A Review of Thermal Safety and Management of Second-Life Batteries: Cell Screening, Pack Configuration and Health Estimation

Abstract

Electric vehicle (EV) adoption is generating a rapidly increasing stream of retired lithium-ion batteries for second-life deployment. However, thermal safety concerns continue to limit their reuse. This paper reviews second-life battery (SLB) thermal safety and management and organizes existing work through a mechanism-to-deployment framework linking four domains: degradation mechanisms, cell screening, pack configuration, and monitoring. Evidence indicates that thermal risk depends on the degradation pathway rather than capacity fade. In fact, cells with comparable capacity can exhibit substantially different trigger temperatures depending on whether lithium plating or solid-electrolyte interphase (SEI) growth dominates. Therefore, capacity-based screening is insufficient because cells that satisfy capacity thresholds may still remain thermally unstable. The four domains are tightly coupled: the degradation pathway determines screening requirements; screening outcomes constrain pack design; pack topology influences fault escalation; and together these factors determine what monitoring can reliably detect. This review highlights three gaps and outlines future research directions in the field of SLB thermal safety and management: limited aged-cell thermal characterization by degradation pathway, insufficient diagnostic validation under industrial-throughput conditions, and the incomplete translation of screening outputs into design rules.

1. Introduction

1.1. Second-Life Battery Markets and Applications

Global EV adoption is accelerating rapidly, with the worldwide EV fleet reaching approximately 58 million vehicles by the end of 2024 and projected to grow to 250 million by 2030 [1]. As EV batteries are typically retired at 70–80% of their original capacity, this growth will generate an increasing stream of retired packs with diverse chemistries and degradation histories. Based on a scenario analysis by the International Council on Clean Transportation (ICCT), in which 50% of end-of-life EV batteries are reused for stationary energy storage, the retired battery capacity available for second-life applications could reach ∼96 GWh by 2030, increase to ∼3000 GWh by 2040, and exceed ∼12,000 GWh by 2050 [2]. Second-life battery energy storage systems (SL-BESSs) can reduce investment costs by up to 25% compared to new batteries while contributing to decarbonization goals and increasing power system flexibility [3]. However, large-scale adoption is constrained by safety concerns arising from unknown first-life usage histories and the resulting uncertainty in degradation state, performance, and reliability [4]. Thermal runaway (TR) remains a primary technical barrier, as aged cells can exhibit reduced thermal stability margins that vary with degradation pathway [5,6]. Ensuring safe reuse therefore requires thermal safety evaluation and management methods specifically adapted to retired battery characteristics.

1.2. Thermal Safety Evaluation Methods for Second-Life Batteries

Thermal safety evaluation for SLBs commonly combines state estimation and screening diagnostics to assess battery condition and identify cells with elevated risk. In practice, this includes estimation of the state of charge (SOC), state of health (SOH), and remaining useful life (RUL) using model-based, feature-based, and data-driven approaches [7,8]. In this review, SOH is defined as the percentage of discharge capacity remaining relative to the initial discharge capacity, $\mathrm{SOH}=(Q/Q_0)\times 100\%$, where Q is the current discharge capacity and $Q_0$ is the initial discharge capacity; capacity loss is $100\%-\mathrm{SOH}$. Capacity-based SOH is used throughout for consistency with published thermal-safety reporting [6]; however, energy-based metrics (usable kWh) may consider more informative for grid applications where discharge profiles vary. SOC estimation supports safe operating boundaries, SOH quantifies degradation severity, and RUL forecasts the remaining window for safe operation [7]. Screening diagnostics further support high-throughput qualification by using capacity, resistance/impedance, and voltage-response features to capture aging-related changes that influence heat generation and thermal margins [9,10]. However, retired batteries exhibit substantial heterogeneity in degradation pathways, and capacity-based SOH alone may not reliably indicate TR initiation sensitivity, particularly when lithium plating or other heterogeneous aging pathways are present [6,11]. A cell that passes capacity-based screening may therefore remain electrically functional yet thermally unstable, creating latent hazards in SL-BESS.

1.3. Thermal Management Methods for Second-Life Batteries

Thermal management systems (TMSs) maintain battery temperature within safe operating limits and minimize cell-to-cell thermal gradients. Common approaches include air cooling, liquid cooling, phase change materials (PCMs), and hybrid designs. Selection depends on heat rejection capability, cost, packaging constraints, and temperature uniformity requirements [12]. Battery management systems (BMSs) complement TMSs by providing real-time monitoring through voltage, current, and temperature measurements, supporting state estimation, and by enforcing protection actions to maintain safe operating boundaries [13]. Digital twin approaches integrate operational data with physics-based or data-driven models to enable online assessment, adaptive control, and decision support [14]. For SLBs, such methods have been proposed for suitability assessment and regrouping decisions by combining operational history with model predictions [15]. Compared with new batteries, SLB packs exhibit higher heat generation rates and substantially larger cell-to-cell variability, requiring TMS and BMS designs that account for evolving thermal loads and heterogeneous degradation states [6,16].

1.4. Related Reviews and Contributions of This Paper

Several reviews address SLB assessment, deployment, and safety.

Table 1. Recent reviews relevant to SLB and thermal safety.

Ref.	Year	Focus	Level	SLB Thermal Safety	Scope/Key Notes
[17]	2022	SLB techno-economics	Grid/ESS	Limited	Economic viability and grid services; thermal safety outside scope.
[18]	2022	Li-ion TR mitigation	Module/Pack	Partial	TR initiation, propagation; emphasis on new-cell contexts.
[19]	2022	Aged Li-ion safet	Cell	Partial	Extensive aged-cell thermal data compilation; SLB-specific deployment not treated.
[13]	2023	SLB grid integration	Grid/ESS	Limited	Power-electronics interfaces; thermal safety not treated in detail.
[9]	2024	SLB diagnostics	ESS	Limited	SOH diagnostics for repurposing; thermal safety not primary focus.
[20]	2024	SLB technologies	ESS	Partial	Degradation modeling and techno-economics; compares plating vs SEI, but TR evidence not central.
[4]	2024	SLB characterization	ESS	Partial	Discusses aging-related risk indicators; pack-level propagation not treated.
[16]	2025	SLB prognosis	ESS	Limited	Prognosis and RUL under data scarcity; thermal safety not addressed.
This review	—	SLB thermal safety	Cell–Pack	Full	Aging-dependent TR by pathway; screening-to-pack linkage; topology and fault escalation; venting behavior; SLB-oriented monitoring.

SLB thermal-safety coverage: Limited = thermal safety or TR not explicitly treated; Partial = TR discussed but does not connect SLB aging to TR behavior or omits pack-level aspects; Full = integrates aging-linked TR mechanisms with screening-to-pack linkage and pack-level propagation, venting, and monitoring.

compares representative works, ordered by publication year, in terms of primary focus, scope level, and treatment of key topics relevant to thermal safety.

As shown in Table 1, existing reviews primarily address either techno-economic feasibility or thermal runaway mechanisms and mitigation strategies, with limited integration between these domains. Domain-specific reviews offer valuable insights into state estimation, diagnostics, characterization, or prognosis but treat screening, topology selection, and monitoring as independent problems rather than coupled stages in a safety workflow.

This review addresses this integration need by organizing the literature around four interdependent areas: (1) degradation mechanisms and their effects on thermal runaway initiation; (2) screening methods that detect thermal hazards before reuse; (3) pack configuration and thermal management that constrain fault escalation; and (4) state estimation and monitoring that track safety margins during operation. Figure 1 illustrates this coupled framework, where each stage constrains the next and weaknesses propagate forward. If degradation mechanisms are not considered (Section 2), screening may apply inappropriate criteria (Section 3). If screening fails to flag a hazardous cell, neither pack design (Section 4) nor monitoring strategies (Section 5) can eliminate the underlying risk. In practice, the degradation mode informs screening needs, screening outputs constrain pack design, topology governs fault escalation, and these architectural choices collectively determine what aspects of cell behavior can be observed or inferred during operation. Consequently, ensuring second-life safety depends on identifying and managing the prevailing degradation pathway rather than relying solely on capacity-based measurements.

