This fragile layer must simultaneously block electrons while allowing sodium ions to pass freely, and its composition, structural heterogeneity, and mechanical integrity largely determine whether a sodium cell cycles efficiently or bleeds capacity with every charge. The problem has been that nobody could actually watch the SEI form. Conventional experimental techniques such as X-ray photoelectron spectroscopy and transmission electron microscopy typically provide ex situ or ensemble-averaged snapshots, struggling to capture the ultrafast, highly reactive interfacial processes occurring under operating conditions.

Atomic-scale simulations offer a complementary route, but each existing approach falls short in its own way. Ab initio molecular dynamics delivers first-principles accuracy yet is confined to a few hundred atoms over tens of picoseconds. Reactive force fields such as ReaxFF extend the timescale, but their predefined functional forms and reaction-specific parameterization limit transferability; ReaxFF simulations of carbonate electrolyte decomposition at alkali-metal electrodes have failed to capture essential single-electron reduction pathways, producing qualitatively incorrect product distributions.

Machine learning potentials trained on quantum calculations promised a way out, but at electrode-electrolyte interfaces they have remained limited to ultrashort durations of roughly 100 picoseconds, far too brief to resolve the nanosecond-scale densification and phase separation that govern SEI maturation. A research team has now reported a way through this methodological bottleneck. In a study centered on sodium metal electrodes in organic electrolytes, the researchers introduced an Interface Reactor sampling strategy to construct a charge-aware neuroevolution potential, abbreviated qNEP, designed specifically for metal electrode-organic electrolyte interfaces.

The strategy rests on three pillars. First, high-temperature ab initio molecular dynamics sampling broadens coverage of the descriptor space, since the descriptor distribution obtained at elevated temperatures nearly completely encompasses what is observed at lower temperatures, enhancing the ergodicity of local atomic environments. Second, short ab initio simulations are injected at critical nodes during active learning, just before a machine-learning-driven trajectory approaches a bifurcation point where an incomplete potential might steer the system along an unphysical energy pathway.

Third, the training set systematically incorporates representative inorganic products such as sodium fluoride and sodium oxide, along with organic decomposition intermediates in their distinct local environments near the substrate, within the interfacial layer, and in the bulk electrolyte. The resulting model was trained using the GPUMD package with a neuroevolution algorithm, taking approximately two days on four RTX 4090 graphics processing units and containing only 22,633 trainable parameters for six elements, a figure the authors highlight as evidence of the approach's data efficiency. The final training set comprises 5,035 configurations totaling 683,526 atoms.

Accuracy metrics are strong: the root mean square errors are 10.50 meV per atom for energy and 251.89 meV per angstrom for forces on the training set, with corresponding test-set values of 12.11 and 250.36 meV per angstrom. For comparison, a widely used ReaxFF force field yields a force error of 9,680 meV per angstrom on the same training set, roughly 38 times larger. In terms of speed, the trained potential runs about one order of magnitude faster than ReaxFF and roughly six orders of magnitude faster than ab initio molecular dynamics on a single RTX 4090 GPU, opening the door to million-atom simulations.

The payoff is a dramatic extension of accessible timescales. In a 100-nanosecond simulation of a sodium electrode-electrolyte system containing roughly 1,000 atoms, the potential remained stable throughout, an improvement of approximately two to three orders of magnitude in simulation duration compared with similar lithium and sodium electrode-electrolyte simulations reported previously. For a larger system of about 27,000 atoms, simulations on the scale of tens of nanoseconds were also achievable, with the longest single-GPU run taking approximately eight days.

Crucially, the model incorporates trainable atomic charges, allowing it to distinguish atoms in different chemical states and to track charge transfer during electrolyte decomposition and sodium-ion migration, a capability that standard machine learning potentials typically lack because they omit explicit charge and long-range electrostatic interactions. With this tool in hand, the researchers simulated five common sodium battery electrolytes, three carbonates (ethylene carbonate, dimethyl carbonate, and propylene carbonate) and two ethers (dimethoxyethane and diglyme), in contact with a sodium metal electrode. The two electrolyte classes decomposed along fundamentally distinct pathways.

Ethylene carbonate proved extremely reactive at both 350 and 450 kelvin, following two characteristic routes: a two-electron transfer that cleaves the ester carbon-oxygen double bond, generating carbonate species on the electrode surface while releasing ethylene gas into solution, and a single-electron process in which two carbon-oxygen single bonds within the carbonate ring break to form disodium ethylene glycolate and carbon monoxide. At 450 kelvin this decomposition neared completion within roughly 100 picoseconds. Dimethoxyethane, by contrast, showed far lower reactivity; at 350 kelvin only slow scission of a single carbon-oxygen bond generating methyl radicals was observed, with small amounts of ethylene and ethane appearing only at 450 kelvin.

In both solvent types, the PF6- anion decomposed relatively quickly and largely independently of the solvent, reacting with sodium to form sodium fluoride and phosphorus-containing compounds. Temperature primarily modulated reaction rates without altering the essential SEI formation pathways. Extended simulations of 100 nanoseconds in systems of roughly 27,000 atoms then revealed two distinct SEI growth mechanisms.

In the ethylene carbonate electrolyte, reactions were largely complete within the first 10 nanoseconds, with sodium carbonate, organic sodium salts, and gaseous products nucleating and growing simultaneously at different sites across the electrode interface, producing a mixed organic-inorganic co-growth structure in which fluorine and hydrogen distributions were scattered nearly randomly throughout the film. In the dimethoxyethane electrolyte, most of the metallic sodium and electrolyte showed no significant reaction even after 100 nanoseconds, with sodium fluoride as the primary product. Fluoride ions released from salt decomposition first dissolved randomly into the sodium matrix, then aggregated and underwent heterogeneous nucleation, growing on the sodium substrate in a spherical cap morphology.

