Testing Whether Polysulfide Spectra Are Really Mixtures

Lithium–sulfur batteries lose capacity through the "shuttle effect." Sulfur on the cathode is reduced into intermediate molecules called polysulfides. These dissolve in the liquid electrolyte, drift across to the lithium anode, and corrode it. To study this, chemists record spectra across a working cell, one spectrum per spatial point (pixel) and per moment in time. They then "unmix" each spectrum as a weighted blend of a few pure species. The weights are read as how much of each species sits in that pixel. A good fit is taken as confirmation, but it is not. Population genetics learned this the hard way. Admixture tools such as ADMIXTURE describe each genome as a blend of ancestral populations, and they fit the data just as well when the true history was a bottleneck or contact with an unsampled "ghost" population. The resulting bar plots are confidently wrong (A tutorial on how not to over-interpret STRUCTURE and ADMIXTURE bar plots).

This proposal carries over the fix that paper introduced, not the mixture model itself. The fix is a consistency test: if a sample truly is a blend, the same blend weights must explain a second, independent kind of measurement. For batteries, the second measurement is a different spectroscopy of the same pixels. Examples are Raman, which reads sulfur–sulfur bond vibrations, and sulfur X-ray absorption, which reads sulfur oxidation state. A species the model never included will leave a residual that cannot be hidden in both views at once. Examples are the radical anion $\text{S}_3^{\bullet-}$, a solid precipitate, or a polysulfide in an unusual solvent environment. The target is lean-electrolyte cells, where such unlisted phases are already known to form and to cause failure.

Core Claim

In lean-electrolyte lithium–sulfur cells, the polysulfide-rich liquid droplets that drive anode failure should appear to standard unmixing as ordinary blends of reference polysulfides. A cross-modal residual test can flag them as a missing species, and can locate them in space and time. The same test should report that the droplets largely disappear when an additive that breaks them up is used.

The chemical motivation comes from Localization of micron-scale lithium polysulfide-rich phases causes anode heterogeneity in lean-electrolyte lithium–sulfur batteries. That paper showed that at an electrolyte-to-sulfur ratio of 2.3 µL/mg, polysulfides separate into micron-scale droplets. These droplets exclude the anions that normally build the anode's protective surface film, and adding lithium iodide fragments them. The evidence there was cryo-electron microscopy, nuclear magnetic resonance, Raman spectroscopy, and optical imaging, examined qualitatively or averaged over the whole cell. No per-pixel test showed where the droplet phase departs from a mixture of dilute species.

This is a "drifted source" problem in the population-genetics sense. The polysulfides in the droplets are chemically the same molecules as in dilute solution, but they sit in a different coordination environment, so their spectra are shifted. A linear unmixing model will absorb such a shifted signature as a strange blend of the listed references. That is exactly how admixture software absorbs a drifted isolate such as the Ari Blacksmiths (A tutorial on how not to over-interpret STRUCTURE and ADMIXTURE bar plots).

The Test

Let $\mathbf{X}$ be the pixel-by-channel matrix from one spectroscopy and $\mathbf{Y}$ the co-registered matrix from a second, physically different one. A true $K$-species mixture satisfies

$$\mathbf{X} \approx \mathbf{Q}\mathbf{F}, \qquad \mathbf{Y} \approx \mathbf{Q}\mathbf{G},$$

The symbols are defined as follows:

  • $\mathbf{Q}$ is the pixel-by-$K$ matrix of nonnegative fractions, with each row summing to one.
  • $\mathbf{F}$ holds the $K$ pure-species spectra in the first view.
  • $\mathbf{G}$ holds the same species' spectra in the second view.

The test runs in four steps:

  1. Estimate $\hat{\mathbf{Q}}$ and $\hat{\mathbf{F}}$ from $\mathbf{X}$ alone.
  2. On a held-out half of pixels, fit $\hat{\mathbf{G}}$ by least squares from $\hat{\mathbf{Q}}$ and $\mathbf{Y}$.
  3. Compute per-pixel residuals $\mathbf{R} = \mathbf{Y} - \hat{\mathbf{Q}}\hat{\mathbf{G}}$ on the other half.
  4. Compare those residuals with a null distribution. The null comes from a parametric bootstrap: resample noise around the fitted model, then refit.

Pixels whose residual energy exceeds the null, and which cluster in space, are flagged as containing a species outside the model. Step 2 plays the role that the second, independent similarity statistic plays in the tutorial's badMIXTURE diagnostic.

The test has a known blind spot, which we state up front. It detects a missing species only when the two spectroscopies respond to chemistry in different, non-linearly related ways. A ghost species whose signatures happen to be the same blend of the references in both views is invisible to it. We quantify this blind spot on synthetic data rather than assume it away.

A second transfer from population genetics addresses uneven sampling. Admixture tools split an over-sampled group into fake subgroups and merge rare groups, as described in the same tutorial. In cell maps, cathode pixels vastly outnumber separator and anode pixels, where the dissolved polysulfides actually are. We therefore reweight pixels by region during fitting and report whether the recovered species change.

Research Program

The work has three stages.