The paper is organized as follows. Section 2 discusses degradation mechanisms and their effects on TR initiation and severity. Section 3 reviews screening methods for detecting thermal hazards. Section 4 addresses pack configuration, including cell matching, topology, propagation mitigation, and thermal management. Section 5 covers state estimation and fault detection. Section 6 discusses the key findings, identifies knowledge gaps, and outlines future directions. Section 7 presents the conclusions.

2. Degradation Mechanisms and Thermal Safety

TR behavior is commonly characterized by initiation sensitivity metrics such as self-heating onset temperature ($T_{\mathrm{onset}}$) and trigger temperature ($T_{\mathrm{trigger}}$), and by severity metrics such as peak temperature ($T_{max}$), total heat release, and peak heat-release rate. For SLB, these metrics vary with degradation pathway in ways that capacity or resistance alone cannot predict. Figure 2 maps the primary mechanism-to-consequence pathways. Multiple pathways can operate simultaneously, and cross-links between them mean that aging rarely follows a single isolated route as discussed throughout this section.

This section first reviews how aging affects TR severity and initiation sensitivity, then discusses three mechanisms, resistive heating, lithium plating, and gas–mechanical coupling, that differentially impact these metrics. Each mechanism produces distinct observable indicators through different internal pathways.

2.1. Aging Effects on TR Severity and Initiation Sensitivity

Severity refers to how destructive a thermal-runaway event is, typically characterized by peak temperature and total heat release. Initiation sensitivity describes how easily thermal runaway begins, commonly assessed by the temperature at which self-heating or runaway first initiates. Figure 3 shows the central relationship reviewed in this section; while TR severity generally decreases with capacity loss, initiation sensitivity diverges depending on the dominant degradation mechanism. Several studies have reported that aging reduces TR severity. Deeply aged cells exhibit lower $T_{max}$ and reduced peak heat-release rates compared with new batteries. First-life operation consumes active lithium and electrolyte, decreasing stored energy; this reduction in severity is consistent with lower total energy available for release during failure [21,22]. This decline is reflected in the severity trend in Figure 3.

Capacity fade in retired cells arises through two principal degradation modes that affect thermal stability differently: loss of lithium inventory (LLI) and loss of active material (LAM). LLI results from cyclable lithium permanently consumed during SEI growth or immobilized as metallic deposits during plating. LAM is driven by structural and mechanical failures, such as electrode particle cracking, binder degradation, or the structural isolation of active regions, reducing the electrodes ability to host lithium [23,24,25]. Their thermal consequences differ. LLI through SEI growth consumes lithium in a stable passivation layer, whereas LLI through plating deposits reactive metallic lithium that lowers exothermic reaction barriers (Section 2.3). LAM reduces electrode structural integrity but does not introduce the low-temperature reactivity of metallic lithium. Capacity fade alone therefore cannot distinguish thermally benign from thermally hazardous degradation states.

Initiation sensitivity, however, does not follow the same trend and instead varies with the dominant degradation pathway. SEI-growth-dominated aging (high-temperature cycling and extended storage) consumes the lithium inventory and can partially suppress early-stage exothermic reactions, maintaining $T_{\mathrm{trigger}}$ near new-cell levels [11,26]. In 18650 NCA cells tested under identical protocols, SEI-growth-aged cells showed self-heating onset rising from 80 ± 5 °C to 98 ± 5 °C, with thermal runaway onset and venting temperatures unchanged; across NCA, NMC, and LFP chemistries, SEI-growth-aged cells exhibited similar or improved safety [27]. This corresponds to the flat SEI-dominated curve in Figure 3. In contrast, plating-dominated aging (low-temperature cycling, fast charging) deposits metallic lithium that lowers kinetic barriers for parasitic reactions [28]. Using the same cell type and protocols, plating-aged cells showed self-heating onset dropping from approximately 80 °C to 35 °C, with consistently higher self-heating rates and elevated European Council for Automotive R&D (EUCAR) hazard levels across all test methods including overcharge, overcurrent, and nail penetration [29]. These results reflect accelerated aging with substantial plating deposits; the critical amount of deposited lithium required to alter safety in field-aged cells remains unknown [29]. This sharp increase in sensitivity corresponds to the rising plating-dominated curve in Figure 3. Absolute onset values vary with cell type and chemistry; the relative shift within each study is the relevant safety indicator. Plating shifts thermal thresholds at multiple stages as summarized in

Table 2. Aging-related changes in thermal behavior and safety-relevant indicators.

Mechanism	New-Cell Behavior	Aged/Second-Life Behavior	Observable Indicators
Impedance heterogeneity	Uniform current distribution	Localized Joule heating, hotspots	Internal temperature gradients
Lithium plating	Lithium intercalated;   $T_{\mathrm{onset}}$ 80–116 °C [6,29,30]; $T_{\mathrm{trigger}}$ ∼$218 °$C [30]	Metallic Li reactive;  $T_{\mathrm{onset}}$ 35–82 °C [6,29,30]; $T_{\mathrm{trigger}}$ ∼$96 °$C [30]	$T_{\mathrm{onset}}$ drops 20–62 °C; $T_{\mathrm{trigger}}$ drops ∼$122 °$C
Venting	Venting at ∼130 °C [31]	Venting at ∼112 °C (plating) [31]	Earlier gas release (∼$18 °$C shift)
Gas + mechanical	SEI-formation gas; intact casing	Accelerated gas; weakened casing	Reduced margin to rupture

. In 18650 cells aged under low-temperature cycling, self-heating onset drops from 102 °C to 82 °C [6]. In pouch cells subjected to fast-charge aging, self-heating onset falls from 116.5 °C to 54.5 °C and thermal runaway trigger temperature from 217.6 °C to 95.8 °C at 50% SOH [30]. In automotive pouch cells, the venting temperature decreases from 130 °C to 112 °C when lithium plating is present [31]. Venting and gas–mechanical behavior also differ between new and aged cells as discussed in Section 2.4.

Plated lithium is not always permanent. Resting cells at room temperature after low-temperature cycling can allow reversibly plated lithium to re-intercalate, which shifts self-heating onset temperatures toward values observed in fresh cells [19]. In contrast, tests conducted immediately after plating–promoting cycling have reported onset temperatures as low as 35 °C [29], compared with 98 °C for SEI-growth-aged cells of the same type [27]. Both the degradation pathway and the interval between retirement and testing therefore affect screening outcomes. Capacity-based SOH alone cannot distinguish thermally benign from hazardous degradation states [32,33] as reflected in the shifts summarized in Table 2.

2.2. Resistive Heating: Localized Joule Heating from Heterogeneous Impedance

Internal resistance in retired cells increases heterogeneously rather than uniformly. While ohmic resistance contributes to power fade, aging is often dominated by non-uniform SEI thickening and the associated rise in charge-transfer resistance. These dgradation mechanisms accelerate once SOH falls below approximately 40% [34]. Morphological analysis of one cell type indicated SEI thickening from 23 nm (new) to 57.5 nm (82% SOH), accompanied by increased anode stiffness and irreversible mechanical swelling [35]. The growing SEI layer also reduces electrode porosity and effective ionic diffusion coefficients [23]. In large-format cells, degradation can initiate near current-collection regions and propagate toward the electrode interior, producing localized high-impedance zones that distort the internal current-density distribution [24]. These spatial gradients promote current constriction and localized Joule heating, as the effective reaction area decreases and current is forced through narrower conductive pathways.

As non-uniform SEI raises the local charge-transfer resistance, the current shifts toward lower-impedance regions, increasing local overpotentials and concentrating both ohmic and kinetic heat generation. This creates a self-reinforcing cycle: localized heating accelerates side reactions and SEI alteration in hot-spot regions, further increasing impedance heterogeneity [24]. The resulting internal temperature gradients compound the problem. Modeling studies showed that interelectrode thermal gradients shift electrode potentials and redistribute plating severity, with through-plane gradients being particularly critical [36]. Experimentally, even a temperature difference of only ±2 °C between electrodes has been shown to cause rapid degradation in cells operated at 35 °C and C/5, conditions that are otherwise considered safe under uniform thermal environments [37]. When the cathode was slightly warmer, lithium plating formed at the anode. When the anode was slightly warmer, degradation occurred at the cathode instead [37]. In retired cells, heterogeneous impedance growth naturally leads to these small internal thermal gradients, providing a direct mechanism that links impedance non-uniformity to the initiation of plating.

