I view the retirement of lithium-ion cells not as an endpoint but as the beginning of a second, highly constrained life. In my assessment, the global energy transition has made the retired lithium-ion cell a strategic asset for communication base stations, distributed microgrids, low-speed electric vehicles, backup power, and residential storage. The critical gatekeeper for this second life is State of Health (SoH), because a lithium-ion cell with unknown degradation cannot be safely or economically reassigned. From my first-person perspective, the technical problem is no longer whether retired lithium-ion cells can be reused, but how quickly, accurately, and robustly their SoH can be estimated under industrial constraints. I have therefore organized this review around direct measurement, model-based estimation, and data-driven estimation, and I add my own synthesis of hybrid architectures, uncertainty quantification, sensor innovation, and multi-physics coupling.
In my experience, SoH for a retired lithium-ion cell is not a single scalar truth. It is a family of condition indicators, each tied to a different failure mode, use case, and measurement pathway. Capacity fade, resistance rise, lithium inventory loss, active material loss, electrolyte depletion, and mechanical degradation can all move independently. Therefore, I treat SoH estimation as a multi-indicator inference problem. The following equation expresses the most common capacity-centric definition I use when screening retired lithium-ion cells:
$$ SoH_C = \frac{C_{aged}}{C_{rated}} \times 100\% $$
I also use a resistance-centric definition, because a retired lithium-ion cell may retain acceptable capacity while exhibiting unacceptable power capability:
$$ SoH_R = \frac{R_{aged} – R_{new}}{R_{new}} \times 100\% $$
For industrial sorting, I prefer a composite health vector rather than a single number. In my framework, the vector includes capacity, direct-current resistance, electrochemical impedance, thermal stability, and self-discharge. A weighted fusion can be written as:
$$ SoH_{composite} = w_C SoH_C + w_R SoH_R + w_E SoH_E + w_T SoH_T + w_S SoH_S $$
where the weights depend on the target application. For a low-speed electric vehicle, I assign more weight to capacity and energy efficiency. For a grid buffer, I assign more weight to resistance and thermal behavior. For a backup system, I assign more weight to self-discharge and calendar stability. This weighting logic is central to my review because a retired lithium-ion cell that passes one test may fail another.
Definitions, Indicators, and Measurement Boundaries
I define the measurement boundary before I compare algorithms. A retired lithium-ion cell arrives with unknown history, unknown partial cycling, unknown storage temperature, and often unknown cell-to-cell variation. The absence of a clean birth certificate makes every estimation method dependent on assumptions. I therefore summarize the key SoH indicators in Table 1.
| Indicator | Physical meaning | Typical measurement | Strength | Limitation |
|---|---|---|---|---|
| Capacity fade | Loss of usable charge | Full charge-discharge, coulomb counting | Direct and interpretable | Slow, accelerates aging |
| Resistance rise | Loss of power capability | HPPC, DC pulse, EIS | Fast, online friendly | Temperature and SoC sensitive |
| Lithium inventory loss | Loss of cyclable lithium | Half-cell, differential voltage, P2D fitting | Mechanistic insight | Requires advanced diagnostics |
| Active material loss | Loss of accessible sites | ICA, DVA, EIS, model fitting | Links to electrode degradation | Indirect and model dependent |
| Impedance growth | Interfacial and diffusion limitation | EIS, multi-frequency excitation | Rich frequency information | Slow full-spectrum scan |
| Self-discharge | Charge leakage and shuttle effects | Rest voltage tracking | Safety relevant | Long rest time required |
| Thermal stability | Heat generation and safety margin | Calorimetry, thermal imaging | Safety screening | Expensive and not online |
In my own screening logic, I treat the retired lithium-ion cell as a stochastic system. The true SoH evolves with cycling and calendar time, while the measured SoH is corrupted by sensor noise, temperature gradients, and incomplete data. I write the general observation model as:
$$ SoH_{measured,k} = SoH_{true,k} + b_k + \epsilon_k $$
where \(b_k\) is a slowly varying bias and \(\epsilon_k\) is random noise. I emphasize this because most published accuracy claims ignore bias. A method that reports low random error but high bias is dangerous for sorting. A method that reports slightly higher random error but bounded bias is often better for industrial deployment.
Direct Measurement Methods
I begin with direct methods because they are the foundation of any validation campaign. Direct methods do not infer SoH from a learned representation; they measure a physical quantity and then map that quantity to health. In my experience, direct methods are indispensable for labeling retired lithium-ion cells, even when they are too slow for full-scale sorting.
