How Early Can Air Poisoning of a Lithium-Air Cathode Be Detected?

Lithium-air batteries take their cathode reactant, oxygen, straight from the air. That is the source of their very high theoretical energy, and also of their main practical weakness. Trace carbon dioxide and water in air react with the discharge product, lithium peroxide, and form lithium carbonate and lithium hydroxide. These crusts barely conduct electricity, so the charging voltage climbs and the cell eventually dies abruptly. Today this poisoning is usually seen only after it has done its damage: the overpotential spikes, or a mass spectrometer that samples the cell's gas registers a large carbonate signal.

This note proposes treating the onset of air poisoning the way particle physicists treat a rare-decay search. The question becomes a counting problem: at what carbon dioxide concentration, and after how many cycles, can carbonate formation be claimed with a calibrated false-alarm rate? Answering it requires three things that battery diagnostics currently lack. The first is a physically motivated template for the signal. The second is a background model constrained by a control channel. The third is an honest correction for the many places in time and voltage where a fluctuation could have been mistaken for a signal.

Core Claim

The proposal makes isotope-tagged carbon dioxide the ground truth and a Poisson likelihood-ratio test the detector. Cells are fed oxygen spiked with graded traces of carbon dioxide made from carbon-13. Any carbonate that forms from atmospheric gas then releases mass-45 carbon dioxide when it decomposes during charge. Carbon dioxide from electrolyte and carbon-cathode decomposition appears mostly at mass 44.

This separation gives the analysis a background region it can measure in advance. Naturally occurring carbon-13, about 1.1% of carbon, puts a predictable leak of electrolyte-derived gas into the mass-45 channel, and its size is set by the measured mass-44 rate. That is the same logic as a sideband control region in a physics search.

On top of this design we build a test statistic that combines ion counts with the charge-voltage profile. We calibrate its global significance by Monte Carlo trials on carbon-dioxide-free control cells. The deliverable is a sensitivity curve: the minimum detectable carbon dioxide fraction as a function of cycle number, at fixed global false-alarm rate and fixed detection power. No such curve exists in the literature in the research file. It would turn "trace carbon dioxide is harmful" into a quantitative detection horizon that can be checked against experiment.

Why This Intersection, and What Is Not New

Three cathode papers motivate the design. A combined density-functional and mass-spectrometry study, Combined DFT and Dems Study On the Role of CO2 Poisoning On the Electronic Conduction and Overpotentials in Li-Air Batteries, showed that carbon dioxide binds selectively at step and kink sites of lithium peroxide. These are also its growth sites. The resulting carbonate shuts off conduction far more strongly than the peroxide does. The implication is that early poisoning is a sparse, site-limited process whose gas signature per cycle is small, which is exactly the regime where counting noise dominates.

A Quasi‐Solid‐State Flexible Fiber‐Shaped Li–CO2 Battery with Low Overpotential and High Energy Efficiency reports that lithium carbonate needs 4.2–4.6 V to decompose. That fixes where in the charge the tagged signal should appear, and it gives us the shape of the signal template.

Coupling Water‐Proof Li Anodes with LiOH‐Based Cathodes Enables Highly Rechargeable Lithium–Air Batteries Operating in Ambient Air ran cells in ambient air and traced their failure to electrolyte oxidation rather than cathode passivation. This acts as a warning: electrolyte-derived carbon dioxide is a real, competing background. The isotope tag exists precisely to separate the two.

Two statistics papers supply the method. INDICATIONS OF INTERMEDIATE-SCALE ANISOTROPY OF COSMIC RAYS WITH ENERGY GREATER THAN 57 EeV IN THE NORTHERN SKY MEASURED WITH THE SURFACE DETECTOR OF THE TELESCOPE ARRAY EXPERIMENT found a 5.1σ local excess that fell to 3.4σ after one million sky-scrambled trials. We reuse that trial-factor procedure, replacing sky windows with windows in cycle and voltage. Branching fractions for psi(2S) -> gamma eta' and gamma eta reported a near-threshold channel as a central value together with a 90% upper limit. We adopt the same convention for sub-threshold cycles rather than declaring them "clean."

The closest machine-learning work is about battery degradation in general, not about gas diagnostics. DITING: A Weak Degradation Listener for Battery Lifetime Early Prediction amplifies faint aging trends in early cycling data. Battery Fault: A Comprehensive Dataset and Benchmark for Battery Fault Diagnosis classifies rare faults injected into electric-vehicle telemetry. BatteryML: An Open-source Platform for Machine Learning on Battery Degradation benchmarks lifetime regression. All three work on closed lithium-ion cells, rely only on voltage, current and temperature, and report average-case accuracy. None of them states a false-alarm rate or a detection limit. The proposal differs on all three points: it uses an open chemistry, it adds a gas-counting channel with isotope ground truth, and its output is a calibrated sensitivity limit rather than a lifetime estimate.

The research file reports no work that connects first-principles poisoning mechanisms to formal sensitivity limits. The obvious recipe at this intersection would turn mass-spectrometer traces into images and train a deep network, as EXO-200 did for its detector signals. We do not make that the core method, for reasons given below.

Research Program

The first component is the measurement design. We build Li-O2 cells with a carbon cathode and a glyme or dimethyl sulfoxide electrolyte, cycled under oxygen containing 0, 10, 30, 100, 300 and 1000 ppm of carbon-13 carbon dioxide. Several replicate cells run at each dose, with continuous mass spectrometry on masses 32, 44 and 45. The zero-dose cells serve as null data. Separate replicate cells are stopped at fixed cycles and titrated for carbonate. Titration is destructive, but it provides an independent check on the onset that the gas analysis claims.

