Lithium-ion batteries underpin electrified transportation, grid buffering, and portable electronics, yet large-scale deployment remains constrained by the coupled requirements of high energy density, short charging time, long lifetime, low cost, and safety. Recent reviews in Advanced Materials and npj Computational Materials show that machine learning (ML) has accelerated the transition from empirical trial-and-error workflows toward data-centric battery science, while also exposing the limitations of purely empirical optimization when data coverage, task relevance, and experimental validation are inadequate^[1-3]. For electric vehicles and stationary storage, the central challenge is therefore no longer limited to discovering better materials; it is to manage each battery system safely and efficiently as it ages under heterogeneous real-world use.

Materials-level and lifecycle evidence further show why battery digital twins should be framed as materials-to-systems representations rather than as battery management system (BMS) software alone. Transport, interfacial stability, manufacturability, safety tolerance, and recycling constraints all influence which hidden states can be observed, which models remain valid, and which lifecycle decisions can be justified.

This systems-level challenge creates a natural role for digital twins. In the battery context, a digital twin should be understood as a continuously updated representation that integrates measurements, mechanistic priors, and inference to estimate internal states, predict degradation, quantify risk, and support operational decisions^[4,5]. Such a representation is valuable only if it links hidden electrochemical and thermal states to actionable outcomes: health estimation, safety warning, fast-charging supervision, maintenance, and lifecycle decisions.

Fast charging illustrates the promise and difficulty of this framework. Closed-loop learning has shown that high-dimensional charging protocols can be searched efficiently^[6]. Plating-aware pressure sensing^[7], thermal-wave diagnostics^[8], ultrasound imaging^[9], and MHz-band electromagnetic measurements^[10] further indicate that the practical limits of fast charging are controlled by latent electrochemical, thermal, and mechanical states rather than by voltage-current thresholds alone. Fast charging should therefore be treated as a state- and risk-dependent control problem, not as a fixed protocol-design problem.

The main barrier is the gap between laboratory observability and field deployment. Vehicle batteries rarely experience ideal full cycles; instead, they operate through fragmented charging, partial-discharge windows, thermal variability, cell-to-cell inconsistency, and sparse labels for true health states. A Communications Engineering study demonstrates that large electric vehicle (EV) datasets can support more realistic state-of-health estimation^[11], while a Nature Communications analysis of open-source EV data further reveals persistent ambiguity in state of health (SOH) definition, pack-level heterogeneity, and operational domain shift^[12].

Existing reviews often cover only one part of this landscape - battery informatics^[2], application-oriented battery ML^[3], physics-guided ML^[4], or probabilistic health prognostics^[5] - and therefore stop short of a unified account of how sensing, modeling, inference, and decision-making should be connected. This revised review adopts a different organizing principle. Instead of cataloging algorithms, it focuses on uncertainty-aware physics-guided digital twins as decision systems for integrated health, safety, and fast-charging management.

What is still missing is a framework that is sufficiently mechanistic to remain meaningful under distribution shift, yet sufficiently adaptive to remain useful under fragmented operating data and evolving battery conditions. This missing middle is precisely where battery digital twins become relevant. In this context, a digital twin should not be reduced to a high-fidelity simulator, a software dashboard, or a machine-learning predictor. It should instead be understood as a decision-oriented representation of the battery that is continuously updated by measurements, informed by mechanistic priors, and evaluated against uncertainty. Such a definition is especially important in electrochemical systems because degradation-mode analysis shows that multiple internal pathways can generate similar terminal voltage and capacity trajectories^[13], while spatially resolved plating simulations further illustrate why internal-state representation is needed beyond scalar terminal signals^[14].

A second unresolved issue concerns scale. Much of the literature on battery diagnostics is still framed around single cells under well-controlled laboratory protocols, whereas practical battery management must operate at the level of packs, fleets, and lifecycle decision chains. Thermal gradients, current maldistribution, module imbalance, charging-station heterogeneity, and incomplete ground truth all alter the meaning of state estimation once deployment moves from the cell cycler to the vehicle or storage system^[11-13]. Consequently, a review focused only on cell-level inference or only on algorithm families risks missing the system's problem that motivates digital twins in the first place.