Cells with internal resistance exceeding 130% of fresh-cell values show increased heat generation during operation, driven primarily by amplified $I^{2}R$ losses [25]. This current redistribution produces internal temperature gradients in which core regions reach higher temperatures than the cell surface, a disparity often missed by external sensors but critical for hotspot initiation [38]. Heterogeneous impedance can also promote lithium plating, further reducing TR initiation thresholds.

2.3. Trigger Temperature Collapse: Lithium Plating and Kinetic Acceleration

Lithium plating in retired cells involves the deposition of metallic lithium on the anode surface. These deposits are typically porous with high surface area; they may become electronically isolated and electrochemically irreversible [39,40,41]. Unlike intercalated lithium, metallic lithium remains chemically reactive and is not effectively passivated by a stable SEI [42].

Its heightened reactivity arises from both morphology and surface state: plated lithium forms high-surface-area structures ranging from needle-like dendrites to porous crystals and mossy layers depending on temperature and cycling conditions [41,42], and residual lithium on the anode surface exists as highly reactive nanoclusters rather than in an ordered intercalated state [40]. This increased surface area amplifies the kinetic contact area for parasitic reactions with the electrolyte. The kinetic consequence is direct: exothermic decomposition rates follow Arrhenius-type temperature dependence, and cells aged near 0 °C showed activation energy of 0.813 eV compared with 0.856 eV for cells aged at 25 °C [28]. Thermokinetic analysis of Accelerating Rate Calorimeter (ARC) data showed that activation energy decreases linearly with the total plating energy, and onset temperature for adiabatic self-heating shows a strong inverse correlation with plating quantity [41]. This means that greater plating deposits systematically lower the thermal barrier, and the resulting heat generation can exceed passive dissipation at temperatures well below conventional thresholds.

The practical consequence is early-stage self-heating, with onset occurring at temperatures well below new-cell stability thresholds. As shown in Table 2, plating-dominated aging reduces self-heating onset from 80–116 °C (new) to 35–82 °C depending on cell format and aging conditions [6,29,30]. This early heat release arises from rapid reactions between electrolyte and non-intercalated lithium that can outpace passive heat dissipation [43]. In situ observations further indicated that residual lithium nanoclusters react with polymer binders (e.g., PVDF) to generate flammable hydrogen, elevating hazard during abuse or failure progression [40]. Trigger temperature also collapses, falling from approximately 218 °C (fresh) to 96 °C under plating-dominated aging (Table 2). Lithium plating therefore makes conventional stability benchmarks (e.g., separator melting temperature) insufficient for bounding TR initiation in retired cells [30,31]. Concurrent gas generation and mechanical degradation, discussed in Section 2.4, further narrow the safety margin between pressure buildup and structural failure.

2.4. Gas Generation and Mechanical Degradation

Gas-generation mechanisms differ between new and aged cells. New cells generate gas primarily during early SEI formation, whereas aged cells experience cathode destabilization that releases transition metals. This dissolution accelerates cathode decomposition and promotes oxygen release [44]. Concurrently, transition metal dissolution reduces thermal runaway trigger temperatures, while lithium plating and proton reduction side reactions generate reductive gases such as hydrogen and methane [45]. The accumulation of decomposition products (including PF5) further intensifies this process by reducing activation barriers for solvent ring-opening reactions and increasing gas-generation rates [46]. Real-time measurements in cylindrical formats showed internal gas pressure rising to approximately 20 bar at venting onset, with secondary peaks reaching 28.7 bar during later failure stages [47].

The mechanical integrity of retired cells degrades due to fatigue accumulated during first-life cycling. Cells aged under conditions promoting lithium plating showed casing stiffness reductions of approximately 16% relative to new counterparts, along with irreversible thickness swelling of approximately 9.1% [42]. This swelling generates persistent internal stress within the jellyroll and lowers resistance to deformation under elevated pressure [35]. The relationship between stiffness loss and burst-pressure reduction across different SOH values has not been systematically studied in the literature.

Together, accelerated gas generation and mechanically weakened containment narrow the margin between pressure buildup and structural failure in aged cells. Each mechanism reviewed in this section produces distinct observable indicators (Table 2) that inform the screening methods discussed in Section 3.

3. Cell Screening and Grading: Detecting Thermal Hazards

Screening and grading form the first line of defense for safe second-life battery deployment. Four diagnostic layers address different aspects of thermal risk, each targeting distinct hazards but also carrying limitations. Figure 4 maps this progressive filtering workflow: incoming retired cells with unknown histories pass through electrical screening, impedance-based assessment, structural and thermal testing, and data-driven risk scoring. At each stage, cells are passed, flagged for manual review, or rejected. Screening outputs directly constrain pack assembly and operational monitoring decisions (Section 4 and Section 5).

3.1. Electrical Screening: Proxies for Heat and Instability

Figure 5 maps the trade-off between screening throughput and thermal-hazard detection capability across the methods reviewed in this section. Faster methods such as electrical DC pulse resistance offer high throughput at low cost but limited thermal-risk visibility, while methods such as ultrasound spectroscopy and computed tomography (CT) can provide higher hazard detection at greater time and expense. The following sections discuss each method, starting with electrical screening.

DC pulse tests and voltage-curve analysis detect where heat generates and whether cells show electrochemical instability. For thermal safety, the objective extends beyond SOH estimation to identifying proxies of initiation sensitivity that capacity metrics alone would miss [6,48]. Heterogeneous impedance growth in aged packs creates current constriction and concentrated Joule losses, while interconnect resistance in retired modules increases due to busbar and weld degradation. Module-level measurements indicate that interconnects can contribute approximately 20% more heat than standalone cells at high current [49]. Module-level assessment therefore captures resistive heat sources that cell-level testing alone would miss.

Traditional characterization relies on full charge–discharge cycles at low C-rates, requiring approximately 1400 min (24 h) per cell; resistance-based approaches reduce this to approximately 2 min per cell [10]. Charge-interrupt (CI) resistance correlates well with SOH without requiring precise SOC information [48]. The electrode polarization coefficient ($\alpha$) has been proposed as a reliability indicator, with cells exhibiting high initial polarization reported to degrade more rapidly and develop interfacial instability [50]. As discussed in Section 2, plating-dominated aging reduces thermal thresholds significantly (Table 2). Incremental Capacity Analysis (ICA) detects LLI, an indirect indicator of lithium plating, and can therefore identify chemically unstable cells that may otherwise pass capacity-based screening. However, ICA accuracy depends on collecting voltage-capacity data at relatively low C-rates (for example, C/6 as a practical compromise for time resolution), which increases test duration compared with simple capacity-based screening [51]. ICA correlates strongly with SOH for LFP cells but is less robust for NMC chemistries due to overlapping voltage features; in such cases, impedance-based methods are often favored in practice [52].

However, many rapid methods compress resistance or polarization into single scalar values that remain insensitive to the physical source of resistance growth [11,48]. DC measurements can hide the true cause of degradation by collapsing different physical processes into a single number [53]. Cells with plating-dominated versus SEI-growth-dominated aging can exhibit similar DC resistance yet vastly different thermal margins (Section 2). This limitation motivates impedance-resolved characterization. Figure 5 maps the trade-off between screening throughput and thermal-hazard detection capability across representative methods in this section.

3.2. Electrochemical Impedance Spectroscopy (EIS) Screening

DC-based methods provide a single measure of lumped resistance. Electrochemical Impedance Spectroscopy (EIS) resolves internal resistive and kinetic processes across the frequency domain, separating ohmic, interfacial, and diffusion contributions. EIS can act as a second-stage screening tool after rapid DC checks when higher mechanism sensitivity is required. Interfacial and diffusion impedances contribute to irreversible heat generation [54] and capture degradation features that DC resistance alone obscures [55,56]. Growth of the SEI layer during first-life operation leads to gradual impedance increase, raising the likelihood of localized heat accumulation. Separating bulk resistance ($R_b$) from SEI resistance ($R_{\mathrm{SEI}}$) provides a clearer and more physically meaningful indicator of interface stability than DC resistance alone [55]. Low-frequency impedance behavior, including Warburg impedance, has been correlated with capacity loss and indicates transport limitation [55,56]. Such impedance heterogeneity can amplify localized heat generation under load [57]. EIS was long considered impractical for large-scale screening due to long measurement times, specialized equipment, and interpretative complexity.