The second component is the test statistic. In each charge step we bin the mass-45 counts in time. The expected background in bin $t$ is $b_t = r\,m_t + c_t$. Here $m_t$ is the simultaneous mass-44 count, $r$ is the natural-isotope leak ratio fitted on null cells, and $c_t$ is the instrument baseline. The expected signal is $\mu s_t$. In this expression $s_t$ is a template that turns on when the cell voltage passes the carbonate decomposition window, and $\mu \ge 0$ is the unknown signal strength. The likelihood-ratio statistic is

$$q = 2\sum_t \left[ n_t \ln\!\frac{b_t + \hat\mu s_t}{b_t} - \hat\mu s_t \right],$$

where $n_t$ is the observed count and $\hat\mu$ is the best-fit strength. Uncertainty in $r$ and $c_t$ is profiled out as nuisance parameters, meaning they are fitted alongside $\mu$ rather than fixed.

We combine $q$ with two further channels in a joint likelihood. One is the deficit in oxygen evolved per electron transferred. The other is the end-of-charge overpotential, modeled as a Gaussian drift. Because the minimum detectable signal grows roughly as the square root of the background counts, which is plain Poisson fluctuation, the sensitivity curve should scale predictably with integration time and instrument baseline. We test that scaling explicitly.

The third component is calibration. We declare onset at the first cycle where the running statistic crosses a threshold. That threshold is chosen from thousands of synthetic null histories, generated by resampling cycles from zero-dose cells, so that the probability of a false onset anywhere in a 100-cycle life is $10^{-3}$. This is the battery version of the post-trial correction used by Telescope Array.

A secondary analysis asks whether detected carbonate onset comes before the overpotential rise. If it does, that supports the proposed chain from adsorption to blocked growth sites to resistance spike. We borrow the falsification idea from Formalizing and Falsifying Causal Pathways of Rare Events: an intermediate step that looks normal while the rare pathway is active counts against the pathway. We keep this analysis modest, because that paper itself warns that the threshold chosen to turn a continuous quantity into a yes/no event can make or break the conclusion.

Baselines

The naive reference point is current practice. We subtract a baseline from the mass-44 trace and declare carbonate wherever the trace exceeds 3σ locally, with no trial correction. It runs on exactly the same cells. Its expected failure is a high false-onset rate on zero-dose cells, which the null data measure directly. Beating it means a lower minimum detectable dose at equal global false-alarm rate.

The second baseline is voltage-only early detection: a weak-signal model in the style of DITING: A Weak Degradation Listener for Battery Lifetime Early Prediction. It is retrained to classify dose from the first cycles of charge-voltage curves, using the same cells, the same splits and the same hyperparameter search budget as our method. It is the strongest available competitor among methods that need no mass spectrometer. That paper reports best or tied-best results on 10 of 16 dataset-metric pairs on the BatteryLife benchmark, but those results are for lithium-ion lifetime prediction. For trace carbonate detection there is no reported result found on Lacuna. To beat it, our method must detect onset at a lower dose or an earlier cycle at matched false-alarm rate.

The third baseline is a learned detector on the mass-spectrometer channels: a convolutional network on time-by-mass images, following the image approach of Search for Neutrinoless Double-beta Decay with the Complete EXO-200 Dataset. That work reports roughly a 25% sensitivity gain over its previous classifier. The network is trained on the same isotope labels, and its threshold is calibrated on the same null resamples.

The ablations remove one ingredient at a time. We drop the mass-44 sideband and fit background from mass 45 alone. We replace the voltage-gated template with a flat one. We remove the oxygen-deficit and overpotential channels. We skip trial calibration. Each ablation isolates how much of any gain comes from that ingredient.

Why not just…?

Why not just train a deep network on the spectrometer traces? It may well find structure we missed. That is why it is a baseline and not a strawman. But when the signal shape is roughly known and the background is Poisson, a likelihood ratio is already the most powerful test available. A network needs large labeled sets to calibrate false-alarm rates near $10^{-3}$, and a few dozen cells cannot supply them. If the network beats the likelihood test at matched calibrated false-alarm rate, our template is wrong. We would then report the network's gain as evidence of unmodeled signal shape, which is informative in itself.

Why not just titrate the cathode for carbonate? Titration is the gold standard for quantity, and we use it as the check. But it destroys the cell and gives one time point per cell. Pinning down onset would require dozens of sacrificed cells per dose, and it can never warn a running cell.

Why not just watch the charging voltage? The cited poisoning study describes the conduction cutoff as abrupt, which suggests voltage is a late indicator. This is a genuine open question, though, and the voltage-only baseline settles it directly. If it matches the gas-based method, the cheaper sensor wins and we will say so.

Why not just use a more sensitive spectrometer or integrate longer? Better hardware lowers the background, and our sensitivity curve shows exactly how much that buys under square-root scaling. It does not remove the need to correct for scanning many cycles. Without that correction, any instrument produces false onsets at the rate the null cells reveal.

Decisive Checks

The first check is calibration. On held-out zero-dose cells, the method's p-values must be uniformly distributed and its false-onset rate must stay near the target of $10^{-3}$. The uncorrected 3σ baseline is expected to violate this badly; if it does not, the look-elsewhere concern is overstated for this setting.

The second check is the headline result: the minimum carbon dioxide fraction detectable at 90% power within 10, 30 and 100 cycles, compared across all baselines. Detected onset must agree with titration onset within one sampling interval.

The third check is the scaling test. Halving integration time should worsen the detection limit by about $\sqrt{2}$; a clear departure would mean the background model is missing a systematic effect.

Finally, the temporal-precedence analysis reports whether carbonate onset leads the overpotential rise across doses. A consistent lead would give the first quantitative support for the poisoning chain as a measured sequence of events. Its absence would count against that chain.