As the amorphous phase transitioned into an ordered crystal, the sodium fluoride grew outward from the underlying metal in a morphology consistent with the calculated Wulff construction, a surface-energy-controlled process aligned with classical nucleation theory. This growth mode promotes preferential lateral extension, forming an ordered, dense, sheet-like overlayer. Sodium ions can thus hardly traverse the film, which self-limits its growth and kinetically suppresses further solvent decomposition.

Experimental measurements corroborated these atomic-scale predictions. Gas chromatography of sodium anodes immersed in the two electrolytes found ethylene at 58.6 percent of the gas products in the ethylene carbonate system, validating the two-electron carbonate pathway, while the dimethoxyethane gas was dominated by ethane at 42.0 percent and methane at 18.7 percent, reflecting carbon-oxygen cleavage and recombination. Quantitative X-ray photoelectron spectroscopy confirmed the SEI origins: the ethylene carbonate system showed higher sodium carbonate content (6.9 percent versus 1.36 percent), consistent with strong carbonate reactivity, while the ether system showed higher organic sodium content (5.08 percent versus 3.92 percent) and detectable sodium fluoride (1.44 percent), consistent with solvent alkylation and salt decomposition.

One discrepancy did emerge: theory based on the PBE functional predicts carbon monoxide to be more favorable than carbon dioxide, but experiments showed higher carbon dioxide, particularly at elevated temperature. The authors attribute this to the inherent limitations of the PBE functional and note that carbon dioxide's higher standard entropy, 214.3 versus 197.8 joules per mole-kelvin, makes its formation increasingly favored at higher temperatures as the Gibbs free energy change becomes more negative. The simulations also connected the bulk electrolyte's solvation structure to the SEI that ultimately forms.

Sodium ions exist in three forms in solution: solvent-separated ion pairs, contact ion pairs, and aggregates. Raising the NaPF6 concentration from 0.5 to 2 molar decreased the intensity of the main sodium-oxygen radial distribution peak and reduced coordination numbers, shifting the balance from solvent-separated ion pairs toward contact ion pairs and aggregates, while carbonate electrolytes exhibited sharper, more ordered solvation shells than the more flexible ether solvents. Because aggregates possess significantly poorer mobility than the other two species, they are left behind in the immediate interfacial region once the SEI forms, and once aggregates contact the interface they can form surface protrusions that act as sites for non-uniform SEI growth.

The electrolyte composition thus dictates the identity and reduction sequence of precursor species arriving at the interface, hierarchically governing SEI composition and growth patterns. Finally, metadynamics simulations probed what happens during charging, when sodium ions cross the interface and deposit into the anode, at a fixed deposition rate of 0.1 nanoseconds per atom. In a sodium carbonate-dominated SEI, the electrolyte was nearly depleted, indicating severe side reactions, and the sodium stored in the substrate appeared in an ionic, light-pink state; even a minor amount of sodium insertion induced metallic-like character in localized regions of the carbonate matrix, triggering rapid electrolyte decomposition.

In the sodium fluoride-dominated system, side reactions were effectively suppressed, the deposited sodium appeared metallic and white, and the amount of stored sodium was significantly greater. Quantitatively, the ether system showed a two-stage deposition behavior: deposited sodium counts remained constant for the first 5 nanoseconds while the fluoride-rich layer acted as a reservoir undergoing a solid-solution process, with metallic deposition beginning only after saturation, whereas the carbonate system rose slowly and steadily from the start. The simulations also illuminate the atomic-scale nature of irreversible sodium loss: some sodium ions fail to deposit as metallic sodium after crossing the SEI and are instead trapped within the layer, where their ionic state enhances local reactivity and drives side reactions with infiltrated electrolyte, converting stored sodium into surface compounds that constitute dead sodium.

The authors note that extending the framework to constant-potential simulations, which would allow direct electric-field effects to be addressed, remains future work, and that for carbon-based electrodes pore size drives a transition from ionic adsorption in small pores to quasi-metallic deposition in larger pores, providing an additional physical basis for sodium loss when strongly bound ionic sodium cannot rejoin the electrochemical cycle. Together, the findings establish an atomistic framework linking solvation structure, interfacial reactivity, and electrode storage behavior, and suggest that tuning SEI composition toward sodium fluoride-rich, self-passivating films is a promising route to improving sodium battery efficiency and cycling stability.

Subject of Research: Atomic-scale simulation of solid electrolyte interphase formation and sodium loss at sodium metal battery electrodes

Article Title: Unveiling Mechanisms of SEI Formation and Sodium Loss in Sodium Batteries via Interface Reactor Sampling

Article References: Liu, Z., Wang, Z., Chen, Z., Sha, J., Pan, F., Zhang, P., & Zhang, Y. (2026). Unveiling Mechanisms of SEI Formation and Sodium Loss in Sodium Batteries via Interface Reactor Sampling. Advanced Science, 13(59), Article e76808. https://doi.org/10.1002/advs.76808

Image Credits: AI Generated

DOI: 10.1002/advs.76808

Keywords: sodium batteries, solid electrolyte interphase, machine learning potentials, molecular dynamics, electrolyte decomposition, sodium fluoride, sodium carbonate, solvation structure, sodium metal anodes, metadynamics, energy storage, Unveiling