Recent advances have made EIS more practical for large-scale screening. Feature-attribution methods identify the frequency regions most strongly linked to battery health, reducing measurement time to under 100 s while maintaining high predictive accuracy [58]. Neural-network-assisted EIS enables rapid SOH diagnosis of retired batteries without historical usage data, completing per-cell assessment in under eight minutes [59]. Comparative studies indicate that although charge-interrupt techniques are faster, EIS is more sensitive to subtle aging mechanisms, making it preferable when reliability and thermal safety are prioritized [60].

EIS features are sensitive to state of charge and temperature, and contact or fixture impedances can distort high-frequency responses. Controlled conditions and standardized protocols are therefore required for consistent large-scale grading [53,61]. Impedance features may not capture hidden physical defects or structural precursors to internal short circuits, underscoring the need for complementary structural diagnostics that provide direct defect visibility.

3.3. Thermal and Structural Characterization

DC and impedance-based screening improve mechanism sensitivity, but hidden mechanical damage and gas-related defects can remain electrically subtle yet strongly elevate internal short-circuit risk. Thermal and structural characterization targets physical precursors to failure, including deformation, delamination, and gas pockets, that can remain invisible to electrical and impedance-based methods. Internal deformations such as jellyroll collapse and electrode delamination are well-established precursors to internal short circuits [62,63].

A CT-derived “CT score” has been proposed for grading structural integrity, with findings indicating that cells scoring below 0.55 exhibit severe internal damage and should be removed from further use [62]. While conventional CT remains slow and resource intensive, recent rapid-XCT developments have reduced scan times from hours to minutes per cell, with production-scale targets of less than $10\mathrm{s}$ per cell [64,65]. It remains uncertain whether these throughput improvements can be reliably extended to the larger form factors and heterogeneous aging states characteristic of second-life cells.

Quantitative ultrasound spectroscopy (QUS) offers a faster, non-invasive alternative that has been validated on field-recovered second-life batteries, showing sensitivity to delamination and gas pockets that electrical screening can miss. QUS reduces per-cell data acquisition from hours to seconds, supporting high-throughput sorting when integrated into practical handling workflows [63]. These methods provide direct visibility of physical defects that electrical and impedance-based tests cannot access, but they remain challenging to deploy at industrial scale due to high equipment cost, long acquisition times, and the need for specialized interpretation. Data-driven methods offer a complementary pathway by detecting rare fault signatures in operational data without requiring direct physical inspection of every cell.

3.4. Data-Driven Fault Prediction

Data-driven methods address heterogeneity and unknown usage history in retired cells by detecting safety-critical fault signatures that are difficult to capture with fixed thresholds alone. In screening contexts, they primarily support offline, high-throughput sorting and chemistry verification before second-life deployment, flagging high-risk outliers for deeper diagnostic checks [66,67].

Truly hazardous cells represent rare anomalies in retired stocks, leading to strongly imbalanced datasets that can bias learning toward healthy cells. Combining the Synthetic Minority Oversampling Technique (SMOTE) with Transformer-based networks reduces this bias and improves the identification of high-risk batteries, working best as an offline layer that prioritizes suspicious cells for deeper checks [66].

Fast pulse testing combined with machine learning accelerates large-scale sorting. Unlike linear regression-based approaches, pulse-test feature learning can capture the nonlinear aging characteristics embedded in transient responses. A Random Forest-based framework achieves high diagnostic accuracy (SOH error 1.79%) even for highly heterogeneous batches, reducing data acquisition time from approximately 150 min to 125 s per cell [67].

The correct identification of cell chemistry is a prerequisite for safe second-life deployment, particularly because mixed-chemistry batches can occur in battery waste streams. Machine learning methods based on partial voltage-curve features distinguish thermally stable lithium iron phosphate (LFP) cells from more reactive nickel manganese cobalt (NMC) cells, helping prevent hazardous chemistry mixing during pack assembly [68].

Data-driven methods enhance screening throughput and early fault visibility under uncertainty, but deployment is constrained by data representativeness, labeling quality, and domain shift across heterogeneous retired stocks.

Diagnostic applicability depends on cell format and scale. DC resistance and pulse methods scale readily across geometries. EIS becomes increasingly difficult with larger cell formats [4,58], and these challenges compound at the module and pack levels due to equipment complexity and limited impedance data for large formats. ICA accuracy depends on voltage resolution, which is harder to achieve in large-format cells with low internal resistance. Structural methods face geometry-dependent constraints: QUS requires specialized analysis for cylindrical cells [63], while the CT scan time, computational burden, and cost can increase significantly with cell sizes [62].

Screening outputs, including resistance distributions, capacity spreads, structural defect flags, and chemistry classifications, define acceptable mismatch and assembly constraints for second-life packs. Electrical tests provide high-throughput filtering, impedance spectroscopy enables mechanism discrimination, structural methods detect physical damage precursors, and data-driven approaches enable rapid sorting when failure examples are scarce. Table 3 links diagnostic indicators to thermal-safety relevance and deployment constraints. These screening outputs translate into the pack-level design requirements discussed in Section 4.

Table 3. Screening method classes, representative safety indicators, and practical limitations for second-life cells.

Method	Safety Indicator	Thermal Relevance	Limitations
Electrical (DC)	Resistance, pulse response	Resistive heating, plating detection; ∼2 min/cell [10]	Misses structural defects
ICA	LLI peaks, phase transitions	Plating detection; ∼hours/cell [51]	Chemistry-dependent; slow
EIS	$R_{\mathrm{SEI}}$, charge-transfer, Warburg	Mechanism-specific risk separation; ∼2–8 min/cell [58,59]	Higher cost; SOC/temperature sensitive
Structural (CT/QUS)	CT score, ultrasound attenuation	Direct defect visibility; seconds (QUS) to hours (CT) [62,63]	High equipment cost, hard to automate
Data-Driven	Anomaly score, chemistry classifier	Rare fault detection, chemistry verification; ∼125 s/cell [67]	Needs labeled data; domain shift

Table 3. Screening method classes, representative safety indicators, and practical limitations for second-life cells.

Method	Safety Indicator	Thermal Relevance	Limitations
Electrical (DC)	Resistance, pulse response	Resistive heating, plating detection; ∼2 min/cell [10]	Misses structural defects
ICA	LLI peaks, phase transitions	Plating detection; ∼hours/cell [51]	Chemistry-dependent; slow
EIS	$R_{\mathrm{SEI}}$, charge-transfer, Warburg	Mechanism-specific risk separation; ∼2–8 min/cell [58,59]	Higher cost; SOC/temperature sensitive
Structural (CT/QUS)	CT score, ultrasound attenuation	Direct defect visibility; seconds (QUS) to hours (CT) [62,63]	High equipment cost, hard to automate
Data-Driven	Anomaly score, chemistry classifier	Rare fault detection, chemistry verification; ∼125 s/cell [67]	Needs labeled data; domain shift

4. Pack-Level Design for Safe Second-Life Operation

Screening establishes cell-level health and safety indicators; this section reviews how these constraints translate into four pack-level design domains: cell matching, electrical topology, mechanical propagation mitigation, and thermal management. Each domain interacts with the others—topology choices shape fault escalation pathways, which in turn determine the requirements for mechanical barriers and thermal control.

4.1. Compatibility and Cell Matching for Pack Assembly

Two mismatch dimensions constrain second-life pack assembly: capacity dispersion ($\Delta C$) and resistance dispersion ($\Delta R$). Both metrics are quantified during screening (Section 3); this section reviews how they translate into topology-specific assembly constraints. Figure 6 illustrates these mechanisms: series voltage divergence from capacity mismatch (Figure 6a), parallel current redistribution from impedance dispersion (Figure 6b), and the resulting electro-thermal feedback loop (Figure 6c).

In series strings, the lowest-capacity cell limits usable energy through early voltage cut-off, with unmitigated mismatch reducing discharge capacity by 20–25% or more [69]. Capacity matching is therefore essential for series utilization. Bilevel equalizers can partially recover capacity when $\Delta C$ bounds cannot be met through sorting alone, though at the cost of added hardware complexity. In repurposed packs, $\Delta C$ matching remains the preferred first-line strategy, with equalization treated as mitigation when residual mismatch is unavoidable [69,70]. In parallel blocks, capacity variance is partially buffered: a 9% difference produces only 4.3% current variation [71], though time-resolved measurements show capacity becomes the dominant load-imbalance driver mid-discharge [72]. Parallel-then-Series topologies deliver higher initial capacities than Series-then-Parallel when cells equalize total capacity across parallel modules [73].

