High-quality quantum chemical data for spin state determination in transition-metal complexes.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Materials Science

Background:

• Machine learning (ML) models excel in organic chemistry using reliable density functional theory (DFT) data.
• ML models are less reliable for transition metal complexes, especially spin-state energetics (SSE), due to DFT's system-dependent inaccuracies.

Purpose of the Study:

• To address the data limitations in ML for transition metal SSE.
• To create a benchmark dataset and develop improved ML models for accurate SSE predictions.

Main Methods:

• Computed spin energy gaps for 50 first-row mononuclear octahedral complexes using high-level CASPT2/CC multireference methods.
• Systematically benchmarked various DFT methods against the high-accuracy dataset.
• Introduced an electronic-structure-based descriptor (Des-δ) and employed a Δ-machine-learning (Δ-ML) framework.

Main Results:

• Demonstrated that the optimal fraction of Hartree-Fock exchange in DFT is dependent on the specific spin-state transition.
• Successfully extrapolated CASPT2/CC-level accuracy to a larger set of 500 complexes using the Δ-ML framework with the new descriptor.

Conclusions:

• The developed benchmark dataset and Δ-ML approach significantly enhance the reliability of ML predictions for transition metal SSE.
• This work provides a pathway to overcome DFT limitations in predicting SSE for transition metal complexes.