A third unresolved issue concerns interpretation. A growing body of recent work argues that the central variables for battery management are not single scalar labels, but latent representations that connect what can be observed to what must be controlled. For health, this means degradation modes rather than capacity alone^[13]. For charging, it means internal plating dynamics and a risk-conditioned action set rather than a universally optimal protocol^[14,15]. For safety, it means a latent state of safety rather than post hoc threshold alarms^[16]. These developments suggest that the most useful future battery-management frameworks will be structured around hidden states, uncertainty, and control relevance instead of around isolated prediction tasks^[5].

This review therefore aims to provide a more closed analytical narrative than is typical in existing battery-ML surveys. Rather than treating health estimation, safety diagnosis, fast charging, sensing, uncertainty quantification, and deployment as parallel themes, the article argues that they are different expressions of the same multi-layered problem. The central thesis is that a deployable battery digital twin must be simultaneously physics-grounded, uncertainty-aware, and decision-oriented. The discussion is organized accordingly: Section SHARED PHYSICOCHEMICAL ORIGINS OF HEALTH, SAFETY, AND FAST CHARGING identifies the shared physicochemical basis of health loss, safety escalation, and fast-charging limitation; Section ARCHITECTURE OF UNCERTAINTY-AWARE PHYSICS-GUIDED DIGITAL TWINS defines the architecture of uncertainty-aware physics-guided digital twins; Section DIGITAL-TWIN-ENABLED HEALTH MANAGEMENT examines digital-twin-enabled health management beyond scalar SOH; Section SAFETY DIAGNOSIS AND RISK-CONSTRAINED FAST CHARGING turns to safety diagnosis and risk-constrained fast charging; Section DEPLOYMENT REALISM, BENCHMARKING, AND DATA GOVERNANCE addresses deployment, benchmarking, and data governance; and Sections OUTLOOK: TOWARD TRUSTWORTHY AND AUTONOMOUS BATTERY TWINS and CONCLUSION synthesize future directions and the main conclusion.

Table 1 positions representative prior work according to the role it plays in this Review: data-centric battery science, probabilistic prognostics, physics-guided modeling, operando observability, field-data deployment, and lifecycle data governance.

Health estimation, safety protection, and fast charging are often treated as separate engineering problems, yet in practice they are different manifestations of one coupled degradation landscape. Mechanism-aware analyses show that capacity fade, power fade, plating propensity, and safety escalation are linked through shared electrochemical and mechanical pathways, including solid electrolyte interphase (SEI) growth, electrolyte depletion, lithium plating, active-material loss, and particle cracking^[5,13]. This observation matters because very different internal states can produce similar external capacity-loss trajectories, especially when only voltage or resistance are used as labels^[13]. A useful digital twin must therefore explain not only how much performance has been lost, but why it has been lost.

This is why degradation modes are more informative than scalar health labels. The now common decomposition into lithium inventory loss (LLI), positive-electrode loss of active material (LAM_PE), and negative-electrode loss of active material (LAM_NE) offers a physically meaningful middle layer between latent mechanisms and measurable performance^[5,13]. Once health is expressed in this intermediate form, state estimation, remaining-life prediction, and charging control can be connected more naturally because the twin can reason over the mechanism space instead of over a single capacity number.

Fast charging is the harshest validation case for such reasoning. Under low-temperature, high-current, heterogeneous internal transport conditions, local overpotentials rise, and lithium plating becomes more likely, while thermal and structural nonuniformity amplifies the damage^[6,7]. Operando pressure measurements^[7], thermal-wave sensing^[8], ultrasound visualization^[9], and MHz electromagnetic diagnostics^[10] further show that plating and related side reactions are spatially distributed, time-varying phenomena rather than discrete threshold events. As a result, the limiting quantity for fast charging is not current alone but the joint evolution of latent electrochemical, thermal, and risk states.