Capacity Testing and Coulomb Counting
Capacity testing is the most intuitive approach. I discharge a retired lithium-ion cell under a controlled protocol, integrate current, and compare the result with the rated capacity. The core equation I use is:
$$ C_{usable} = \int_{t_0}^{t_1} \eta_c(t) I(t) dt $$
If I include dynamic efficiency compensation, I rewrite the efficiency factor as a function of temperature, current, SoC, and age:
$$ \eta_c(t) = f(T, I, SoC, age) $$
In my view, enhanced coulomb counting can reach approximately \(\pm 1.5\%\) capacity estimation error, but only when the cell is brought to a known thermal state and the current sensor is well calibrated. The disadvantages are severe for retired lithium-ion cells: a complete test can require four to eight hours, and repeated deep cycling further degrades the very asset being evaluated. I therefore use capacity testing as a laboratory reference, not as a high-throughput industrial method.
Internal Resistance Testing
Internal resistance is negatively correlated with SoH, so I often use it as a fast proxy. The simplest calculation is:
$$ R_{dc} = \frac{\Delta V}{\Delta I} $$
For a retired lithium-ion cell, I measure the voltage response to a current pulse and extract ohmic, charge-transfer, and diffusion contributions. Hybrid pulse power characterization is widely used because it fits industrial time budgets. However, I have found that conventional second-scale pulses miss fast dynamics. Short pulses in the microsecond range can excite intrinsic responses and reduce concentration polarization interference. In my interpretation, the pulse duration must match the physical process being probed. I summarize this in Table 2.
| Pulse scale | Dominant process | Information gained | Risk |
|---|---|---|---|
| Microsecond | Ohmic and interfacial | Intrinsic resistance | Sensor bandwidth limits |
| Millisecond | Charge transfer | Kinetic limitation | Mixed polarization |
| Second | Diffusion and concentration | Mass transport | Long test and heating |
| Multi-frequency | Broad spectral response | Robust impedance features | Complex excitation design |
I also note that resistance rise in a retired lithium-ion cell is not uniform. The solid electrolyte interphase may thicken, contact resistance may increase, and pore clogging may reduce effective transport. Therefore, a single resistance value is insufficient. I use a resistance vector:
$$ \mathbf{R}_{health} = [R_o, R_{ct}, R_{SEI}, R_{diff}, R_{contact}]^T $$
The mapping from this vector to SoH can be linear, nonlinear, or probabilistic. In my experience, a monotonic mapping is often assumed but not guaranteed, especially when different aging modes compensate each other.
Electrochemical Impedance Spectroscopy
Electrochemical impedance spectroscopy is the most information-rich direct method I use. I apply a small sinusoidal perturbation over a wide frequency range and measure the complex impedance:
$$ Z(\omega) = \frac{V(\omega)}{I(\omega)} = Re(Z) + j Im(Z) $$
A typical equivalent circuit representation for a lithium-ion cell is:
$$ Z(\omega) = R_o + \frac{R_{ct}}{1 + j \omega R_{ct} C_{dl}} + \frac{1}{j \omega C_{SEI}} + Z_W(\omega) $$
In my analysis, high-frequency impedance is closely related to ohmic resistance, while mid-frequency arcs reveal charge transfer and interfacial effects. Low-frequency tails reflect diffusion. I have observed that ohmic resistance can correlate linearly with capacity fade in some retired lithium-ion cell populations, but the correlation is chemistry dependent. A practical linear model I may test is:
$$ SoH = \alpha + \beta R_o + \gamma R_{ct} + \delta C_{dl} + \epsilon $$
Traditional full-spectrum EIS is slow, which limits online use. I therefore favor operando EIS, sparse frequency sampling, and multi-frequency excitation. Sparse sampling reduces test time, while operando EIS captures dynamic impedance during charge or discharge. In my view, the future of EIS for retired lithium-ion cells lies in selecting a small set of frequency points that maximize aging information per second.
Model-Based Methods
I now turn to model-based methods. These methods build a mathematical or physical representation of the lithium-ion cell and then estimate SoH from model parameters or states. In my framework, the general pipeline is model construction, parameter identification, and state estimation. I write this as:
$$ \theta_{k+1} = \theta_k + r_k $$
$$ y_k = h(x_k, u_k, \theta_k) + v_k $$
where \(\theta\) contains health-relevant parameters, \(x\) contains electrochemical or electrical states, \(u\) is current, and \(y\) is voltage or temperature. The parameter drift is the hidden signature of aging. I classify model-based methods into equivalent circuit models, electrochemical models, and mathematical models.
Equivalent Circuit Models
Equivalent circuit models are my preferred starting point for online estimation because they are computationally efficient. The Rint model contains only an open-circuit voltage source and an ohmic resistance:
$$ V_t = OCV(SoC) – I R_o $$
The Thevenin model adds one RC branch:
$$ V_t = OCV(SoC) – I R_o – V_{RC} $$
$$ \frac{dV_{RC}}{dt} = -\frac{V_{RC}}{R_p C_p} + \frac{I}{C_p} $$
I can extend this to multiple RC branches for better transient fidelity. For a retired lithium-ion cell, I often observe that ohmic resistance and polarization resistance increase exponentially with aging, while polarization capacitance decreases. A simple empirical trend is:
$$ R_o(N) = a + b \exp(cN) $$
$$ R_p(N) = d + e \exp(fN) $$
$$ C_p(N) = g – hN $$
where \(N\) is equivalent full cycles. In my experience, these trends are useful for interpolation but risky for extrapolation. I summarize common equivalent circuit structures in Table 3.