Synthetic benchmark. We build spectra with a known answer. They are generated from Raman and sulfur X-ray absorption references for $\text{S}_8$, $\text{Li}_2\text{S}_8$, $\text{Li}_2\text{S}_6$, $\text{Li}_2\text{S}_4$, and $\text{Li}_2\text{S}$, laid out along simulated diffusion profiles between cathode and anode. We then plant three kinds of ghost:

We vary ghost abundance, noise level, and the imbalance between cathode and separator pixels.

Intercalation sanity check. We rerun the analysis of the 2D X-ray absorption maps of $\text{Li}_{0.5}\text{CoO}_2$ electrodes from Non-negative matrix factorization for 2D-XAS images of lithium ion batteries, if that dataset can be obtained. This electrode does not dissolve, so it should not produce droplet-like ghosts. The false-alarm rate on real, non-shuttling data is the quantity of interest.

Real cells. We run co-registered operando Raman and sulfur X-ray absorption mapping of cells across the separator and anode region. The cells use the conditions of the localization paper: the baseline electrolyte (FN) versus the lithium iodide electrolyte (FNI), at 2.3 µL/mg. Optical microscopy of droplets serves as an independent spatial ground truth.

Baselines

Naive reference: residuals in one view only. The simplest detector fits simplex-constrained unmixing to one spectroscopy. It flags pixels whose reconstruction error exceeds a threshold calibrated on the same bootstrap. It shares our null model and data but ignores the second view. Beating it means higher detection power at a matched 5% false-alarm rate on the synthetic ghosts, especially the coordination-shifted one. We predict single-view residuals will miss that ghost because it can be absorbed into the blend.

Strongest existing method: orthogonal factorization with cross-validated $K$. This is the one-sided orthogonal nonnegative matrix factorization of Non-negative matrix factorization for 2D-XAS images of lithium ion batteries. It assigns pixels to non-overlapping domains and chooses the number of components by bi-cross-validation. It is the strongest competitor because it is the only battery-imaging unmixing method in the file. It also has a model-selection step that could, in principle, "discover" a ghost by adding a component. On synthetic data with eight true domains it recovered "eight distinct spatial domains and eight corresponding spectral profiles." For intercalation, it found ten domains on each real electrode.

We run it on the same concatenated two-view spectra, with the same preprocessing and the same budget of restarts. We count it as detecting a ghost if one selected component localizes to the planted region. Beating it means that when it absorbs the ghost into existing domains or into a spurious split, our test still flags the correct pixels. No reported result was found on Lacuna for ghost-species detection in lithium–sulfur spectral maps, so that comparison is new.

Chemometrics standard: multivariate curve resolution. Multivariate curve resolution with a sum-to-one constraint and model-order selection is standard in chemometrics. We run it as an additional competitor with identical constraints, although no paper for it appears in the file.

Ablations. Two ablations isolate the key ingredient:

  • The "second view" is replaced by a different energy window of the same spectroscopy. This tests whether physical independence matters or merely more channels.
  • The region reweighting is removed. This isolates the correction for uneven sampling.

Why not just…?

Why not just increase $K$ until the residual vanishes? This is the most serious threat. Adding components will eventually absorb a ghost. The trouble is that the same move also splits over-sampled cathode pixels into fake species. The genetics tutorial shows that fit quality alone does not separate these outcomes. The synthetic benchmark settles the question directly by scoring how often cross-validated $K$ recovers the planted ghost as a component versus producing a spurious split. If cross-validated $K$ matches our test's detection power, we will report that the cross-modal test adds nothing beyond model selection.

Why not just measure reference spectra for every candidate species and fit them directly? Reference fitting works only for species someone thought to measure. The droplet phase from the localization paper has no clean reference, because its signature depends on local coordination that cannot be prepared in a dilute standard. The synthetic arm tests this by holding one ghost out of the reference library. If supervised fitting with a complete library matches our results, that is expected. The comparison that matters is the incomplete library.

Why not just image the droplets optically or by cryo-electron microscopy, as the localization paper did? Imaging shows where the droplets are, but not their composition pixel by pixel during operation, nor whether the unmixing used elsewhere is mis-assigning them. We use optical imaging as ground truth for our flags, so this is an evaluation tool, not a competitor.

Why not use an amortized autoencoder like Neural ADMIXTURE for rapid genomic clustering? That model speeds up admixture fitting for very large genomic cohorts. Speed is not the failure here, and an autoencoder falls into the same over-simple-fit trap. Its simplex-constrained decoder can be slotted in as the unmixing step, but it does not replace the diagnostic.

Decisive Checks

On synthetic data, the cross-modal test must show two things:

  • higher detection power than both single-view residuals and cross-validated orthogonal factorization at matched false-alarm rates, for the coordination-shifted ghost;
  • a measured blind-spot region where it fails.

On the intercalation maps, it must raise few or no flags.

On real cells, flagged pixels must overlap optically imaged droplets in the FN electrolyte. The flagged area must shrink in the FNI electrolyte, consistent with that additive extending pouch-cell life from fewer than 50 cycles to 134 cycles in the localization paper. The flagged area should also grow near the anode as cycling proceeds.

If flags do not track droplets, or appear equally in both electrolytes, the claim that unmixing hides the droplet phase is wrong for these spectroscopies. We will report that outcome.