Prediction and Experimental Verification of Electrolyte Solvation Structure from an OMol25-Trained Interatomic Potential

This study demonstrates that machine learning interatomic potentials trained on the chemically diverse OMol25 dataset outperform traditional inorganic-based models in accurately predicting the density, structure, and solvation dynamics of Na-ion battery electrolytes, a finding validated by experimental data and used to reveal how temperature and solvent topology influence ion-solvation behavior.

The Gist

The Big Picture: Building a Better Battery Without the Guesswork

Imagine you are trying to build a better battery for your phone or an electric car. The secret sauce inside a battery is the electrolyte—a liquid soup that allows ions (tiny charged particles) to swim back and forth, storing and releasing energy.

For decades, scientists have been stuck in a "Goldilocks" dilemma when trying to understand this soup:

• The "Super-Accurate" Method (DFT): This is like using a microscope to look at every single atom. It's incredibly precise, but it's so slow and expensive that you can only study a tiny drop of liquid for a split second. It's like trying to map a whole city by walking every single street on foot.
• The "Fast" Method (Classical Force Fields): This is like using a satellite map. It's fast and covers a huge area, but the details are blurry. It often misses the subtle interactions between atoms, leading to wrong predictions. It's like trying to navigate a city using a map drawn by someone who has never visited it.

The Solution: This paper introduces a new "AI Coach" (a Machine Learning Interatomic Potential, or MLIP) that acts like a super-smart GPS. It learns from millions of examples so it can predict how atoms behave with the accuracy of the microscope but the speed of the satellite map.

The Star of the Show: The "OMol25" Dataset

The researchers tested three different AI models. Two of them were trained on data about crystals and rocks (inorganic materials). The third one, called UMA-OMol, was trained on a massive new dataset called OMol25.

• The Analogy: Imagine you want to learn how to drive a car.
  • The first two models were trained by reading textbooks about trains and bicycles. They know a lot about wheels and metal, but they don't really understand how a car engine works or how to handle a slippery road.
  • The UMA-OMol model was trained on millions of hours of actual driving footage, specifically including rainy days, highway traffic, and tricky intersections (which represent the complex liquid chemistry of battery electrolytes).

What They Discovered

The team used this new AI coach to simulate Sodium-ion batteries (a cheaper, more abundant alternative to the Lithium-ion batteries in our phones) and compared the results to real-world experiments.

1. The Density Test (The "Weight" Check)

They asked the AI to predict how heavy the battery liquid would be.

• The Result: The models trained on rocks (trains/bicycles) got the weight wrong, often guessing the liquid was much lighter than it actually was. They also crashed the simulation (the "car" broke down).
• The Winner: The UMA-OMol model (trained on driving footage) got the weight almost perfectly right. It understood that the liquid molecules pack together tightly, just like in the real world.

2. The X-Ray Test (The "Structure" Check)

They used X-rays to see the arrangement of atoms in the liquid and compared it to the AI's prediction.

• The Result: The UMA-OMol model could "see" the tiny details of how the molecules wiggled and bounced off each other. The other models were too blurry to see these details.

3. The Temperature Effect (The "Heat" Factor)

They turned up the heat in the simulation.

• The Discovery: As the battery gets hotter, the "dance" between the sodium ions and the liquid changes. The ions loosen their grip on the liquid and start sticking to each other (forming "contact ion pairs").
• Why it matters: This is like a crowded dance floor. When it's cool, everyone keeps their personal space. When it gets hot and sweaty, people bump into each other more and form tight groups. The AI correctly predicted that this "bumping" increases with heat, which changes how well the battery conducts electricity.

4. The Solvent Shape (The "Key and Lock" Effect)

They tested different types of liquid solvents (some are short chains, some are long and flexible).

• The Discovery: The shape of the liquid molecule matters immensely.
  • Short, stiff molecules act like a tight hug, keeping the sodium ion isolated and free to move.
  • Long, flexible molecules act like a loose blanket. They wrap around the ion but leave gaps where other ions can sneak in and stick together.
• The Insight: By changing the "shape" of the liquid, you can control whether the ions stay free (good for speed) or clump together (bad for speed). The AI figured this out without being explicitly told the rules; it just learned from the data.