| Model | Structure | Parameters | Strength | Weakness |
|---|---|---|---|---|
| Rint | OCV source plus \(R_o\) | \(R_o\) | Simple and fast | No polarization dynamics |
| Thevenin | OCV plus \(R_o\) and one RC | \(R_o, R_p, C_p\) | Good online compromise | Limited transient fit |
| Dual polarization | OCV plus two RC branches | \(R_o, R_1, C_1, R_2, C_2\) | Better dynamic fidelity | More identification burden |
| PNGV | OCV plus \(R_o\), RC, and capacitance | Extended parameter set | Captures cumulative charge | Model order issues |
| Nonlinear ECM | Voltage-current nonlinear mapping | State-dependent parameters | Wide operating range | Complex parameterization |
For parameter identification, I use recursive least squares, Kalman filtering, or particle filtering. The discrete Kalman filter form I use is:
$$ \hat{x}_{k|k} = \hat{x}_{k|k-1} + K_k (y_k – h(\hat{x}_{k|k-1})) $$
$$ K_k = P_{k|k-1} H_k^T (H_k P_{k|k-1} H_k^T + R_k)^{-1} $$
I also use square-root unscented Kalman filters when the nonlinearity is strong. In my assessment, equivalent circuit models are excellent for online SoH tracking in a known lithium-ion cell family. Their main weakness is generalization: a model tuned for one chemistry, format, or electrode design may fail on another retired lithium-ion cell.
Electrochemical Models
Electrochemical models are the most physically meaningful methods I use. The pseudo-two-dimensional model couples solid-phase diffusion, electrolyte transport, and charge conservation. The solid-phase diffusion equation is:
$$ \frac{\partial c_s}{\partial t} = \frac{D_s}{r^2} \frac{\partial}{\partial r} \left( r^2 \frac{\partial c_s}{\partial r} \right) $$
The electrolyte transport equation is:
$$ \frac{\partial c_e}{\partial t} = \nabla \cdot (D_e \nabla c_e) + \frac{a_s j}{F} $$
The Butler-Volmer equation links current density to overpotential:
$$ j = i_0 \left[ \exp \left( \frac{\alpha_a F \eta}{RT} \right) – \exp \left( – \frac{\alpha_c F \eta}{RT} \right) \right] $$
I find that the pseudo-two-dimensional model can explain capacity fade through lithium inventory loss and active material loss. In my own interpretation, solid electrolyte interphase growth is a dominant aging mode in many retired lithium-ion cells. The single-particle model simplifies the pseudo-two-dimensional model by assuming uniform electrolyte concentration and spherical particles:
$$ \frac{dc_s}{dt} = \frac{3}{R_p} j $$
The single-particle model reduces computation by orders of magnitude, but it sacrifices electrolyte dynamics. I use it when I need a fast electrochemical observer. The pseudo-two-dimensional model is better when I need mechanistic diagnosis. Table 4 compares these electrochemical models.
| Model | Fidelity | Computation | Parameterization | Best use |
|---|---|---|---|---|
| P2D | High | High | Difficult | Mechanistic degradation analysis |
| SPM | Medium | Low | Moderate | Online electrochemical observer |
| Reduced-order EM | Medium | Low to medium | Moderate | Embedded battery management |
| Multi-scale coupled | Very high | Very high | Very difficult | Research and digital twin |
I want to stress that electrochemical models require parameters that are rarely available for retired lithium-ion cells. Porosity, particle radius, diffusion coefficients, and reaction rates may be unknown. Therefore, I often combine electrochemical models with online parameter adaptation. The model becomes a structured prior rather than a fixed truth.
Mathematical Models
Mathematical models simplify aging into empirical or probabilistic relationships. I use them when I need fast deployment and interpretable trends. Arrhenius-type equations describe temperature dependence:
$$ k(T) = A \exp \left( – \frac{E_a}{RT} \right) $$
Calendar aging is often written as:
$$ Q_{loss} = B \exp \left( – \frac{E_a}{RT} \right) t^z $$
Cycle aging may be represented as:
$$ Q_{loss} = \alpha N^b + \beta t^c $$
I also use Markov chains to represent transitions among health stages:
$$ P_{ij} = P(X_{t+1} = j \mid X_t = i) $$
In my view, mathematical models are useful for screening and lifetime projection, but they are not a substitute for direct measurement. Their main risk is hidden extrapolation. A retired lithium-ion cell may have experienced a usage pattern not represented in the fitting data.
Data-Driven Methods
I now consider data-driven methods. These methods learn a mapping from measured features to SoH without requiring an explicit physical model. In my experience, they are powerful for complex duty cycles and multi-factor coupling, but they are vulnerable to dataset shift. A model trained on one retired lithium-ion cell population may fail on another because the aging distribution has changed.