The same logic extends from cell to pack scale. In practical EV systems, the relevant state space includes not only cell-level state of change (SOC) and degradation modes, but also pack-level thermal gradients, current maldistribution, worst-cell margin, and evolving operational context^[11,12]. The lab-to-field gap therefore changes the task definition of a digital twin: the objective is no longer the accurate fitting of curated cycling data, but robust state reconstruction and decision support under partial observability, domain shift, and sparse ground truth.

Two analytical consequences follow from this observation. The first is that identifiability becomes a central issue. If different combinations of LLI, LAM_NE, interfacial impedance growth, electrolyte depletion, and structural heterogeneity can reproduce similar external degradation curves, then capacity-only model fitting is intrinsically underdetermined. A management framework built only on terminal variables may therefore remain blind to whether the limiting factor is lithium inventory, active-material accessibility, transport resistance, or plating risk. This is not a semantic nuance. It directly affects whether a battery should be charged more conservatively, thermally conditioned, derated, repurposed, or retired^[10,13].

The second consequence is that health, safety, and fast charging cannot be separated by time scale as cleanly as they often are in engineering practice. Slow interphase growth, transition-metal dissolution, and active-material isolation reshape local transport pathways^[13,21]; those transport limitations alter overpotential distributions during fast charging^[6,7]. Elevated local overpotentials, in turn, increase the likelihood of plating, gas generation, or localized heat release^[9,22]. In this sense, fast charging is not an independent objective layered on top of a healthy battery. It is an operational regime that exposes the evolving internal state of the battery more aggressively than mild cycling does. Recent work on air-exposure failure of single-crystal high-nickel cathodes further illustrates that storage history can alter interfacial chemistry before apparent capacity loss becomes severe^[23].

This coupling is particularly visible when degradation is viewed through the lens of propagation. At the cell level, the relevant progression may be from interfacial growth to transport limitation, then to plating or self-heating. At the pack level, the same progression is amplified by nonuniform cooling, electrical interconnection, and balancing constraints. A cell that is only moderately aged in isolation can become the worst cell in a series-connected string, limiting both pack power and safe charging current. The cell-to-pack transition is therefore not a mere scaling problem; it changes which latent variables become safety-critical and which observables retain diagnostic value^[11,12].

Recent safety-focused studies have begun to articulate this bridge more explicitly. Degradation-mode analysis and spatially resolved plating models connect ageing to mechanism evolution^[13,14], while pack diagnostics and fault-diagnosis frameworks link local faults to system-level risk^[24,25]. That viewpoint is highly relevant here because battery digital twins must eventually operate across the same continuum. A useful twin should therefore not be designed only to estimate gradual ageing or only to detect rare events. It should instead be able to represent how long-term ageing modifies near-term hazard thresholds and how abnormal operation reshapes future degradation trajectories. Similar scale coupling is evident in abnormal-thermal studies of all-solid-state batteries, where cracking and lithium-dendrite growth remain coupled to local thermal and mechanical states^[26].

Figure 1 summarizes this coupled landscape from application need to latent mechanism. The application map in Figure 1A sets the management requirements. Figure 1B and C then show that capacity fade and power fade arise from interacting degradation submodels and degradation modes rather than from a single ageing coordinate^[13]. Figure 1D adds spatially resolved plating dynamics^[14], and Figure 1E adds electrothermal nonuniformity under spatial thermal gradients^[22]. Figure 1F-H show how these mechanisms become difficult to infer in EV field data^[12]. Figure 1I therefore frames the digital twin as a systems-level response to a many-to-one inverse problem: multiple latent degradation pathways can project onto similar terminal signals, so management requires an internal-state representation rather than a scalar label alone.

A useful way to make this coupling more concrete is to separate the latent failure landscape from its observable manifestations. At the latent level, the primary actors are interphase growth, active-material isolation, electrolyte consumption, plating, crack nucleation, oxygen release, and heat generation. At the observable level, the quantities seen by the BMS are far more compressed: voltage trajectory, apparent capacity, impedance rise, relaxation behavior, pressure, temperature, and, occasionally, gas emission. The mapping from the first set to the second is many-to-one. The same terminal voltage depression can arise from lithium loss, transport limitation, thermal imbalance, or contact degradation. Conversely, a small change in one observable, such as dP/dQ, may carry disproportionate information about plating-sensitive mechanics^[7], field-data ambiguity^[12], degradation-mode/plating non-uniqueness^[13,14], thermal-gradient effects^[22], or early-trajectory uncertainty^[27]. The digital-twin problem is therefore fundamentally an inverse problem under structural non-uniqueness.

