictions of exchange spin coupling, especially for extrapolation.
• Chemical intuition in feature selection can be highly effective, rivaling automated methods for specific chemical problems.
• This study underscores the importance of balancing descriptor complexity with predictive power and extrapolation capability in machine learning for materials science.