Feature Engineering for SoH
Before I train any model, I extract features. For a lithium-ion cell, I commonly use charge curve features, voltage plateau statistics, incremental capacity peaks, differential voltage peaks, impedance features, temperature rise, and rest voltage relaxation. I summarize these in Table 5.
| Feature family | Examples | Physical link | Online feasibility |
|---|---|---|---|
| Voltage curve | Mean voltage, slope, curvature | Polarization and capacity | High |
| Incremental capacity | Peak height, peak position, area | Phase transitions and lithium loss | Medium |
| Differential voltage | Peak shift, valley depth | Electrode balancing | Medium |
| Impedance | Ohmic, charge-transfer, diffusion | Interfacial aging | Low to medium |
| Thermal | Temperature rise, gradient | Heat generation and safety | High |
| Relaxation | Time constant, voltage decay | Diffusion and self-discharge | Medium |
In my first-person workflow, I prefer features that are robust to sensor offset and temperature. I also normalize features by cell chemistry, nominal capacity, and operating conditions. Without normalization, a data-driven model may learn the test bench rather than the retired lithium-ion cell.
Neural Networks
Neural networks are my most flexible nonlinear tool. A feedforward network can be written as:
$$ y = \sigma(W_L \cdots \sigma(W_1 x + b_1) + b_L) $$
For temporal data, I use recurrent networks, long short-term memory, bidirectional long short-term memory, and temporal convolutional networks. A typical training loss is:
$$ \mathcal{L} = \frac{1}{N} \sum_{i=1}^{N} (SoH_i – \hat{SoH}_i)^2 + \lambda \|\theta\|_2^2 $$
I have found that combining incremental capacity analysis with bidirectional long short-term memory improves accuracy because the feature extractor captures electrode-level changes and the sequence model captures trajectory dependence. However, I do not treat neural networks as black boxes. I inspect saliency, permutation importance, and error by operating condition. If a model relies on a single temperature artifact, I reject it.
Fuzzy Logic
Fuzzy logic is useful when I need interpretable rules under uncertainty. I define membership functions:
$$ \mu_A(x): X \rightarrow [0,1] $$
A typical rule for a retired lithium-ion cell is:
$$ \text{If } R_o \text{ is high and } R_{ct} \text{ is high, then } SoH \text{ is low.} $$
I use fuzzy inference when the relationship between impedance and capacity is not crisp. The method handles parameter drift and multi-condition sensitivity well. Its weakness is rule explosion: as I add features, the number of rules grows quickly. I therefore use fuzzy logic as a fusion layer rather than a standalone estimator.
Regression and Probabilistic Learning
Regression methods remain my most reliable data-driven baseline. Gaussian process regression is especially attractive because it provides uncertainty:
$$ f(x) \sim \mathcal{GP}(m(x), k(x,x’)) $$
$$ k(x,x’) = \sigma_f^2 \exp \left( – \frac{(x-x’)^2}{2l^2} \right) $$
$$ \mu_* = k_*^T (K + \sigma_n^2 I)^{-1} y $$
$$ \sigma_*^2 = k_{**} – k_*^T (K + \sigma_n^2 I)^{-1} k_* $$
I use Gaussian process regression when I have limited data and need calibrated confidence intervals. For larger datasets, I use random forests, gradient boosting, and support vector regression. I evaluate all models with:
$$ RMSE = \sqrt{\frac{1}{N} \sum_{i=1}^{N} (SoH_i – \hat{SoH}_i)^2} $$
$$ MAE = \frac{1}{N} \sum_{i=1}^{N} |SoH_i – \hat{SoH}_i| $$
$$ R^2 = 1 – \frac{SS_{res}}{SS_{tot}} $$
I also report maximum error and bias, because a low RMSE can hide a dangerous tail. For retired lithium-ion cells, the worst-case error matters more than the average error.
Hybrid and Fusion Architectures
In my own view, no single method is sufficient. Direct methods provide ground truth, model-based methods provide structure, and data-driven methods provide flexibility. I therefore design hybrid architectures. A general fusion model is:
$$ \hat{SoH}_{fusion} = \sum_{m=1}^{M} w_m \hat{SoH}_m + b $$
where the weights can be static, adaptive, or learned. I have tested feature-level fusion, decision-level fusion, and parameter-level fusion. Table 6 summarizes my assessment.
| Fusion level | Mechanism | Advantage | Risk |
|---|---|---|---|
| Feature-level | Concatenate voltage, impedance, thermal features | Rich representation | Dimensionality and noise |
| Model-level | Use ECM parameters as NN inputs | Physical regularization | Parameter identification error |
| Decision-level | Weighted average of estimators | Robust and simple | Weight tuning |
| Physics-informed | Add physical loss to training | Better extrapolation | Complex loss design |
| Transfer-level | Adapt source model to target cell | Small-sample capability | Negative transfer |
I also use probabilistic fusion. If each estimator provides a mean and variance, I combine them with inverse-variance weighting:
$$ w_m = \frac{1/\sigma_m^2}{\sum_{j=1}^{M} 1/\sigma_j^2} $$
This is my preferred approach when the retired lithium-ion cell population is heterogeneous. It prevents an overconfident model from dominating the result.
