Semi-supervised Machine Learning Enables the Robust Detection of Multireference Character at Low Cost

Area of Science:

• Computational chemistry
• Quantum chemistry
• Materials science

Background:

• Single-reference (SR) methods like density functional theory (DFT) struggle with strongly correlated electronic structures.
• Multireference (MR) diagnostics, often using wave function theory (WFT), are computationally expensive and can be inconsistent.
• Accurate prediction of molecular properties requires reliable identification of strong correlation.

Purpose of the Study:

• To develop a cost-effective and accurate method for identifying strong electron correlation.
• To overcome the computational cost and inconsistency issues of traditional multireference diagnostics.
• To enable reliable high-throughput screening of molecules with complex electronic structures.

Main Methods:

• A semi-supervised machine learning (ML) model using virtual adversarial training (VAT) was developed.
• The model was trained using 15 multireference (MR) diagnostics from WFT and DFT calculations.
• The approach was adapted to use regression predictions instead of WFT diagnostics to reduce computational cost.

Main Results:

• The VAT-based MR classifier demonstrated superior performance compared to alternative methods.
• The model accurately distinguished between SR and MR electronic structures, reflected in distinct molecular property distributions.
• Replacing WFT diagnostics with regression predictions maintained high performance, significantly reducing computational expense.

Conclusions:

• The developed MR decision engine offers a low-cost, high-accuracy solution for automatic strong correlation detection.
• The approach shows promise for predictive high-throughput screening in computational chemistry.
• The method is transferable to larger molecules and diverse chemical compositions.