Some lithium-ion batteries now carry a fire retardant sealed inside tiny polymer capsules. The capsules stay closed during normal use and burst when the cell overheats. The idea comes from pharmaceutics, but the battery work has so far taken only the container and not the mathematics drug designers use to decide whether a dose arrives in time and at the right strength. This proposal treats a capsule-protected cell the way a pharmacologist treats a patient on a narrow-margin drug. Two things must hold together. Over hundreds of cycles, the retardant that leaks out must stay below the level that damages the cell, which plays the role of the toxic ceiling. And after an abuse event, the released retardant must reach an effective concentration at the hot spot before the cell crosses its runaway trigger temperature, which plays the role of the minimum effective dose.
The central claim is that whether a capsule design suppresses runaway is governed by a race between two timescales. One is how long the payload takes to escape the shell and diffuse through the separator and electrode pores to the hot spot. The other is how long the cell's self-heating chemistry takes to carry it from the capsule's trigger temperature to the point of no return. The ratio of these two times should predict success or failure across capsule designs, states of charge, and cell ages far better than the rule the field uses now, which is simply to choose a shell that melts at the right temperature. If the claim holds, capsule design becomes a dosing problem that can be calculated, and it also predicts something uncomfortable: capsules tuned on fresh cells may quietly stop working as cells age.
What exists and what is missing
The closest work is Bioinspired Thermal Runaway Retardant Capsules for Improved Safety and Electrochemical Performance in Lithium‐Ion Batteries. It builds temperature-responsive core-shell capsules, reports nail-penetration peak temperature rises of about 5 °C against 50 °C for a reference cell, and reports capacity retention often above 95% after 50 cycles. This paper motivates the proposal and also supplies its starting hardware. It treats release as instantaneous and leaves open how the payload spreads through porous cell components. It also leaves open a standard way to weigh safety gain against electrochemical penalty, which it only sketches as variation coefficients. The present proposal adds the missing transport and kinetics model and turns the trade-off into an explicit window.
The cell-side kinetics come from calorimetry. Thermal Runaway Characteristics and Modeling of LiFePO4 Power Battery for Electric Vehicles measures the staged onset temperatures $T_1$, $T_2$ and $T_3$ in adiabatic calorimetry. It reports a self-heating activation energy of 110–113 kJ/mol and a maximum heating rate 535 times higher at 100% than at 25% state of charge, with model errors below 3%. That 535-fold spread is the key constraint: the time available for rescue must collapse at high charge, so any fixed trigger-temperature rule should fail somewhere.
The 3D internal-short simulations in Three-Dimensional Modeling for the Internal Shorting Caused Thermal Runaway Process in 20Ah Lithium-Ion Battery show two things. Short-circuit resistance and convection dominate the outcome, and only an impractical convection coefficient of 10,000 W/(m²·K) suppresses runaway externally. That result argues for intervening inside the cell.
On the aging side, Prognostic Modeling of Thermal Runaway Risk in Lithium-Ion Power Batteries Based on Multivariate Degradation Data models how $T_1$ erodes with cycling. It finds runaway risk exceeding 10% around 1000 cycles, but it borrows its $T_1$ relationship from offline calibration rather than measuring it. We measure it directly.
The pharmacokinetics literature supplies the lens rather than the data. Narrow-index drugs such as colchicine need delivery that keeps exposure between tight bounds (Delivery Strategies for Colchicine as a Critical Dose Drug). Extended-release products can "dose dump", meaning they release their payload early or unevenly under environmental stress such as a high-fat meal. Examples include the opposite-direction food effects in Pharmacokinetic comparison of sitagliptin and metformin HCl extended-release tablets versus JANUMET® XR and the altered peaks under fed conditions in Pharmacokinetics and bioequivalence study of esomeprazole magnesium enteric-coated tablets.