Resistance mismatch drives current imbalance and localized Joule heating in parallel groupings. A 20% $\Delta R$ can reduce cycle life by approximately 40% [74], and 30% mismatch can increase peak cell currents by up to 60%, with branch currents exceeding rated limits by 53%, elevating lithium-plating risk [75,76]. This sensitivity is chemistry dependent: NMC cells exhibit current deviations up to 80% under pulsed loads, compared with approximately 40% for LFP [77], requiring stricter impedance tolerances for NMC to manage thermal gradient sensitivity. Chemistry-dependent $\Delta R$ thresholds across cell formats and aging states lack standardization, complicating universal screening criteria for parallel-connected second-life assemblies [74,77]. In highly parallel configurations, interconnect resistance can dominate over internal cell resistance in driving current redistribution [71,78], making strict resistance matching indispensable.

Table 4. Mismatch impacts and matching priorities for second-life pack assembly.

Mismatch Type	Series Impact	Parallel Impact	Matching Priority
Capacity ($\Delta C$)	Weakest-cell constraint; 20–25% energy reduction if unmitigated [69]	Partially buffered; 9% $\Delta C$ → 4.3% current variation [71]	$\Delta C$ for series utilization; capacity-balanced parallel grouping
Resistance ($\Delta R$)	Increased string impedance; distorted voltage monitoring	Current maldistribution; 20% $\Delta R$ → ∼40% life reduction [74]; hotspot and plating risk. Chemistry-dependent: NMC up to 80% deviation, LFP ∼40% [77]	Strict $\Delta R$ for parallel safety; tighter than new-cell standards; stricter for NMC
Interconnect contribution	Aging contacts distort voltage readings	Dominates current redistribution in high-parallel assemblies	Include interconnect resistance in total mismatch budget

summarizes these mismatch effects across series and parallel topologies.

Beyond enforcing $\Delta C$ and $\Delta R$ bounds, second-life pack assembly increasingly relies on multi-parameter grouping. Screening outputs from Section 3—electrochemical features, structural integrity scores, and remaining useful life estimates—serve as inputs to grouping algorithms that minimize dispersion across multiple dimensions [79]. Support Vector Machine (SVM)-based grouping using high-rate IC-curve features reduces current inconsistency after regrouping [80], “Equal-number” Support Vector Clustering (SVC) with RUL as a grouping feature reduces within-cluster RUL dispersion [81], and two-stage frameworks combining Density-Based Spatial Clustering of Applications with Noise (DBSCAN) with Principal Component Analysis-Self-Organizing Map (PCA-SOM) improve grouping outcomes relative to single-stage schemes [79]. Structural indicators from non-destructive CT imaging further support assembly decisions; a Structural Similarity Index Measure (SSIM) above 0.65 supports reuse, whereas scores below 0.55 indicate scrapping [62].

Cell matching and grouping establish mismatch bounds for pack assembly. Topology then determines whether residual mismatch manifests primarily as voltage divergence (series) or current redistribution (parallel), shaping both utilization and fault escalation pathways.

4.2. Electrical Topology and Thermal Safety

The mismatch effects described in Section 4.1 manifest differently depending on electrical topology, which also determines fault escalation severity.

4.2.1. Series and Parallel Connections

In series strings, current continuity enforces identical charge throughput, so heterogeneity manifests as voltage and SOC divergence rather than branch-current imbalance [69,82]. This divergence accumulates with cycling due to coulombic efficiency differences and initial SOC offsets [82,83], with aged modules reporting end-of-discharge voltage spreads approaching 0.9 V that force the weakest cell into disproportionately deep discharge [84]. These conditions promote localized heat generation and accelerate degradation. The BMS cut-off effectiveness depends on accurate voltage prediction. Widely used equivalent-circuit models can produce errors exceeding 0.76 V when BMS constraints such as minimum SOC limits are not accounted for in pack-level representations [85]. These models further struggle to capture nonlinear aged-cell behavior, as degradation trends become polynomial near the aging knee and first-life parameterizations lose validity for second-life applications [86]. Series voltage sensitivity can, however, support fault detection, as incipient failure may appear as a distinct terminal-voltage drop [70].

Parallel topologies promote SOC equalization through voltage clamping, but in heterogeneous assemblies the dominant risk shifts toward current-sharing imbalance and thermally coupled feedback [82,83]. Layout choices shape branch-current distribution: common-end designs produce charge disparities up to 21.5%, compared with 7.8% in opposite-end configurations [87]. High interconnect resistance ratios can produce temperature spreads up to 47.7 °C within a module [88], triggering a destabilizing feedback loop in which warmer cells draw more current and heat further [57,89]. Concealed re-balancing currents up to 3.5 A circulate after load removal, sustaining ohmic heating that standard BMS strategies may not detect [87]. Parallel architectures also enable hazardous failure escalation: during internal short circuits, healthy cells discharge into the faulty cell, and transferring only 2.56 Ah (about 4.6% of the cell’s total capacity) can lower the thermal runaway onset temperature from 221.7 °C to below 170 °C [90]. Parallel-connected modules show the most severe failure characteristics among common topologies, including peak temperatures approaching 720 °C [91].

4.2.2. Combined Architectures and Propagation Pathways

Combining series and parallel connections into Series-Parallel (SP, series-first) or Parallel-Series (PS, parallel-first) architectures introduces a trade-off between capacity utilization and electro-thermal safety. Figure 7 contrasts SP and PS architectures: SP enforces uniform series current (Figure 7a) with heat-transfer-governed fault propagation (Figure 7b), while PS divides current by impedance (Figure 7c) and enables parallel neighbors to discharge into failing cells (Figure 7d).

PS designs are often favored in repurposing because capacity averaging partially mitigates variability and improves cell utilization [73,92], achieving about 10% higher energy throughput over 400 cycles [84]. However, PS topologies are more sensitive to parameter deviations and develop larger intra-module current imbalances than SP designs [82,93]. In aged modules, 2P2S yields the greatest mass loss (55.7 g) and highest average heat transfer (20.16 kJ) during thermal runaway [91], and current inhomogeneity grows with the interconnect-to-cell impedance ratio [57]. Whether the increased internal resistance of retired cells attenuates fault-energy transfer or whether aging-induced dispersion instead intensifies localized heating remains unresolved [88,90].

Table 5. SP and PS topology comparison for second-life modules.

Performance Metric	Series-Parallel (SP)	Parallel-Series (PS)	Second-Life Takeaway
Energy Utilization	Weakest-cell limited	Capacity averaging; ∼10% higher throughput [84]	PS maximizes utilization; needs robust screening
Current Distribution	Equal current enforced	Impedance-dependent splitting; prone to maldistribution	SP stable; PS vulnerable to dispersion
Voltage Consistency	Divergence up to 0.9 V; deep-discharge ris [84]	Voltage clamping; self-balancing	SP divergence aids detection; PS masks faults
TR Initiation and Propagation	Heat-transfer governed; limited energy transfer	Neighbor-to-fault energy transfer; TR onset lowered >50 °C [90]; peak ∼720 °C [91]	PS more severe escalation
Interconnect Sensitivity	Aging contacts distort voltage monitoring	Higher impedance ratios worsen $\Delta T$ (up to 47.7 °C) [88]	PS depends on assembly quality
Monitoring and BMS	Cell-level voltage visibility	Reduced observability; faulty cells masked	SP preferable for diagnostics

compares SP and PS topologies across six performance dimensions.

Topology determines how mismatch manifests as current and thermal redistribution; mechanical design determines whether single-cell thermal runaway remains isolated or cascades through the module.

4.3. Mechanical Design and Propagation Mitigation

Passive mechanical strategies, including cell spacing, thermal barriers, and venting management, delay propagation after a cell enters thermal runaway, while active thermal management (Section 4.4) targets prevention during normal operation. These strategies must account for aged-cell characteristics (Section 2.4): swelling reduces effective spacing, altered gas generation changes venting dynamics, and structural weakening narrows the margin between pressure buildup and failure.

