Data-Driven Approaches Can Overcome the Cost-Accuracy Trade-Off in Multireference Diagnostics.

Area of Science:

• Computational chemistry
• Quantum chemistry
• Materials science

Background:

• Density functional theory (DFT) struggles with strongly correlated electronic structures.
• Multireference (MR) correlated wavefunction theory (WFT) is accurate but computationally expensive and difficult to automate.
• Existing diagnostics for MR character in single-reference (SR) calculations yield conflicting results.

Purpose of the Study:

• To evaluate the performance of 15 multireference diagnostics across a large dataset of organic molecules.
• To assess the ability of diagnostics to predict correlation energy recovery (%E_corr).
• To develop a machine learning approach for accurate and cost-effective prediction of MR diagnostics.

Main Methods:

• Computed 15 MR diagnostics (DFT-based and WFT-based) for 3165 small organic molecules.
• Evaluated diagnostic performance in predicting %E_corr.
• Developed kernel ridge regression models using DFT diagnostics and a 3D geometric representation to predict WFT-based diagnostics.

Main Results:

• Significant conflicts and low correlations were observed among existing diagnostics.
• DFT-based diagnostics were poor predictors of %E_corr compared to WFT-based diagnostics.
• Machine learning models accurately predicted WFT-based diagnostics, matching their computed accuracy and outperforming DFT-based methods.

Conclusions:

• Existing MR diagnostics show limitations in accuracy and consistency.
• Machine learning offers a viable solution to the cost-accuracy trade-off in computational screening.
• ML-predicted diagnostics enhance the reliability of high-throughput screening for complex molecular systems.