Benchmarking foundation potentials against quantum chemistry methods for predicting molecular redox potentials

This study benchmarks machine learning foundation potentials against quantum chemistry methods for predicting molecular redox potentials, revealing their high accuracy for proton-coupled electron transfer but limitations for multi-electron transfers, and proposes a hybrid workflow combining efficient potential-based geometry optimization with single-point DFT energy refinement to enable scalable high-throughput screening.

The Gist

Imagine you are a chef trying to invent a new recipe for a dish that captures carbon dioxide from the air to help save the planet. To do this, you need to find the perfect "ingredient" (a molecule) that can grab the carbon and let it go again easily. The key to finding the right ingredient is knowing its "redox potential"—basically, how much energy it takes to make the molecule change its state to grab the carbon.

In the past, figuring out this energy level was like trying to bake a cake by weighing every single grain of flour and sugar with a microscopic scale. It was incredibly accurate, but it took so much time and computer power that you could only test a few recipes a year. This is what scientists call Quantum Chemistry (specifically a method called DFT).

The New Shortcut: "Foundation Potentials"

Recently, scientists developed a new type of AI tool called Foundation Potentials (FPs). Think of these FPs as a super-smart, trained assistant who has read millions of cookbooks (DFT calculations). Instead of weighing every grain of flour yourself, you ask the assistant, and they give you a very good guess instantly. Two specific assistants were tested in this paper: MACE-OMol and UMA.

The researchers wanted to know: Can we trust these AI assistants to find the perfect carbon-capturing ingredients without doing the slow, expensive work ourselves?

The Test Kitchen

To find out, the researchers set up a "taste test" using three different groups of molecules:

1. The "Simple Switch" Group (Electron Transfer): Molecules that just gain or lose an electron, like flipping a light switch.
2. The "Team Effort" Group (Proton-Coupled Electron Transfer): Molecules that gain an electron and a proton (a hydrogen ion) at the same time, like a team passing a ball and a bat together.
3. The "Non-Polar" Group: Molecules that don't like water, similar to oil.

What They Found

1. The "Team Effort" Group: The Assistants Were Perfect
When it came to the molecules that needed both an electron and a proton (PCET), the AI assistants were amazing. They predicted the energy levels almost exactly as well as the slow, microscopic-scale method.

• Analogy: It's like the assistant knowing exactly how much sugar to add to a cake just by looking at the picture, with zero error.

2. The "Simple Switch" Group: Good, but with a Catch
For molecules that just swap electrons (ET), the assistants were mostly good, but they stumbled when the molecule had to swap two electrons at once, especially if the resulting molecule was a reactive ion (a charged particle).

• The Problem: The AI had never seen enough examples of these specific "double-swap" charged molecules in its training data.
• The "Hallucination": When the AI tried to predict the shape of these tricky double-swapped molecules, it got confused. It essentially "hallucinated," predicting a shape that looked like a neutral molecule instead of the charged one it was supposed to be. It was like the assistant trying to bake a cake but accidentally making a loaf of bread because they had never seen a cake recipe with two eggs before.

3. The Speed Boost
Even when the AI wasn't perfect on the energy numbers, it was incredibly fast at figuring out the shape of the molecule and how it vibrates.

• Analogy: The AI could sketch the outline of the cake in seconds, while the old method took hours to measure every curve.

The Winning Strategy: The "Hybrid Workflow"

The researchers realized they didn't have to choose between "fast but sometimes wrong" and "slow but perfect." They proposed a hybrid workflow (a best-of-both-worlds approach):

1. Let the AI do the heavy lifting first: Use the fast AI assistants to quickly figure out the shape of the molecule and how it vibrates. This saves 99% of the time.
2. The "Final Check": Once the shape is set, run the slow, expensive, high-precision quantum calculation just one time on that specific shape to get the final, perfect energy number.
3. Add the "Water Factor": Since the AI was trained on dry molecules, they added a specific mathematical correction to account for how the molecule behaves in water (solvation).

The Bottom Line

This paper shows that these new AI tools are powerful enough to speed up the search for sustainable materials, but they aren't perfect on their own. They are like a brilliant apprentice who can do 90% of the work instantly but needs a master chef to do the final taste test for the trickiest recipes.

By combining the AI's speed with a final, precise check, scientists can now screen thousands of potential carbon-capturing molecules in the time it used to take to screen just a few. This makes the discovery of materials for a greener future much faster and more practical.