Why This Matters for You

This paper proves that we can now use AI to design better batteries much faster than before.

• No More Trial and Error: Instead of mixing chemicals in a lab for months to see what works, scientists can now run thousands of "virtual experiments" on a computer in a day.
• Cheaper Batteries: Sodium is everywhere (it's in table salt), unlike Lithium, which is rare and expensive. This research helps us figure out how to make Sodium batteries work as well as Lithium ones.
• The Future: The authors suggest that in the future, we could have "Digital Twins" of batteries—perfect virtual copies that we can test, break, and fix before we ever build the real thing.

In short: The researchers built a "Crystal Ball" for battery chemistry. It's trained on the right kind of data (liquid molecules, not just rocks), and it's telling us exactly how to mix the ingredients to build the next generation of powerful, affordable batteries.

Technical Summary

1. Problem Statement

Developing next-generation battery chemistries, particularly for Sodium-ion (Na-ion) batteries, requires a molecular-level understanding of electrolyte solvation structures and ion-ion correlations.

• The Accuracy-Scale Trade-off: Traditional ab initio molecular dynamics (AIMD) based on Density Functional Theory (DFT) offers high accuracy but is computationally prohibitive for the length and time scales required to study mesoscale transport and structure. Conversely, classical force fields are fast but often lack transferability, require labor-intensive parameterization, and fail to capture many-body effects and electronic polarization in chemically heterogeneous environments.
• The MLIP Gap: Machine Learning Interatomic Potentials (MLIPs) offer a promising middle ground. However, existing large-scale pre-trained MLIPs are predominantly trained on inorganic crystalline materials (e.g., Materials Project, OMat24). These models often fail to accurately predict properties of liquid electrolytes because they lack training data on liquid-phase configurations, solvation environments, and specific electrolyte chemistries (e.g., Na, organic solvents).
• The Need: There is a critical need for a universal, transferable MLIP trained on diverse molecular data that can accurately predict electrolyte density, structure, and dynamics without system-specific re-parameterization.

2. Methodology

The study employs a hybrid approach integrating computational modeling with rigorous experimental validation.

• Models Benchmarked:
  • UMA-OMol: A Universal Model of Atoms trained on the Open Molecules 2025 (OMol25) dataset. This dataset includes tens of millions of molecular DFT configurations with broad elemental coverage, specifically targeting electrolytes.
  • Orb-OMat & SevenNet-OMat: State-of-the-art models trained on inorganic materials datasets (OMat24, MPtrj, Alexandria).
• Simulation Protocol:
  • Systems: Simulations of Na-ion electrolytes (NaPF₆, NaOTf, NaTFSI) in various solvents (DME, DEGDME, TEGDME, PC, DMC, THF) and concentrations (0.1 M to 1.0 M). Li-ion and K-ion systems were also included for comparison.
  • Conditions: NPT ensemble simulations at temperatures ranging from 253 K to 353 K.
  • Initialization: Systems were equilibrated using classical OPLS-AA force fields before switching to MLIP-MD.
  • Analysis: Calculations of liquid density, X-ray structure factors ($S(Q)$), radial distribution functions ($g(r)$), coordination numbers, and ion-pair fractions (Free ions, Contact Ion Pairs [CIPs], Solvent-Separated Ion Pairs [SSIPs]).
• Experimental Validation:
  • Density: Measured using an Anton Paar DMA 4500 M oscillating U-tube instrument across various temperatures and concentrations.
  • X-ray Scattering: Total scattering measurements performed at the Advanced Photon Source (APS) to obtain experimental structure factors $S(Q)$ for 1 M sodium salt glyme electrolytes.

3. Key Contributions

1. Validation of OMol25-trained MLIPs: The study demonstrates that MLIPs trained on large-scale molecular datasets (OMol25) significantly outperform those trained on inorganic materials for liquid electrolyte simulations.
2. Systematic Benchmarking: It provides the first comprehensive benchmark of pre-trained MLIPs against experimental density and X-ray scattering data for Na-ion electrolytes across diverse chemistries.
3. Microscopic Insights: Using the validated UMA-OMol model, the authors map the complex interplay between cation identity, anion chemistry, salt concentration, solvent topology, and temperature on solvation structure.
4. Interface Characterization: The study extends the application of MLIPs to solid-liquid interfaces (Graphite/DME), revealing temperature-dependent structuring relevant to Solid Electrolyte Interphase (SEI) formation.