4.3.1. Cell Spacing and Thermal Barriers

Spacing alone is generally insufficient in enclosed modules. For cylindrical 18,650 cells, horizontal gaps exceeding 2 mm in open air typically stop propagation, whereas vertical arrangements require spacing greater than 8 mm regardless of enclosure. In closed packs, horizontal spacing must exceed 4 mm [94]. Even with increased spacing, convection and radiation dominate in tight enclosures: 0 mm to 2 mm extends the propagation window from 32 s to 165 s [95] but can fail to prevent full-module failure [95,96]. Under high-density packing, a failing cell transfers 35.3% to 72% of its total heat to neighbors, and high SOC drives peak temperature more strongly than geometry can offset [97]. Widening the gaps from 0.3 mm to 0.9 mm delays propagation but is insufficient without additional mitigation [98].

Interstitial thermal barriers address these limitations, with effectiveness scaling predictably with thickness. Nanofiber aerogels show this clearly: 2.0 mm stops propagation, 0.5 mm doubles time to failure, and 1.0 mm extends it sixfold [99]. A 2 mm flame-retardant polypropylene layer prevented propagation where air gaps were ineffective [95], and a 3 mm epoxy composite maintained neighboring-cell temperatures below 100 °C [100]. Thinner insulation (e.g., 1 mm aerogel) offers measurable delay but can remain insufficient in large-format pouch configurations [101].

Phase change materials (PCMs) buffer transient heat release rather than providing sustained rejection. Submerging cells in PCM delayed runaway onset from 160 s to 330 s, yet subsequent ignition led to rapid failure [102]. Hybrid approaches therefore combine PCMs with structural constituents: graphite–wax composites reduced neighboring-cell temperatures to 109 °C versus 189 °C under air cooling [103], and trifunctional barriers combining nano-ceramics, PCM, and mica tolerated 4 MPa with <20% deformation [104]. Additional concepts use vaporization or radiant shielding, including hydrogel layers and zirconia-aerogel composites [105,106]. Vent jets (high-velocity hot gas discharge) and ejecta (expelled particles and fragments) can bypass conduction- and radiation-dominated isolation, requiring barrier selection to consider vent-path management.

4.3.2. Venting Pathways and Ejecta Management

For cylindrical cells, sidewall rupture is more thermally severe than top-cap rupture, with average maximum surface temperatures of 919 °C versus 653 °C, motivating directional containment [107]. Aging alters venting behavior: cells at 70% SOH produce the largest total gas volume, shifting toward more carbon monoxide and hydrocarbons, and exhibit a two-step violent jet fire in which the second jet has a larger diffusion area but lower intensity than the first [21]. These characteristics can exceed the capacity of pressure-relief valves and exhaust paths sized for new-cell behavior [21,95]. If vent gases are not routed through dedicated channels, jet-driven convection can become the main driver of propagation even in modules with air gaps [95].

Aged-cell venting characteristics—particularly multi-stage ejection and elevated gas production near intermediate SOH—are not yet incorporated into second-life design guidelines or safety standards. Propagation risk is also SOC-dependent: thermal runaway propagation is highest within 40–60% SOC [108], indicating that second-life deployment may require operational constraints avoiding this window for packs with elevated heterogeneity.

4.4. Thermal Management Systems (TMS) for Aged Batteries

Aged cells generate more heat and distribute it less uniformly than new cells due to irreversible resistance growth and impedance dispersion (Section 2.2). This uneven heat generation reinforces the electro-thermal feedback described in Section 4.1, and thermal gradients alone can increase degradation by 5.2% relative to uniform-temperature operation [109]. Second-life TMS design must therefore prioritize temperature uniformity ($\Delta T$) over average temperature reduction [74,109].

Forced-air cooling remains attractive for its simplicity but may be insufficient to suppress gradients in dense, heterogeneous modules. Air cooling is most defensible for low-power stationary duty when mismatch has been tightly controlled through screening [110]. Liquid-based approaches offer substantially higher heat-rejection capability: single-phase immersion cooling can increase heat transfer coefficients by up to 1000 times relative to passive air, rising to 10,000 times for two-phase systems, and can maintain cell temperatures near 35 °C even under 10C cycling [111]. Counterflow liquid-cooling arrangements reduce thermal gradients further, with $\Delta T_{max} = 2.70 °\mathrm{C}$ at discharge rates up to 2C [112]. In TMS applications, MXene-doped PCM limits maximum temperature to 57.03 °C at 3C-rate (current normalized to capacity; 1C = full discharge in 1 h) discharge, while hybrid PCM-based approaches improve uniformity compared with air cooling alone [113,114]. Higher-power duty or larger health variation typically motivates active liquid or hybrid systems to constrain $\Delta T$ and suppress hotspot-driven degradation [111,112]. A remaining gap is life-consistent thermal sizing: designs derived from new-cell heat generation can under-predict demand as internal resistance increases over second life, and whether oversizing at the design stage or adaptive control is the more robust pathway remains unclear [74,115,116].

Table 6. Integrated pack-level design recommendations for SLB safety.

Design Domain	Recommended Strategy	Safety Rationale	Primary Engineering Objective
Cell Matching	Multi-parameter matching; tight $\Delta R$ (parallel), $\Delta C$ (series)	Resistance dispersion drives hotspots; capacity dispersion limits usable energy	Define mismatch bounds; prioritize $\Delta R$ for parallel safety
Electrical Topology	PS with fault protection for utilization; SP for observability	PS enables fault energy transfer, reducing TR onset >50 °C; SP maintains voltage visibility	Balance utilization against fault-escalation risk
Propagation Mitigation	2–4 mm spacing + barriers (≥2 mm) + directional vent routing	Spacing alone fails enclosed; barriers block conduction/radiation; venting prevents convection bypass	Contain propagation across all transport modes
Thermal Management	Liquid/hybrid for high-power or large SOH variation; air/passive for low-rate (<0.5C-rate)	Non-uniform $I^{2}R$ heating in mismatched cells; $\Delta T$ suppression more critical than average T	Maintain temperature uniformity under evolving SOH

summarizes the recommended strategies across these four design domains.

These four design layers form an integrated safety system in which weaknesses at one stage propagate forward. Pack-level design choices establish the physical risk envelope and observability constraints that motivate the state-estimation frameworks discussed in Section 5.

5. Thermally Coupled State Estimation

The physical safeguards established in Section 4 require real-time monitoring to detect incipient faults and track battery health. This challenge is acute for second-life packs: heterogeneous aging limits observability, low-cost BMS hardware constrains computation, and conventional state estimators assume stable parameters and uniform cells—assumptions that often fail for retired packs. We address these requirements across five domains. Section 5.1 discusses why conventional estimation fails under heterogeneity and unknown aging histories. Section 5.2 reviews adaptive electrical state estimation that tracks time-varying parameters. Section 5.3 reframes prognostics around time-to-unsafe boundaries rather than capacity retention. Section 5.4 discusses fault detection under topology-driven observability constraints, particularly parallel masking. Section 5.5 surveys data-driven strategies for rare-event monitoring under domain shift and sensor uncertainty.

5.1. Second-Life Estimation Challenges

SLB deployment confronts three operational constraints: unknown first-life degradation histories that limit parameter initialization, pronounced cell-to-cell heterogeneity that invalidates uniform-cell assumptions, and low-cost BMS hardware that constrains real-time computation. The coefficient of variation in capacity across second-life packs is approximately four times higher (11.3%) than in new counterparts (3.0%) [86,117]. Cells with similar SOH can exhibit distinct degradation pathways—lithium plating, SEI growth, or gas generation—that produce vastly different thermal margins [16,118]. Although extensive characterization could reduce this uncertainty, conventional testing requires days per cell, making mass screening impractical [10].

Most state-estimation algorithms were developed for new cells with slowly varying parameters, which creates three fundamental limitations for aged batteries. First, Coulomb counting accumulates initial offsets and becomes increasingly biased as coulombic efficiency changes with aging, leading to SOC drift [119,120]. Second, Equivalent Circuit Models with fixed parameters cannot track nonlinear resistance evolution and polarization changes, producing voltage prediction errors that degrade estimator consistency [121,122]. Third, separating capacity fade from resistance growth can yield inaccurate SOC estimates by neglecting coupled electro-thermal interactions in degraded cells [123]. At the pack level, these errors are amplified by architecture-dependent aggregation. Approximating parallel-connected groups as single voltage nodes can introduce errors of up to 5%, masking cell-level anomalies that become critical during fault progression [85].

