Rapid Detection of Strong Correlation with Machine Learning for Transition-Metal Complex High-Throughput Screening

Area of Science:

• Computational chemistry
• Quantum chemistry
• Materials science

Background:

• Approximate density functional theory (DFT) is widely used in chemical discovery but is unreliable for systems with strong multireference (MR) character, such as open-shell 3d transition metals.
• Predictive discovery workflows require automated, cost-effective methods to identify chemical spaces where DFT is applicable and where it is not.

Purpose of the Study:

• To evaluate affordable, finite-temperature DFT fractional occupation number (FON)-based diagnostics for assessing MR character in open-shell transition-metal complexes.
• To develop and apply machine learning (ML) models for predicting MR character and guiding the use of DFT in large-scale chemical discovery.

Main Methods:

• Curated a dataset of over 4800 open-shell transition-metal complexes from high-throughput DFT studies.
• Evaluated finite-temperature DFT fractional occupation number (FON)-based multireference (MR) diagnostics.
• Trained machine learning (ML) models to predict HOMO-LUMO gaps and FON-based MR diagnostics.

Main Results:

• Intuitive measures like the HOMO-LUMO gap are not predictive of MR character, as determined by FON-based diagnostics.
• ML models revealed distinct metal and ligand sensitivities for HOMO-LUMO gaps and FON-based diagnostics.
• Rapid evaluation of MR character across ~187,000 theoretical complexes identified trends in spin-state-dependent MR character.

Conclusions:

• FON-based diagnostics and ML models offer a reliable approach to assess MR character in transition-metal complexes.
• These methods enable the identification of suitable regions in chemical space for DFT applications, improving predictive accuracy in chemical discovery.
• The study highlights the limitations of simple metrics and emphasizes the need for advanced computational tools for complex chemical systems.