Desolvation as a Decay Cascade: Branching Fractions of Transient Lithium Coordination States

Battery electrolyte design rests on a simple story. If the salt's negative ions (anions) sit in the lithium ion's first shell of neighbors, then the anions are reduced first at the electrode. That reduction builds a tough, inorganic protective film, and the cell lasts longer. The field measures this with equilibrium snapshots: the fraction of "contact ion pairs" and "aggregates" seen in bulk Raman spectra, an average coordination number, or a single lumped desolvation activation energy. These snapshots miss the moment that actually matters. That moment comes when the lithium ion sheds its neighbors one at a time at the electrode, and whichever ligand is still attached when the electron arrives decides the chemistry. We propose to describe that shedding the way particle physicists describe multi-body decays. A messy many-body transition is factored into a sequence of two-body steps through short-lived intermediate states. Each intermediate has a lifetime and a branching fraction (the probability that the cascade passes through it). These quantities are fitted from data, and the ambiguity of the fit is explicitly accounted for.

The concrete claim is testable. One measured quantity should predict lithium plating efficiency better than any equilibrium shell descriptor: the fraction of simulated desolvation cascades that still hold an anion when the lithium ion reaches the inner interfacial layer, which we call the anion-retention branching fraction. If that holds, it explains a known puzzle. In lithium–sulfur work, weaker solvation helps the interphase but slows kinetics Solvation-property relationship of lithium-sulphur battery electrolytes. Surface additives that never change the bulk shell can still halve the measured transfer barrier Enabling an Inorganic-Rich Interface via Cationic Surfactant for High-Performance Lithium Metal Batteries. Both results point to the pathway mattering more than the starting shell.

Core Claim

Two electrolytes can have the same equilibrium contact-ion-pair fraction yet differ in their decay cascades, and the cascades are what govern interphase quality. The cascades can differ in whether solvent or anion leaves first and in how long the anion-retaining intermediate survives near the electrode. We will build an "isobar-style" decomposition of desolvation trajectories into discrete intermediate coordination states. In particle physics, the isobar model factors a three-body decay into a two-body decay plus a resonance that later splits P. Rubin et al.. In our version, each step removes one ligand.

We will extract branching fractions and lifetimes from this decomposition and test whether they predict measured Coulombic efficiency (the fraction of plated lithium recovered per cycle) out of distribution. The test is whether they do better than bulk descriptors, and better than a generic state-clustering of the same trajectories that ignores the sequential structure.

What We Borrow From Particle Physics, and What We Drop

We do not borrow quantum interference. The isobar model sums complex amplitudes with relative phases because the decay paths end in indistinguishable final states and stay coherent. A solvated lithium ion in a room-temperature liquid loses phase coherence far faster than ligands exchange. Claiming interference between desolvation channels would be physically unjustified, and a chemistry reviewer would rightly reject it. Our cascade therefore sums path probabilities, not amplitudes.

Three things do transfer.

The first is factorization. A path probability is written as an initial-shell weight times a product of step-wise branching fractions:

$$ P(\text{path}) = w_{S_0}\prod_{k=1}^{K} b(S_{k-1}\to S_k \mid z_k), $$

Here $S_k$ is the coordination state after $k$ ligand losses (for example, three solvent molecules plus one anion). $w_{S_0}$ is the equilibrium weight of the starting shell. $b(\cdot\mid z_k)$ is the probability of that step given the ion's distance $z_k$ from the electrode.

The second is line-shape fitting with widths set by lifetimes. A resonance appears as a Breit–Wigner peak, a bell-like curve whose width equals the inverse of the resonance's lifetime. Correspondingly, an intermediate coordination state that decays exponentially contributes a Lorentzian component of width $\Gamma_k = 1/\tau_k$ to the frequency spectrum of the coordination-number time series. Here $\tau_k$ is the state's mean residence time. This Lorentzian form follows directly from Fourier-transforming an exponential decay. It gives us a second, independent route to the lifetimes that does not depend on choosing distance cutoffs.

The third, and in practice the most valuable, is how physicists handle fit ambiguity. Dalitz-plot analyses know that several sets of intermediate states can fit the same data equally well. To cope with this, they fit in two-dimensional joint variables rather than one-dimensional projections, report likelihood scans, and compare alternative state sets. They do not quote a single best fit. The isobar studies of charm decays rely on exactly this practice to separate overlapping broad states such as the $\kappa(800)$ P. Rubin et al., P. Rubin et al.. Electrolyte spectroscopy currently splits overlapping Raman bands into ion-pair and aggregate populations by single best-fit Gaussian deconvolution. We will import the multi-hypothesis practice wholesale.