Uncertainty Quantification and Risk-Aware Sorting
I never deploy a SoH estimator for retired lithium-ion cells without uncertainty quantification. The sorting decision is often a threshold decision: reuse, remanufacture, recycle, or reject. If the uncertainty is wide, I must be conservative. I define the lower confidence bound as:
$$ SoH_{LCB} = \hat{SoH} – z_{1-\alpha} \sigma_{SoH} $$
I then sort by the lower confidence bound rather than the point estimate. This reduces the risk of placing a weak lithium-ion cell into a demanding second-life application. I also use Bayesian neural networks and Gaussian process regression because they provide posterior predictive variance. For high-volume sorting, I approximate the posterior with Monte Carlo dropout or ensembles.
Small-Sample and Transfer Learning
The most frustrating problem I encounter is the small-sample bottleneck. Retired lithium-ion cells often lack complete labels. Full capacity tests are expensive, and destructive tests are impossible for reused cells. I therefore use transfer learning. A source model is trained on laboratory cells, and a target model is adapted to field-retired lithium-ion cells. A generic transfer objective is:
$$ \mathcal{L}_{trans} = \mathcal{L}_{source} + \lambda \mathcal{L}_{target} + \gamma \Omega(\theta) $$
I use fine-tuning, domain adaptation, adversarial alignment, and meta-learning. In my experience, transfer learning works best when the source and target share chemistry but differ in duty cycle. It works poorly when the chemistry changes, because the aging mechanisms may not be shared. I summarize my transfer strategy in Table 7.
| Scenario | Recommended method | Expected gain | Main failure mode |
|---|---|---|---|
| Same chemistry, different duty | Fine-tuning with frozen early layers | High | Overfitting target noise |
| Same format, different chemistry | Domain adaptation plus impedance features | Medium | Negative transfer |
| Very few labels | Meta-learning and Gaussian processes | Medium to high | High variance |
| No labels | Self-supervised and physics-informed learning | Low to medium | Unidentifiable health |
Multi-Physics Coupling and Digital Twins
I believe the next generation of SoH estimation must couple electrochemistry, thermal behavior, and mechanics. A lithium-ion cell ages through chemical, thermal, and mechanical pathways that interact. I write a simplified thermal balance as:
$$ \rho C_p \frac{\partial T}{\partial t} = \nabla \cdot (k \nabla T) + Q_{ohm} + Q_{rev} + Q_{entropy} $$
$$ Q_{ohm} = I^2 R_o $$
$$ Q_{rev} = I T \frac{\partial U}{\partial T} $$
Mechanical stress affects diffusion and degradation:
$$ \sigma = E \epsilon – \Omega c $$
$$ \frac{\partial c}{\partial t} = D \nabla^2 c – \frac{\Omega}{RT} \nabla \cdot (c \nabla \sigma_h) $$
In my digital twin concept, the retired lithium-ion cell is represented by a state-space model that updates with field data:
$$ \hat{SoH}_{k+1} = \mathcal{F}(\hat{SoH}_k, u_k, T_k, \theta_k) + K_k (y_k – \hat{y}_k) $$
The digital twin does not replace measurement. It organizes measurement into a coherent, updatable health record. I see this as the most promising route to scalable second-life certification.
Sensor and Hardware Innovations
From my engineering perspective, algorithm advances are limited by sensor quality. I need reliable current, voltage, and temperature at the cell level. I also need new sensing modalities. Table 8 summarizes the sensors I consider relevant.
| Sensor type | Measured quantity | Value for SoH | Deployment challenge |
|---|---|---|---|
| Precision shunt | Current | Accurate coulomb counting | Thermal drift |
| Cell voltage monitor | Voltage | OCV and polarization | Channel count |
| Thermistor array | Temperature | Thermal aging and safety | Placement and gradients |
| Impedance-temperature composite | Impedance and temperature | Online EIS proxy | Calibration complexity |
| Surface acoustic wave | Mechanical and thermal | Early degradation | Packaging and cost |
| Fiber-optic sensor | Strain and temperature | Mechanical aging | Integration |
| Gas sensor | Ventilation products | Safety and abuse | False positives |
I also emphasize sensor fusion. A single temperature sensor cannot capture the thermal gradient inside a large lithium-ion cell. A single voltage sensor cannot separate ohmic and diffusion effects. Therefore, I use multiple modalities and a Bayesian estimator.