The methodological borrowing is from Machine learning predicts the effect of food on orally administered medicines. Across 311 drugs, a single dimensionless dose number, built from dose, solubility and fluid volume, carried most of the predictive power for food effects. We seek the analogous dimensionless number for retardant rescue. Our version is derived from physics rather than learned, which matters because battery abuse data are scarce.
Core Claim
Let $\tau_d$ be the delivery time: the time from shell trigger until the retardant concentration at the hot spot reaches its effective value $C_{\text{eff}}$. It includes shell opening time plus diffusion across a distance $L$ with effective diffusivity $D_{\text{eff}}$, which is the bulk diffusivity reduced by porosity and tortuosity. Let $\tau_r$ be the runaway time: the time for adiabatic self-heating to carry the cell from the trigger temperature $T_{\text{trig}}$ to $T_2$. It is computed by integrating the calorimetry-fitted Arrhenius heating rate. We define the rescue number
$$\mathcal{R} = \frac{\tau_{\text{open}} + L^2/D_{\text{eff}}}{\int_{T_{\text{trig}}}^{T_2} \frac{dT}{\dot T(T,\,\text{SOC},\,\text{age})}}$$
where $\dot T$ is the measured self-heating rate as a function of temperature, state of charge and aging state. The prediction is that suppression succeeds when $\mathcal{R}<1$ and fails sharply above it. The sharpness follows because the denominator shrinks exponentially as Arrhenius heating accelerates, while the numerator is set by geometry and is roughly fixed.
The toxicity side is a matching constraint. The concentration that has leaked into the electrolyte after $N$ cycles must stay below the free-additive level at which capacity fade begins. A design is admissible only if both constraints hold, which is the battery analogue of a therapeutic window.
Research Program
The program has three parts.
The first part calibrates the kinetics. We run adiabatic calorimetry on commercial LFP and NCM cells at 25, 50, 75 and 100% state of charge, following the protocol of the LFP calorimetry study. We also test cells pre-aged on the XJTU-style cycling protocol (2C charge, 1C discharge) to 0, 500 and 1000 cycles. This gives measured $\dot T(T)$ curves and measured $T_1$ and $T_2$ shifts with age, which directly fills the gap the prognostic paper conceded.
The second part characterizes the capsules. We fabricate them by in-situ polymerization following the bioinspired capsule work. The design matrix varies shell thickness, payload loading and shell melting temperature independently, so that trigger temperature is not confounded with release speed. We measure release kinetics with heated-stage microscopy and fit standard controlled-release forms, such as the Higuchi square-root-of-time law and the Korsmeyer–Peppas power law, to obtain $\tau_{\text{open}}$. Separator and electrode tortuosity give $D_{\text{eff}}$, and we measure passive leakage during cycling by analysing the electrolyte.
The third part builds the coupled model. A one-dimensional release-diffusion-consumption model is joined to the calorimetry-fitted heat balance. The consumption term covers retardant lost to radical scavenging and to venting, which corresponds to clearance in pharmacokinetics. We check this model against a 3D internal-short simulation in the style of the 20Ah modeling study.
Baselines
The naive reference is a cell with no retardant. It anchors peak temperature rise, which the capsule paper reports as 50 °C under nail penetration for its reference cell, and the fraction of abuse tests that reach runaway.
The strongest existing method is the original capsule design from the bioinspired capsule paper. It reports up to a 90% peak temperature reduction (5 °C versus 50 °C), self-extinguishing in about 10 seconds, and above 95% retention after 50 cycles. We reproduce it in our cells under our abuse protocol. Beating it means one of two outcomes. Either a rescue-number-optimized design at equal or lower capsule mass fraction suppresses runaway in conditions where the original design fails, especially at 100% state of charge and in aged cells. Or it matches suppression at lower mass fraction; the original external capsules use 5 wt%.
The free-additive baseline puts the same retardant directly in the electrolyte. It shows the toxicity side of the window, and the capsule paper argues it damages the interphase layer that forms on the electrode surface.
