Benchmarking Foundation Potentials against Quantum Chemistry Methods for Predicting Molecular Redox Potentials.

Area of Science:

• Computational chemistry
• Materials science
• Sustainable chemistry

Background:

• High-throughput virtual screening is crucial for identifying molecules for sustainable applications like electrochemical carbon capture.
• Accurate quantum chemistry calculations, essential for this process, are computationally expensive.
• Machine learning foundation potentials (FPs) offer a computationally efficient alternative to density functional theory (DFT).

Purpose of the Study:

• To benchmark MACE-OMol-0 and UMA FPs against DFT functionals for predicting molecular redox potentials in electron transfer (ET) and proton-coupled electron transfer (PCET) reactions.
• To identify limitations of FPs in predicting redox potentials, particularly for ET reactions.
• To propose an optimized hybrid workflow for accelerating virtual screening in sustainable chemistry.

Main Methods:

• Benchmarking MACE-OMol-0 and UMA FPs against DFT functionals.
• Evaluating FP accuracy for ET and PCET reactions using experimental molecular redox potentials.
• Developing and testing a hybrid workflow involving FPs for geometry optimization and DFT for energy refinement.

Main Results:

• FPs demonstrated exceptional accuracy for PCET processes, comparable to the target DFT method.
• FP performance decreased for ET reactions, especially multielectron transfers involving underrepresented reactive ions.
• The proposed hybrid workflow showed robust and scalable performance for accelerating high-throughput virtual screening.

Conclusions:

• While FPs show promise, their accuracy for ET reactions is limited by training data representation.
• A hybrid approach combining FPs and DFT provides a pragmatic and efficient strategy for sustainable chemistry applications.
• This workflow enhances the scalability and robustness of virtual screening for identifying redox-active molecules.