Comparative Assessment of Major Methods
I now provide my direct comparison. Table 9 summarizes accuracy, speed, generalization, and industrial readiness.
| Method | Typical accuracy | Speed | Generalization | Industrial readiness |
|---|---|---|---|---|
| Capacity test | \(\pm 1\%\) to \(\pm 3\%\) | Very low | High | Reference only |
| HPPC and DC pulse | \(\pm 3\%\) to \(\pm 8\%\) | Medium | Medium | High |
| EIS | \(\pm 3\%\) to \(\pm 10\%\) | Low to medium | Medium | Medium |
| ECM with KF | \(\pm 2\%\) to \(\pm 6\%\) | High | Low to medium | High |
| Electrochemical model | \(\pm 2\%\) to \(\pm 5\%\) | Low | Medium | Research to pilot |
| Neural network | \(\pm 1\%\) to \(\pm 5\%\) | High | Low to medium | Medium |
| Gaussian process regression | \(\pm 2\%\) to \(\pm 6\%\) | Medium | Medium | Medium |
| Hybrid physics-data | \(\pm 1\%\) to \(\pm 4\%\) | Medium to high | Medium to high | Emerging |
I caution that these ranges are context dependent. The accuracy of any method on a retired lithium-ion cell depends on the training set, the age distribution, the temperature range, and the reference standard. A method that achieves excellent accuracy on a narrow laboratory dataset may fail on an industrial sorting line.
Challenge One: The Efficiency-Precision Trade-Off
The first challenge I identify is the efficiency-precision trade-off. Capacity testing is accurate but slow and destructive to remaining life. Impedance testing is fast but indirect. I express the trade-off as a constrained optimization:
$$ \min_{\mathcal{M}} \left( \alpha E_{SoH} + \beta T_{test} + \gamma C_{test} \right) $$
subject to:
$$ E_{SoH} \leq E_{max} $$
$$ T_{test} \leq T_{max} $$
where \(\mathcal{M}\) is the estimation method, \(E_{SoH}\) is estimation error, \(T_{test}\) is test time, and \(C_{test}\) is cost. In my view, the industry needs a tiered protocol. Tier one uses fast impedance and voltage features to screen all retired lithium-ion cells. Tier two uses partial charge-discharge and model-based estimation for borderline cells. Tier three uses full capacity and electrochemical diagnostics for high-value or safety-critical cells.
Challenge Two: Generalization Across Chemistries and Formats
The second challenge is generalization. A model trained on one lithium-ion cell chemistry may not transfer to another. The open-circuit voltage curve, phase transitions, and degradation modes differ. I write the domain shift as:
$$ P_{source}(X, Y) \neq P_{target}(X, Y) $$
To address this, I use domain-invariant features, chemistry-aware normalization, and meta-learning. I also use physical constraints. For example, I enforce monotonic capacity fade under identical conditions:
$$ \frac{\partial SoH}{\partial N} \leq 0 $$
I enforce resistance growth:
$$ \frac{\partial R_o}{\partial N} \geq 0 $$
These constraints reduce the hypothesis space and improve transfer. In my experience, physics-informed regularization is more effective than purely data-driven domain adaptation when labels are scarce.
Challenge Three: Black-Box Models and Trust
The third challenge is trust. A black-box model may achieve low error but provide no explanation. For retired lithium-ion cells, I need to know why a cell is classified as healthy or unhealthy. I use explainable AI methods such as Shapley values, partial dependence, and counterfactual explanations. I also use hybrid models where the first layer is physical and the second layer is statistical. This gives me interpretable parameters and flexible residual correction:
$$ SoH = SoH_{physics}(\theta) + \Delta SoH_{data}(x) $$
In my view, the most trustworthy estimator is one that can say, “This lithium-ion cell is at 78% SoH because its ohmic resistance increased by 32%, its charge-transfer resistance increased by 45%, and its incremental capacity peak shifted by 18 mV.” That kind of explanation is essential for certification.
Challenge Four: Small Samples and Label Scarcity
The fourth challenge is label scarcity. I cannot perform full capacity tests on every retired lithium-ion cell. I therefore use semi-supervised learning, active learning, and self-supervised learning. In active learning, I select the most informative cells for full testing:
$$ x^* = \arg\max_{x} \left( \sigma_{SoH}(x) + \lambda \rho(x, \mathcal{D}) \right) $$
where \(\sigma_{SoH}\) is predictive uncertainty and \(\rho\) is diversity. I also use data augmentation based on physical transformations, such as temperature scaling and current scaling. However, I must ensure that augmented data remain physically plausible for a lithium-ion cell.
Challenge Five: Safety and Second-Life Certification
The fifth challenge is safety. A retired lithium-ion cell may have localized lithium plating, separator damage, or gas generation that is not captured by capacity alone. I therefore include safety features in the SoH vector. I define a safety margin:
$$ M_{safe} = \min \left( \frac{T_{vent} – T_{max}}{T_{vent}}, \frac{V_{max} – V_{cell}}{V_{max}}, \frac{R_{threshold} – R_o}{R_{threshold}} \right) $$
I reject any cell with \(M_{safe} \leq 0\). I also use electrochemical impedance to detect soft shorts and gas evolution. In my experience, safety screening must be independent of capacity screening. A lithium-ion cell with 80% capacity can still be unsafe.