These estimator limitations become safety critical because thermal stability depends on degradation mode rather than capacity alone—a divergence that Section 5.3 quantifies in detail [11,31,124]. Parallel configurations are particularly vulnerable: impedance dispersion drives current maldistribution and localized heating, while fault escalation pathways may not be captured by single-cell assumptions [87,90]. Despite these risks, many standards and BMS protocols still treat cells as uniform units, limiting adaptation of thresholds to the governing aging mechanism [125]. Data-driven estimators face an additional constraint: performance degrades significantly once cells enter the nonlinear aging knee, where trajectories depart from training data [126]. Rapid, non-invasive methods that can infer degradation-mode fingerprints without historical data remain needed [127].

5.2. SOC Estimation for Aged Cells

Adaptive co-estimation frameworks address parameter drift by separating fast electrical dynamics from slower aging evolution [119,121]. Adaptive ECM identification using Recursive Least Squares (RLS) enables online tracking of ohmic resistance and polarization elements, and when coupled with nonlinear observers such as Unscented Kalman Filters, updates SOC relative to the cell’s actual impedance state rather than assuming new-cell behavior [121,123]. Enhanced Coulomb counting with periodic calibration of coulombic efficiency provides a lower-complexity alternative [120], and data-driven approaches including cluster-based learning can represent nonlinear behaviors that fixed-structure models struggle to capture [128]. However, safety-critical deployment remains constrained by training-data requirements and generalization across heterogeneous aging pathways [129]. Hybrid physics-informed strategies that retain interpretability while adapting key parameters online are therefore preferable for second-life operation [119].

Computational feasibility constrains pack-level estimation for second-life modules with low-cost BMS hardware. For series-connected packs, simplified aggregation strategies such as Cell Mean Models combined with adaptive filtering provide a practical compromise between fidelity and embedded feasibility [82,130].

Adaptive frameworks improve SOC tracking but do not account for the thermal consequences of resistance evolution. Thermally coupled estimation addresses this gap by integrating temperature dependence into electrical state estimation [121]. Internal core temperature can exceed surface temperature by approximately 4 °C during discharge, and this gradient widens with aging due to degraded thermal conductivity [131]. Temperature-assisted estimators that treat surface temperature as an auxiliary input can reduce electrical-state error under dynamic conditions [132]. Thermally coupled co-estimation enables the real-time inference of localized Joule heating ($I^{2}R$) through jointly estimated resistance parameters, allowing the BMS to implement adaptive current limits—derating—that prevent hotspot formation in high-resistance cells [122].

Most thermally coupled frameworks remain validated at the single-cell level. Scaling to large heterogeneous second-life packs is constrained by embedded computation, sensing availability, and the need to run multiple adaptive filters in parallel [7,133]. Full thermal estimation on every cell is computationally infeasible for low-cost BMS hardware, motivating hierarchical implementations in which thermally informed estimation and derating are prioritized for higher-risk strings or modules, while lightweight pack-level observers maintain real-time feasibility.

5.3. From SOH/RUL to Time-to-Unsafe

Capacity-based SOH is an incomplete proxy for thermal safety. Capacity fade reduces the total chemical energy available during a failure event. However, resistance growth can occur in parallel and increase Joule heating ($I^{2}R$), thereby elevating thermal stress even at reduced capacity [122,124]. As a result, two cells with similar SOH (e.g., 80%) can have substantially different safety margins when their dominant degradation modes differ [11]. In particular, lithium-plating-dominated aging degrades thermal stability across multiple stages of failure progression, as quantified in Section 2.1 [6,30,31]. These safety-relevant distinctions cannot be resolved using capacity-based metrics alone. Mechanical abuse tests reinforce this point: aged cells may appear mechanically stiffer yet exhibit larger voltage drops and faster energy release once an internal short initiates [134]. Safety evolution can also be non-monotonic across SOH. As discussed in Section 4.3, combustible gas generation peaks near 70% SOH in NCA cells—aligning with typical second-life entry points—and aged cells exhibit a two-step violent jet-fire behavior that produces a larger diffusion area than new-cell failure modes [21]. These observations motivate prognostic targets that extend beyond end-of-capacity to a time-to-unsafe boundary.

For second-life deployment, RUL estimation is safety relevant only if it can forecast a transition to an unsafe state, not solely an end-of-capacity threshold. A practical formulation is time-to-unsafe, where the unsafe boundary is associated with the onset of the aging knee—a nonlinear acceleration in degradation reflecting processes such as lithium plating saturation or electrode pore clogging [135]. One approach fits a linear trend using the previous 500 cycles and flags the knee when measured capacity deviates by more than 3% or DCIR by more than 6% from the prediction [86]. Non-invasive diagnostics such as Differential Thermal Voltammetry (DTV) and ICA can track internal degradation states including Loss of Lithium Inventory [136,137], and predictive models can estimate knee-point timing using early-cycle features such as capacity-variance evolution [138]. Translating these outputs into real-time BMS safety logic remains underdeveloped; dynamic derating or operational constraint policies that respond to evolving probabilities of plating or hazardous gas generation are not yet integrated into standard BMS architectures [139].

5.4. Early Fault Detection in Packs

Observability limitations constrain early thermal-fault detection in second-life packs, particularly in parallel architectures where busbar averaging masks cell-level anomalies. Conventional BMS implementations often treat each parallel block as a single voltage node, concealing the severe current imbalance arising from parameter mismatch [140]. The current redistribution and charge disparities quantified in Section 4.1 and Section 4.2 [75,76,87] can exceed typical BMS detection thresholds based on block-level voltage alone.

Figure 8 illustrates these masking effects. During normal operation (Figure 8a), branch currents redistribute according to impedance mismatch while terminal voltage remains clamped. During an internal fault (Figure 8b), parallel neighbors supply current to the failing cell, concentrating heat while pack-level voltage remains relatively stable. From the BMS perspective (Figure 8c), voltage averaging obscures cell-level anomalies: an apparently normal pack voltage can mask a growing hazard.

Fault escalation can also be topology enabled in ways that are difficult to infer from terminal measurements. During internal short-circuit events, transferred electricity from parallel neighbors can concentrate heating near tabs, lowering thermal-runaway onset temperature (Section 4.2.1) [90]. Interconnect–resistance ratios further compound the detection challenge: increasing these ratios raises current inhomogeneity logarithmically and reduces usable capacity by up to 42% [88].

Surface temperature sensing exhibits latency, and aging widens internal–surface thermal gradients (Section 5.2), motivating alternative precursors. Mechanical expansion monitoring provides earlier warning than temperature during low-rate external heating, and combining thermal and mechanical signals further improves early-warning performance [141]. The Connectivity-Based Outlier Factor (COF) method has detected high-resistance internal short circuits (100 $\mathrm{\Omega }$) within 207 s by amplifying subtle voltage deviations [142]. Multi-physics fusion—combining electrical, thermal, and mechanical indicators—can improve detection under these observability constraints, but practical BMS implementations lack integrated frameworks that remain robust to sensor noise and distinguish benign aging from incipient thermal runaway [141,143].

5.5. Data-Driven Safety Monitoring Under Uncertainty

Thermal runaway and related safety faults represent rare but high-consequence events, creating a class-imbalance problem in which failure samples are scarce relative to normal-operation data [66,144]. Reconstruction-error methods address this by learning healthy baselines and flagging departures without requiring dense failure labels [142,144]. A Bi-LSTM with attention achieved warning accuracy 14.3 percentage points higher than K-means clustering for thermal runaway events [144], autoencoder-based monitoring achieved zero false alarms with an average detection delay of 108.2 s for oil-leakage faults [145], and optimized Random Forest classifiers achieved strong anomaly-detection accuracy with zero false negatives under embedded resource constraints [146].

Second-life packs exhibit evolving baselines as degradation progresses, making conservative threshold selection essential to balance missed detections against false alarms [6,142,144]. Sensor corruption further compounds the challenge: under noise-corrupted thermal imaging, autoencoder-based denoising improved fault-recognition accuracy by up to 32% while achieving segmentation performance of mean Average Precision (mAP) >97% [143].