Recent degradation-and-safety studies further suggest that ageing modifies not only the mean performance trajectory but also the shape of the hazard landscape^[13,28-30]. Aged cells may exhibit lower tolerance to mechanical abuse^[31,32], altered separator resilience, earlier self-heating onset^[28,29], modified gas-evolution chemistry^[33], and different thermal-runaway heat release^[34,35]. This means that degradation cannot be treated as a neutral background trend that simply lowers capacity. It actively changes the triggering conditions, escalation pathways, and observability of extreme events^[30,36]. Put differently, ageing does not merely precede safety failure; it conditions the entire fault-to-hazard transition. For a digital twin, this implies that safety states should be conditioned on health states, and health states should be updated with awareness of safety-relevant operating history.

A related point is that fast charging should be understood as an information-rich perturbation. Under mild cycling, many harmful mechanisms evolve slowly and remain partly latent^[13,21]. Under fast charging, closed-loop high-rate operation amplifies kinetic and thermal sensitivity^[6]; pressure, thermal-wave, ultrasound, and electromagnetic measurements then make otherwise subtle local reactions more observable^[7-10]. This is one reason why fast charging is a stringent validation scenario for battery twins: it magnifies both the opportunity for state discrimination^[15] and the penalty for model misspecification under thermal and electrochemical heterogeneity^[22,37]. A twin that remains well calibrated under fast charging is likely to be informative under gentler conditions; the converse is not necessarily true.

The coupling among degradation, fault development, and catastrophic failure can also be viewed through a timescale hierarchy. Slow processes such as SEI growth^[13,38], lithium inventory loss^[13], and contact degradation^[21] reshape the baseline electrochemical state over weeks to months. Intermediate processes, such as localized lithium deposition^[7,10], gas generation^[33], or thermal imbalance under repeated high-rate operation^[22], evolve over cycles or hours. Very fast processes, such as internal short circuits^[31,32], violent gas release^[33,35], or thermal runaway^[39,40], can unfold in minutes or seconds. A battery digital twin that aims to be decision-useful must therefore preserve continuity across these timescales. It is not enough to estimate long-term health without relating it to short-term hazard, nor to detect short-term anomalies without situating them within the ageing state that conditioned them^[13,14].

This continuity matters because interventions also operate on different timescales. Design changes, electrolyte choices, and formation protocols act before the battery is deployed^[41]. Thermal conditioning, charging control, and balancing operate during use. Maintenance, warranty, repurposing, and recycling decisions act later in the lifecycle^[42]. The same hidden state may influence all of them, but in different ways. For example, a pack with modest average capacity fade, but a strongly degraded worst-cell may still be acceptable for some use scenarios while being inappropriate for extreme fast charging or second-life deployment. An integrated digital twin is attractive precisely because it can preserve this causal continuity across lifecycle stages rather than allow the battery to be reinterpreted independently at each stage^[19,20].

The degradation-to-safety bridge can also be interpreted in terms of progressive versus abrupt failure regimes. In a progressive regime, latent damage accumulates through repeated cycling, thermal gradients, local plating, gas accumulation, or contact degradation^[13,21], until the battery enters a state in which even routine operation becomes hazardous. In an abrupt regime, an acute trigger such as severe short-circuiting or mechanical damage drives the battery into failure on a short timescale^[24,39]. Yet the severity of that event is still conditioned by the pre-existing degradation state^[40,43]. This distinction is useful because it shows that battery safety is neither wholly slow nor wholly sudden. It is conditioned by ageing but revealed by events. A practical digital twin must therefore represent both the gradual movement of the battery toward risk and the acute transitions by which that risk becomes visible^[14].