Research Program

Desolvation is ligand exchange without bond breaking. That puts it inside the domain of a non-reactive machine-learned force field, which is where such models are reliable. We will run molecular dynamics with BAMBOO A predictive machine learning force-field framework for liquid electrolyte development, a graph equivariant transformer force field fitted to quantum data and aligned to experimental densities. It reports 0.01 g/cm³ density error, 17% viscosity deviation, and 26% conductivity deviation. Its acknowledged weakness on bond-breaking is not a problem for this step.

We will simulate each electrolyte twice. One run is bulk liquid. The other is a slab against a charged model wall, so that cascades are sampled as ions approach the interface. The wall fidelity is the largest modeling risk and is addressed in the checks below.

From these trajectories we will build two-dimensional "Dalitz-like" maps for every lithium ion over time. The two axes are the anion and solvent oxygen counts inside the first shell, and the maps are binned by distance from the wall. We will then fit competing state sets with explicit likelihoods. The minimal set has only the bulk solvent-separated pair, contact pair, and aggregate classes. Larger sets add transient desolvation intermediates. For each state set we will report branching fractions, lifetimes from both residence-time counting and the Lorentzian spectral fit, and likelihood-ratio intervals.

The key output is the anion-retention branching fraction $B_{\text{ret}}$: the probability that the ion is still anion-coordinated at the innermost layer.

The reactive consequence is scored separately and cheaply. For each terminal intermediate we will look up the reduction thermochemistry of its coordinated anion and solvent fragments in the LIBE dataset. LIBE holds 17,190 DFT-computed electrolyte fragment species with explicit monodentate-to-tetradentate lithium coordination Quantum chemical calculations of lithium-ion battery electrolyte and interphase species. Where entries are missing, we will use LIBE's bond-dissociation model, which reaches 0.022 eV MAE on its benchmark. This weights each cascade by whether its terminal state yields inorganic or organic products.

The analysis is deliberately different from LIBE itself. LIBE scores static minima; we score how often a minimum is actually reached.

Baselines

The naive reference point is bulk equilibrium shell composition: the contact-pair-plus-aggregate fraction computed from the same bulk BAMBOO trajectories, together with average lithium–anion coordination number. This is the descriptor the solvation-engineering literature uses implicitly. It explains the diluent study well Solvation structure modulation via dipole–dipole interactions for high-rate lithium metal batteries exceeding 400 Wh kg−1, where a polar diluent pulls solvent away and pushes anions into the shell. To count as a win, our branching fraction must give lower held-out error in predicting Coulombic efficiency than this descriptor.

The second baseline is a thermodynamic descriptor set: lithium solvation free energy and the LUMO gap between anion and solvent. These are the quantities behind the lithium–sulfur solvation correlation Solvation-property relationship of lithium-sulphur battery electrolytes.

The strongest existing learned method we can run is a graph model trained directly on formulation-to-outcome data in the Open Electrolyte Databank Open Electrolyte Databank: Unlocking Molecular-Mixture Intelligence for Battery Electrolyte Discovery through a Standardized, Multimodal Foundation. Its Coulombic-efficiency set is roughly an order of magnitude larger than earlier public collections. We found no reported Coulombic-efficiency prediction result on Lacuna for this databank, so we will train this baseline ourselves.

The fair comparison gives every model the same formulation splits, held out by solvent family in the spirit of the solvent-split protocol on MNSolv Relational Invariant Learning for Robust Solvation Free Energy Prediction. On MNSolv, the RILOOD method reports RMSE 0.6784 against 0.7662 for its CIGIN baseline. Those numbers are for solvation free energy, not efficiency, so they set no target here. Every model also gets the same hyperparameter-search budget. Our descriptors enter as extra input features to an otherwise identical model, which isolates their contribution.

The ablation that isolates the key ingredient is a Markov state model of the same trajectories. It clusters frames into states and fits transition rates, but it does not impose the one-ligand-per-step cascade or condition on distance to the electrode. If that clustering gives descriptors as predictive as $B_{\text{ret}}$, the sequential factorization adds nothing. A second ablation computes $B_{\text{ret}}$ from bulk trajectories only. It tests whether the interfacial conditioning is doing the work.

Why not just…?

Why not just use the bulk contact-pair and aggregate fraction? For many systems it may suffice, and the experiment is designed to find out. There is already evidence it is incomplete. The cationic surfactant study halved the lithium transfer activation energy, from 63.3 to 30.8 kJ/mol, and extended symmetric-cell life from 500 to 1,300 hours Enabling an Inorganic-Rich Interface via Cationic Surfactant for High-Performance Lithium Metal Batteries. It did so by acting at the electrode's charged layer rather than by changing the bulk shell. A bulk descriptor should be nearly blind to that additive, whereas $B_{\text{ret}}$ at the wall should register it. If the bulk fraction matches our predictive accuracy on held-out solvent families, we will report that the cascade picture is unnecessary.