Technical Summary

Technical Summary: Benchmarking Foundation Potentials Against Quantum Chemistry for Molecular Redox Potentials

Problem Statement
The computational high-throughput virtual screening of redox-active molecules is critical for sustainable technologies, particularly electrochemical carbon capture. However, accurate prediction of molecular redox potentials relies on quantum chemistry methods like Density Functional Theory (DFT), which incur prohibitive computational costs for large chemical spaces. While machine learning foundation potentials (FPs) trained on extensive DFT datasets offer a computationally efficient alternative, their reliability for predicting derived electrochemical properties—specifically those involving subtle energy differences between distinct charge, spin, and protonation states—requires rigorous validation.

Methodology
The authors benchmarked two representative foundation potentials, MACE-OMol-0 (MACE-OMol) and UMA-s, against a hierarchy of conventional DFT functionals (B3LYP-D3(BJ), M06-2X-D3, ωB97X-D3(BJ), and ωB97M-D3(BJ)). The evaluation focused on predicting experimental redox potentials for two reaction types:

1. Electron Transfer (ET): Direct electron addition/removal.
2. Proton-Coupled Electron Transfer (PCET): Reactions involving simultaneous proton and electron transfer.

Three experimental datasets were utilized:

• Test Set A: Lewis base molecules in DMSO (1e⁻ and 2e⁻ ET).
• Test Set B: Quinones with polar functional groups in aqueous media (PCET at pH=0).
• Test Set C: Quinones with non-polar substituents (1e⁻ ET and PCET).

The computational workflow involved conformer sampling, geometry optimization, and vibrational frequency analysis. The authors proposed and tested a hybrid workflow: utilizing FPs for efficient geometry optimization and thermochemical analysis (Gibbs free energy corrections), followed by a single-point energy refinement using high-level DFT (target level: ωB97M-V/def2-TZVPD) and an implicit solvation correction (SMD model) to derive the final redox potential.

Key Results

• PCET Performance: Both MACE-OMol and UMA-s achieved exceptional accuracy for PCET reactions, rivaling their target DFT methods. For Test Set B and C (PCET), the Mean Absolute Errors (MAEs) for the FPs were approximately 0.04–0.10 V, comparable to the best-performing DFT functionals. The single-point DFT correction provided only marginal improvements for these systems, indicating the FPs already capture the necessary energetics for neutral or ion-functionalized molecules.
• ET Performance (1e⁻ vs. 2e⁻): Performance diverged significantly based on the number of electrons transferred.
  • 1e⁻ ET: UMA-s demonstrated strong intrinsic performance (MAE ~0.05 V), while MACE-OMol showed moderate errors (MAE ~0.16 V). However, applying the single-point DFT correction to MACE-OMol reduced its MAE to 0.029 V, matching the accuracy of high-level DFT.
  • 2e⁻ ET: Both FPs exhibited significant failures for multi-electron transfers involving reactive ions (e.g., BNSN²⁻), with MAEs exceeding 1.4 V. Analysis of Hessian matrices revealed that the FPs incorrectly mapped the 2e⁻ reduced species to geometries resembling the neutral ground state, a phenomenon the authors attribute to "hallucination" caused by the underrepresentation of dianions in the OMol25 training data.
• Computational Efficiency: The hybrid workflow leveraged the speed of FPs for Hessian calculations. For a 13-atom system, FPs computed Hessians in seconds, whereas DFT with analytical Hessians required hundreds of seconds, and numerical Hessians required over 3000 seconds.

Significance and Proposed Workflow
The paper establishes that while foundation potentials are highly effective for PCET and single-electron transfer processes (when corrected), they currently suffer from out-of-distribution limitations for multi-electron transfers involving reactive ions.

To address this, the authors propose an optimal hybrid workflow for sustainable chemistry screening:

1. Use FPs for rapid geometry optimization and vibrational frequency analysis.
2. Apply a crucial single-point DFT energy refinement to correct electronic energies, particularly for reactive ions.
3. Incorporate an implicit solvation correction (SMD) compatible with gas-phase optimized structures.

This pragmatic approach balances computational efficiency with accuracy, offering a scalable strategy to accelerate the discovery of materials for electrochemical carbon capture and energy storage, provided that the limitations regarding multi-electron transfer species are acknowledged and mitigated through DFT refinement.