My Proposed Integrated Framework
I propose an integrated framework that combines direct measurement, model-based estimation, and data-driven correction. The framework has four layers:
Layer one is rapid screening. I measure open-circuit voltage, temperature, and a short multi-frequency impedance signature. I compute a preliminary health score:
$$ SoH_{screen} = f_1(OCV, R_o, T, \Delta V_{relax}) $$
Layer two is model-based initialization. I fit an equivalent circuit model or a reduced electrochemical model to a partial charge-discharge curve. I estimate:
$$ \theta_0 = [R_o, R_p, C_p, C_{usable}]^T $$
Layer three is data-driven correction. I feed the model parameters and raw features into a hybrid estimator:
$$ \Delta SoH = f_2(\theta_0, x_{voltage}, x_{thermal}, x_{impedance}) $$
Layer four is uncertainty-aware decision. I compute the posterior distribution and sort by lower confidence bound:
$$ SoH_{final} = SoH_{model} + \Delta SoH_{data} $$
$$ SoH_{decision} = SoH_{final} – z_{1-\alpha} \sigma_{final} $$
This framework is my answer to the three-way requirement of accuracy, speed, and robustness. It uses the strengths of each method and compensates for their weaknesses.
Experimental Design for Validation
I cannot claim a method works unless I validate it on a representative retired lithium-ion cell population. I therefore design experiments with the following variables: chemistry, format, nominal capacity, calendar age, cycle number, average temperature, depth of discharge, charge rate, discharge rate, and storage SoC. I summarize my validation matrix in Table 10.
| Factor | Levels I test | Purpose |
|---|---|---|
| Chemistry | NMC, LFP, LCO, NCA | Chemistry generalization |
| Format | Cylindrical, prismatic, pouch | Thermal and mechanical effects |
| Temperature | 0, 15, 25, 35, 45 °C | Kinetic and diffusion effects |
| Current | 0.2C, 0.5C, 1C, 2C | Polarization and heating |
| Depth of discharge | 20%, 50%, 80%, 100% | Mechanical stress |
| Storage SoC | 30%, 50%, 80% | Calendar aging |
| Cycle number | 0 to 2000 | Trajectory diversity |
For each retired lithium-ion cell, I record the reference capacity, HPPC resistance, EIS spectrum, and incremental capacity curve. I then compare every estimator against the reference. I use cross-validation by cell, not by sample, to avoid data leakage. I also use leave-one-chemistry-out validation to test generalization.
Mathematical Summary of Error Propagation
I often need to propagate measurement error into SoH error. If SoH is a function of several measured variables:
$$ SoH = g(x_1, x_2, \ldots, x_n) $$
then the first-order uncertainty is:
$$ \sigma_{SoH}^2 = \sum_{i=1}^{n} \left( \frac{\partial g}{\partial x_i} \right)^2 \sigma_{x_i}^2 + 2 \sum_{i<j} $$=""
For capacity estimation from current integration, the error is:
$$ \sigma_C^2 = \left( \int \sigma_I dt \right)^2 + \left( \int I \sigma_{\eta} dt \right)^2 + \sigma_{timing}^2 $$
For resistance estimation from a pulse, the error is:
$$ \sigma_R^2 = \left( \frac{\sigma_{\Delta V}}{\Delta I} \right)^2 + \left( \frac{\Delta V \sigma_{\Delta I}}{\Delta I^2} \right)^2 $$
I use these equations to decide which sensor to improve. If current sensor error dominates, I invest in a better shunt. If voltage error dominates, I improve the measurement front end. If temperature error dominates, I improve thermal sensing. This is a practical use of error propagation for retired lithium-ion cell sorting.
Algorithmic Selection Guide
I provide my own selection guide in Table 11. The table is based on the assumption that the user has limited time and limited labels.
| Condition | Preferred method | Reason |
|---|---|---|
| High-throughput screening | Short pulse resistance plus OCV | Fast and low cost |
| Known chemistry and format | ECM with Kalman filter | Accurate and online |
| Unknown history | EIS plus Gaussian process regression | Rich features and uncertainty |
| Small labeled dataset | Physics-informed neural network | Regularization and transfer |
| Safety-critical application | Full capacity plus EIS plus thermal test | Reference-grade confidence |
| Field-deployed battery management | Reduced-order ECM plus adaptive data correction | Computation and robustness |
Future Research Directions I Prioritize
I prioritize five research directions. First, multi-physics coupling. I want models that combine electrochemistry, thermal transport, and mechanics in a computationally tractable form. Second, physics-informed machine learning. I want neural networks that respect conservation laws and monotonic degradation. Third, intelligent sensing. I want low-cost impedance, strain, and gas sensors that can be embedded in a pack. Fourth, transfer learning and meta-learning. I want a model that can adapt to a new retired lithium-ion cell population with ten labeled cells rather than ten thousand. Fifth, uncertainty-aware certification. I want standards that accept a probabilistic SoH statement rather than a single number.