Resilience to domain shift is critical given diverse first-life histories. Models trained on one population may degrade when deployed to another because normal-behavior baselines vary with chemistry, aging mode, and operating context. Maximum Mean Discrepancy (MMD)-based adaptation reduced estimation error by more than 47% relative to non-adaptive methods [136], though generalization can deteriorate near aging-knee transitions [126]. Handling spatiotemporal dependency is equally essential in parallel assemblies where voltage clamping conceals cell-level anomalies [87,140]. A Transformer-based multimodal framework reported a 28.3% reduction in missed pre-runaway cases alongside 93.5% spatial degradation localization accuracy, with a 37% reduction in sensing power consumption via reinforcement-learning-based adaptive sensing [147]. Such deep architectures can exceed low-cost BMS resources without compression or offloading [129].

Balancing detection fidelity against computation cost motivates hierarchical monitoring: lightweight edge models provide continuous anomaly scoring [146], while more intensive diagnostics are triggered selectively or offloaded to cloud-assisted digital-twin infrastructures [148].

Low-overhead implementations have reported approximately 1% CPU and 49.3% memory usage for optimized ensemble inference on embedded hardware [146]. Physics-based frameworks have also demonstrated rapid processing, completing per-test analysis in 2.9 s on a Raspberry Pi [149]. In practice, these constraints compound: parameter drift introduces bias, resistance growth creates gradients surface sensors miss, parallel topology masks divergence, aging-knee transitions depart from training baselines, and all estimates must run on low-cost hardware. Their combination requires integrated solutions that coordinate adaptive estimation, multi-physics sensing, and hierarchical architectures.

Table 7. Monitoring strategies mapped to second-life operational constraints.

Estimation Challenge	Method Class	BMS-Actionable Output
Parameter drift under unknown aging histories	Adaptive co-estimation; thermally coupled observers; temperature-assisted estimation	Robust SOC tracking; resistance-informed derating; adaptive current limits
Safety-oriented prognostics beyond capacity fade	Time-to-unsafe estimation; DTV/ICA diagnostics; knee-point detection	Dynamic operational constraints; retirement triggers; degradation-mode-aware thresholds
Parallel masking and interconnect-driven imbalance	Multi-physics fusion; anomaly scoring; virtual sensing; mechanical expansion monitoring	Early fault detection; at-risk group localization; multi-signal early warning
Rare-event detection under domain shift and compute constraints	Unsupervised anomaly detection; domain adaptation; hierarchical edge–cloud monitoring	Continuous low-overhead scoring; selective high-fidelity diagnostics; noise-tolerant warning

maps these strategies to operational constraints and BMS-actionable outputs.

Combined with mechanism-aware screening and topology-constrained pack design, these monitoring capabilities complete the integrated safety framework for second-life deployment.

6. Discussion, Challenges and Future Directions

6.1. Key Findings and Knowledge Gaps

Section 1 framed second-life thermal safety around four interdependent areas and highlighted that thermal risk follows the degradation pathway rather than capacity fade. Evidence across Section 2, Section 3, Section 4 and Section 5 supports this and shows the tight coupling among these areas: degradation mechanisms shape screening outputs, which in turn constrain pack-level design and monitoring effectiveness. Plating-dominated aging substantially reduces thermal runaway trigger temperatures, even at SOH levels comparable to those of SEI-growth-dominated cells that retain near-new-cell stability. Capacity-based screening cannot resolve this distinction, creating the possibility that electrically functional cells enter packs with latent initiation risk. Weaknesses introduced at any stage propagate forward and reduce safety margins cumulatively as illustrated in Figure 9. In a representative scenario, a plating-prone cell cleared by capacity-only screening enters a parallel module where voltage averaging masks divergence. Under fault conditions, neighboring cells discharge into the failing cell; heat release exceeds barrier capacity. Module-level monitoring detects the event only after irreversible progression has begun. Downstream measures redistribute this risk but do not correct upstream errors.

Despite progress in understanding degradation mechanisms and integrating screening into pack-level controls, three knowledge gaps continue to limit confident second-life deployment.

Thermal characterization: Systematic thermal-safety baselines for cells at 70–80% SOH remain scarce in the published literature. Trigger temperatures, venting characteristics, gas evolution, and heat release profiles organized by degradation pathway are needed to distinguish plating-dominated from SEI-growth-dominated aging across different battery chemistries. Heat-generation evolution with SOH has not yet been mapped systematically across degradation pathways, creating the risk that thermal management sized from new-cell data is inadequate for aged populations.

Diagnostic validation: High-throughput screening approaches (capacity and resistance) do not resolve the degradation-mode differences that determine initiation sensitivity. More informative diagnostics (EIS, ICA, and partial pulse tests) show promise but remain insufficiently validated at industrial throughput across mixed chemistries and formats. Structural methods (CT and QUS) provide direct defect visibility but lack standardized acceptance thresholds, and their correlation with thermal-safety metrics such as trigger temperature has not been established. At the system level, detecting faults within parallel assemblies requires multi-signal sensing (voltage, temperature and expansion), yet current BMS implementations often lack this observability.

Translation: Standardized procedures to convert screening outputs into pack-level design rules remain limited. Validated $\Delta R$ limits for parallel modules, $\Delta C$ tolerances for series strings, and structural damage rejection thresholds have not been established. Mismatch tolerances vary by chemistry and format but lack standardized bounds. Pack models often assume constant interconnect resistance, whereas aging can increase this resistance, potentially shifting failure modes toward connector-related faults.

Collectively, these gaps translate into three engineering challenges that define the near-term research agenda, alongside the corresponding research directions summarized in Figure 10.

6.2. Challenges and Future Directions

Progress follows a dependency chain (Figure 10, bottom): experimental databases enable diagnostic validation, validated diagnostics inform design rules, and established rules support standards revision.

Balancing screening throughput against mechanism-level visibility: Mass screening still relies on capacity and resistance proxies that miss initiation risk. Resolving this requires shared thermal-abuse databases for cells at typical second-life entry points under controlled aging pathways. Priority measurements include trigger temperature and heat-release profiles across aging modes, vent gas composition at different SOH levels, and venting dynamics in moderately degraded cells. First-life history data (cycling conditions, thermal exposure, and fast-charge events) would further improve screening reliability. Pack-relevant data on current redistribution, temperature gradients, and fault-energy transfer as functions of resistance dispersion and interconnect condition are also needed.

Scaling diagnostics from laboratory to industrial deployment: Laboratory validation typically uses uniform cells under controlled conditions, whereas industrial screening faces heterogeneous stocks with unknown histories at much higher throughput. Methods proven in controlled studies must therefore be tested under conditions representative of heterogeneous retired stocks. Chemistry and format-specific thresholds are needed for resistance mismatch (parallel), capacity spread (series), and structural damage rejection. Integrating multi-signal monitoring into BMS would improve fault detectability in parallel architectures where voltage averaging masks cell-level anomalies.

Translating screening outputs into defensible design rules: Second-life qualification protocols should incorporate mechanism-aware screening and topology-specific safety assessment. Battery history tracking that captures thermal exposure and fast-charging events, not only cycle count, would improve risk classification. Safety testing protocols reflecting aged-cell venting behavior, including multi-stage gas ejection, remain underdeveloped. Without shared characterization databases organized by degradation pathway, diagnostic validation and standards development will likely remain fragmented.

7. Conclusions

This review surveyed thermal safety and management across four coupled domains for second-life lithium-ion batteries: degradation mechanisms (Section 2), cell screening (Section 3), pack configuration (Section 4), and state estimation and monitoring (Section 5), organized through the mechanism-to-deployment framework introduced in Section 1. The central finding is that thermal safety follows the degradation pathway rather than capacity fade: cells with comparable capacity can exhibit substantially different trigger temperatures depending on whether lithium plating or SEI growth dominates. This mechanism dependence shapes all downstream decisions: degradation pathway sets the screening requirements, screening outcomes constrain pack design, topology affects fault escalation, and together these determine the observability limits that monitoring must address. Three gaps—thermal characterization, diagnostic validation, and screening-to-design translation—must be addressed in sequence, as each enables the next. Coordinated effort from research, industry, and regulatory communities will be required to close these gaps and ensure safety keeps pace with second-life deployment.