Why not just run ab initio molecular dynamics at the interface? Direct quantum dynamics cannot reach the system sizes needed for nanometer-scale clusters. The diluent study reports clusters of about 2.53 nm, and statistical counts of rare desolvation paths are also out of reach, a limitation BAMBOO was built to address A predictive machine learning force-field framework for liquid electrolyte development. We will use a small number of quantum trajectories only to check that BAMBOO's interfacial step ordering is right.

Why not just cluster trajectories with a Markov state model? This is the real threat. It is the same family of tools, and our cascade can be viewed as a constrained Markov state model. We therefore run it as the ablation above. The honest prediction is that it will match us on in-distribution formulations. It may lose out of distribution, because unconstrained states fitted to one solvent family need not correspond to anything in another. Our states are defined by ligand counts, which transfer across families. If the Markov state model wins anyway, we will drop the factorization and keep only the interfacial-conditioning and fit-ambiguity contributions.

Why not just fine-tune a graph model on the efficiency data? A learned model can find correlates, but it cannot say which pathway failed. RILOOD's motivation was that such models latch onto spurious solvent features across solvent splits Relational Invariant Learning for Robust Solvation Free Energy Prediction. The Databank baseline tests whether our physically defined features improve the model's error on solvent families it never saw.

Proposed Experiments

The first check is internal. On a small set of documented systems, the extracted state set and branching fractions should be stable across likelihood scans and should agree between the residence-time and Lorentzian lifetime estimates. The systems are: three diluents of different polarity, TTE, toluene, and hexafluorobenzene (fluorinated, mildly polar, and nonpolar), at fixed salt; the weakly solvating dicarbonate DMDOHD with and without the additive FEC A dicarbonate solvent electrolyte for high performance 5 V-Class Lithium-based batteries; and an ester electrolyte with and without the surfactant CTAB. Where two state sets fit equally well, we report both, as Dalitz analyses do.

The second check is ordinal against published numbers. TTE is reported to give 99.0% lithium-to-copper Coulombic efficiency Solvation structure modulation via dipole–dipole interactions for high-rate lithium metal batteries exceeding 400 Wh kg−1. DMDOHD gives 92% without FEC and about 98% with it. CTAB gives 99.4% average efficiency. The predicted $B_{\text{ret}}$ ranking must reproduce the direction of each within-paper comparison: TTE above the less polar diluents, FEC above no FEC, CTAB above blank. CTAB is the case where bulk descriptors are expected to fail. The file reports no Coulombic-efficiency figures for toluene or hexafluorobenzene, so only the direction of that comparison is tested.

The decisive check is out-of-distribution regression on the Open Electrolyte Databank, held out by solvent family, with the baselines above. The cascade descriptor succeeds if adding it lowers held-out efficiency error beyond seed-to-seed variance relative to the bulk-composition, thermodynamic, and Markov-state-model features. It fails if the Markov state model or bulk composition matches it.

A final physical sanity check runs a handful of short quantum trajectories at the charged wall for two contrasting electrolytes. In those runs, the first-to-leave ligand should match BAMBOO's dominant branch. If it does not, the interfacial results are discarded and only the bulk results stand.

References

  1. Solvation-property relationship of lithium-sulphur battery electrolytes
  2. Enabling an Inorganic-Rich Interface via Cationic Surfactant for High-Performance Lithium Metal Batteries
  3. Search for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>C</mml:mi><mml:mi>P</mml:mi></mml:math>violation in the Dalitz-plot analysis of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mo>±</mml:mo></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>±</mml:mo></mml:msup></mml:math>
  4. First Observation and Dalitz Analysis of the<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msubsup><mml:mi>K</mml:mi><mml:mi>S</mml:mi><mml:mn>0</mml:mn></mml:msubsup><mml:mi>η</mml:mi><mml:msup><mml:mi>π</mml:mi><mml:mn>0</mml:mn></mml:msup></mml:math>Decay
  5. A predictive machine learning force-field framework for liquid electrolyte development
  6. Quantum chemical calculations of lithium-ion battery electrolyte and interphase species
  7. Solvation structure modulation via dipole–dipole interactions for high-rate lithium metal batteries exceeding 400 Wh kg−1
  8. Open Electrolyte Databank: Unlocking Molecular-Mixture Intelligence for Battery Electrolyte Discovery through a Standardized, Multimodal Foundation
  9. Relational Invariant Learning for Robust Solvation Free Energy Prediction
  10. A dicarbonate solvent electrolyte for high performance 5 V-Class Lithium-based batteries