I also see a role for reinforcement learning in test protocol design. Instead of applying a fixed HPPC profile, I can learn a sequence of pulses that maximizes information gain per unit time. The objective is:
$$ \max_{\pi} \mathbb{E} \left[ I(SoH; y_{1:T} \mid \pi) – \lambda T \right] $$
where \(I(\cdot)\) is mutual information and \(\pi\) is the test policy. This is a promising way to shorten testing while preserving accuracy.
Data and Standardization Needs
I cannot overstate the need for standardized data. Retired lithium-ion cell datasets are fragmented. Different laboratories use different protocols, different reference capacities, and different temperature definitions. I propose a minimum metadata schema: chemistry, format, nominal capacity, nominal voltage, cutoff voltages, cycle history, temperature history, storage history, and reference test protocol. I also propose a standard SoH reporting format that includes capacity SoH, resistance SoH, and uncertainty. Table 12 shows my proposed schema.
| Field | Example value | Purpose |
|---|---|---|
| Chemistry | NMC 811 | Model selection |
| Format | 21700 cylindrical | Thermal modeling |
| Rated capacity | 4.8 Ah | SoH reference |
| Reference capacity | 3.9 Ah | Capacity SoH |
| Reference resistance | 28 mΩ | Resistance SoH |
| Temperature history | 25 °C mean, 45 °C max | Calendar aging correction |
| Cycle history | 850 equivalent full cycles | Cycle aging correction |
| Uncertainty | \(\pm 2.1\%\) | Risk-aware sorting |
Economic and Environmental Implications
In my analysis, accurate SoH estimation directly affects economics. If I underestimate SoH, I prematurely recycle a usable lithium-ion cell and lose value. If I overestimate SoH, I risk early failure and warranty cost. I define the expected value of a sorting decision as:
$$ EV = P(SoH > SoH_{threshold}) V_{reuse} + P(SoH \leq SoH_{threshold}) V_{recycle} – C_{test} – C_{error} $$
I want to maximize \(EV\). This equation shows that uncertainty has a cost. Reducing \(\sigma_{SoH}\) increases \(EV\) because it reduces misclassification. The environmental benefit follows the same logic: extending the life of a lithium-ion cell by three to five years reduces life-cycle carbon emissions by more than forty percent in many scenarios. Therefore, SoH estimation is not only a technical problem; it is an economic and environmental lever.
Discussion of Limitations in My Review
I acknowledge several limitations in my own synthesis. First, I have not exhausted every published method, because the field is expanding rapidly. Second, I have emphasized methods that I can implement with available sensors and computing. Third, I have not provided a single universal accuracy number, because accuracy is always conditional on the dataset and reference. Fourth, I have focused on lithium-ion cells, but other chemistries such as sodium-ion and solid-state cells may require different estimators. Fifth, I have not fully addressed second-life pack-level effects, such as cell-to-cell imbalance and interconnection resistance. I include pack-level SoH as a future extension.
Pack-Level SoH and Cell-to-Cell Variation
When I move from a single lithium-ion cell to a pack, the problem changes. A pack’s SoH is limited by its weakest cell and by its imbalance. I define pack SoH as:
$$ SoH_{pack} = \min_i SoH_i – \kappa \max_{i,j} |SoH_i – SoH_j| $$
where \(\kappa\) penalizes heterogeneity. I also use the ratio of available pack capacity to nominal pack capacity:
$$ SoH_{pack,C} = \frac{\min_i C_i}{C_{nom}} $$
For a retired lithium-ion cell pack, I measure individual cell voltages, temperatures, and impedances. I then use a hierarchical estimator. The cell-level estimators produce a distribution, and the pack-level estimator aggregates the distribution. This is more robust than estimating pack SoH from pack voltage alone.
Final Synthesis
I conclude that retired lithium-ion cell SoH estimation is moving from single-dimension measurement to multi-modal fusion and from experience-driven modeling to data-physics co-driven modeling. Direct measurement remains the foundation for accuracy. Model-based methods remain the core for mechanistic insight. Data-driven methods remain the利器 for complex operating conditions, but they must be constrained by physics and uncertainty. In my first-person assessment, the winning architecture will be a tiered, uncertainty-aware hybrid that uses fast screening, adaptive model parameters, and transferable machine learning. It will estimate not only a point value but also a confidence interval. It will be validated on a diverse population of retired lithium-ion cells. It will be supported by standardized metadata and intelligent sensors. Most importantly, it will be practical enough for industrial sorting lines and safe enough for second-life deployment. I therefore recommend that future work focus on multi-physics coupling, physics-informed learning, small-sample transfer, explainable models, and standardized validation. These are the pillars that will turn retired lithium-ion cells from an uncertain liability into a reliable, sustainable, and economically valuable resource.