Among predictive baselines, the main competitor is the trigger-temperature rule: suppression is predicted whenever $T_{\text{trig}} < T_2$. Two ablations isolate the ingredients of the rescue number. One assumes instantaneous release, which is the implicit assumption in the capsule paper. The other keeps release kinetics but drops pore transport. All predictors are scored on the same abuse outcomes. The model is fitted only on calorimetry and release measurements, never on the abuse outcomes it must predict.
We also include a data-driven classifier trained on design parameters, a random forest following the food-effect study, given the same data budget. It is scored by leave-one-condition-out evaluation.
Why not just…?
Why not just choose a shell that melts below $T_1$? This is the real threat. If trigger temperature alone separates success from failure, the rescue number adds nothing. The design matrix settles it directly. At fixed melting temperature we vary shell thickness and loading. If outcomes do not change, the claim fails. The 535-fold heating-rate spread across state of charge in the LFP calorimetry study suggests the outcomes will change, because a capsule that opens in time at 25% charge may lose the race at 100%. That is a prediction, not yet evidence.
Why not just add retardant directly to the electrolyte? The capsule paper's motivation is that free retardants degrade the interphase layer and accelerate fade. Our free-additive arm measures this as a dose-response curve, and that curve is the ceiling of our window rather than a competitor to it.
Why not just cool the pack harder? The 20Ah internal-short simulation found that external convection suppresses runaway only near 10,000 W/(m²·K), which its authors consider unattainable on commercial cells. We still include the strongest practical cooling condition in the 3D simulations to show where internal delivery is needed.
Why not just learn the outcome from data? Abuse tests are destructive and few, much like the costly crossover trials behind the 311-drug food-effect dataset. Even with a well-chosen dimensionless descriptor, that study reached about 70% accuracy. A physics-derived number should extrapolate to unseen charge and age conditions. The leave-one-condition-out comparison against the random forest tests exactly this.
Proposed Experiments
The decisive test is out-of-sample prediction. The rescue number is computed from calorimetry and release data alone. It is then asked to predict nail-penetration and 180 °C oven outcomes across the full grid of capsule designs, states of charge and ages. We report the area under the curve for success versus failure for the rescue number, the trigger-temperature rule, both ablations and the random forest. We also report whether the empirical boundary falls near $\mathcal{R}=1$.
The second check is prospective design. Using the model, we choose the lightest capsule design predicted to satisfy both window constraints at 100% charge after 1000 cycles. We then test it against the reproduced original capsules on peak temperature rise, self-extinguishing time and retention after 50 and 500 cycles.
The third check concerns aging. We test whether capsules that succeed on fresh cells fail on aged cells as the measured $T_1$ shift predicts. A positive result would show that capsule dosing must be re-evaluated over battery life, much as drug dosing is adjusted as a patient's clearance changes.
There is no reported result on Lacuna for capsule suppression in aged cells or at multiple states of charge, so these comparisons will be new measurements rather than comparisons to published figures.
References
- Bioinspired Thermal Runaway Retardant Capsules for Improved Safety and Electrochemical Performance in Lithium‐Ion Batteries
- Thermal Runaway Characteristics and Modeling of LiFePO4 Power Battery for Electric Vehicles
- Three-Dimensional Modeling for the Internal Shorting Caused Thermal Runaway Process in 20Ah Lithium-Ion Battery
- Prognostic Modeling of Thermal Runaway Risk in Lithium-Ion Power Batteries Based on Multivariate Degradation Data
- Delivery Strategies for Colchicine as a Critical Dose Drug: Reducing Toxicity and Enhancing Efficacy
- Pharmacokinetic comparison of sitagliptin and metformin HCl extended-release tablets versus JANUMET® XR in healthy volunteers under fasting and fed conditions
- Pharmacokinetics and bioequivalence study of esomeprazole magnesium enteric-coated tablets 20 mg in healthy Chinese subjects under fasting and fed conditions
- Machine learning predicts the effect of food on orally administered medicines