4. Key Results

A. Macroscopic Properties: Density and Structure Factors

• Density Prediction:
  • Inorganic-trained models (Orb-OMat, SevenNet-OMat): Showed substantial deviations from experimental densities ($R^2 \approx 0.34–0.45$), often underestimating values. Many simulations became unstable, leading to unphysical box breaking.
  • UMA-OMol: Achieved excellent agreement with experimental densities ($R^2 = 0.98$). While it exhibited a systematic ~8.5% overestimation (attributed to the $\omega$B97M-V functional's overbinding of dispersion and lack of explicit long-range electrostatics), it maintained high fidelity across all chemistries without re-parameterization.
• X-ray Structure Factors ($S(Q)$):
  • UMA-OMol accurately reproduced the dominant features and high-$Q$ oscillations of experimental $S(Q)$, capturing bond-stretching and torsional motions.
  • The mean absolute deviation ($\langle|\Delta S(Q)|\rangle$) for UMA-OMol was 10–30% lower than inorganic-trained models.
  • Inorganic-trained models frequently failed to converge or produced unstable trajectories for specific electrolyte combinations (e.g., NaOTf in DEGDME).

B. Microscopic Solvation Structure

• Anion Effects:
  • A clear hierarchy of ion-pairing strength was observed: OTf⁻ > PF₆⁻ > TFSI⁻.
  • OTf⁻ showed strong, sharp contact ion pairing ($g_{max} \approx 45$), while TFSI⁻ showed weak structuring.
  • Solvent Topology: Shorter glymes (DME) promoted stronger Na⁺–OTf⁻ association, whereas longer glymes (DEGDME) competed more effectively for Na⁺ coordination, suppressing direct anion binding.
• Temperature Effects:
  • Increasing temperature (253 K $\to$ 353 K) weakened cation-solvent binding and increased thermal fluctuations.
  • This led to a transition from fully solvated free ions to Contact Ion Pairs (CIPs). The fraction of free Na⁺ ions decreased significantly, while CIP formation increased (e.g., from negligible at 253 K to ~27% at 353 K).
• Concentration Effects:
  • Higher salt concentrations shifted the equilibrium from free ions to CIPs and larger aggregates. At high concentrations (3.14 M), the free-ion fraction dropped to ~10%, while CIPs dominated.
• Solvent Topology & Ion Aggregation:
  • DEGDME vs. TEGDME: Despite similar coordination numbers, their structural outcomes differed. DEGDME formed compact, octahedral chelates that sterically excluded PF₆⁻, favoring free ions. TEGDME adopted flexible conformations that created voids, allowing PF₆⁻ insertion and promoting CIP formation.
• Solid-Liquid Interface:
  • At the Graphite/DME interface, UMA-OMol predicted pronounced solvent and ion layering at 300 K, with DME oxygens coordinating Na⁺ near the surface.
  • At 353 K, thermal disorder reduced this layering, enhancing ion mobility and altering accessibility for electron transfer.

5. Significance and Impact

• Paradigm Shift: The study establishes that molecular-specific training data (OMol25) is essential for accurate electrolyte modeling, surpassing the limitations of inorganic-material-trained models.
• Predictive Power: UMA-OMol serves as a reliable, high-throughput surrogate for DFT, enabling the screening of complex electrolyte formulations (solvent mixtures, salt types, concentrations) that were previously inaccessible due to computational cost.
• Design Guidelines: The findings provide actionable insights for battery design:
  • Selecting solvents with specific topologies (e.g., chain length, flexibility) can tune ion-pairing and conductivity.
  • Temperature management is critical for controlling the free-ion vs. CIP equilibrium, directly impacting transport properties.
• Future Outlook: While UMA-OMol represents a major advance, the authors note areas for improvement, including handling long-range electrostatics explicitly, improving potential energy surface smoothness to prevent simulation crashes, and integrating experimental data directly into model training (Digital Twins) to further refine accuracy.

In conclusion, this work validates OMol25-trained MLIPs as a powerful, practical tool for accelerating the discovery of next-generation Na-ion battery electrolytes, bridging the gap between computational prediction and experimental reality.