A remaining scientific gap is to translate degradation modes into risk thresholds that are both physically identifiable and operationally actionable. Studies on localized high-temperature instability show that similar capacity loss may correspond to different thermal-stability envelopes^[21,44,45]. Work on side-reaction amplification^[46] and thermal-runaway onset^[35,47] further indicates that heat release and self-heating pathways depend strongly on interfacial chemistry and abuse history. Reports on aged-cell risk asymmetry^[28,29] and sudden-death failure^[30] show that the same nominal SOH can conceal very different escalation routes. In other words, the relevant state is not merely how much the battery has degraded, but how that degradation reshapes the latent pathways by which faults escalate. This is precisely the level of closure a useful digital twin must achieve: it must connect LLI/LAM-type degradation coordinates to transport limitation, self-heating propensity, short-circuit susceptibility, and thermal-runaway severity rather than treating these as independent labels.

One of the clearest consequences of this framework is that health management can no longer be organized around a single scalar SOH definition. Vehicle-oriented discussions of SOH emphasize that capacity-based, energy-based, and resistance-based definitions each capture only part of the future usefulness of a battery^[11]. A twin that is meant to support charging, safety, second-life grading, or warranty decisions therefore needs a vector-like health representation that can simultaneously encode capacity loss, power limitation, degradation-mode composition, and risk margins^[54].

Practical health management begins with deployable health indicators. Early machine-learning pipelines demonstrated that carefully engineered features extracted from partial charge curves can estimate SOH with usable confidence intervals^[79]. More recent domain-knowledge-guided work moves this idea closer to the field by constructing indicators that remain meaningful under fragmented charge windows and real driving discharges^[56]. This is an important conceptual shift: useful health indicators are not simply statistically predictive; they must also remain computable under realistic operating constraints.

A second shift is from capacity estimation to degradation diagnostics. If multiple internal pathways can produce similar capacity-loss curves, then maintenance and charging decisions require more than a scalar estimate. Mechanism-aware studies now argue that LLI, LAM_NE, and LAM_PE should be treated as explicit diagnostic targets, not merely as post hoc explanations^[5,13]. In this context, information-rich measurements such as electrochemical impedance can act as compressed fingerprints of hidden electrochemical processes and support probabilistic forecasting even when detailed usage history is unavailable^[54].

The third shift concerns generalization. Health models trained on single chemistries or narrow laboratory protocols often fail when transferred to new manufacturers, broader operating conditions, or field data. Domain-adaptive deep learning has shown that usable SOH estimation can be obtained without repeating full degradation experiments for every target battery^[77], while inter-cell deep learning has demonstrated improved lifetime prediction across diverse ageing conditions and even across chemistries^[80]. These results suggest that the right question is no longer whether transfer is possible, but how much mechanistic and statistical structure is needed to make transfer trustworthy.

Field data are now sufficiently rich to make this transition concrete. Open-source EV datasets have enabled multimodal SOH estimation from large numbers of vehicles, exposing both the opportunity and the complexity of pack-level health assessment^[12]. At the same time, the historical lesson from early-life prediction is still important: deliberately designed datasets and informative early-cycle features remain powerful when paired with physically meaningful targets^[55]. For digital twins, the implication is that cell-level diagnosis, pack-level heterogeneity management, and fleet-level learning should be treated as a single service continuum rather than as isolated modeling problems.

This is also where hybrid digital twins show practical strength. Fast-charging-specific twins can translate observed aging into mechanism-level interpretations^[15], while mechanistically guided residual learning offers a route to continuous monitoring throughout life without repeated offline recalibration^[18]. Taken together, these developments point to a health-management architecture in which lightweight onboard monitoring is periodically corrected by richer cloud-level inference, and both are anchored to mechanism-aware latent states rather than to a single empirical health score.

The notion of vector health is useful precisely because it resolves several ambiguities that have historically been hidden inside scalar SOH. Capacity loss does not distinguish between loss of lithium inventory and loss of active material^[13]. Resistance increase does not by itself identify whether the limiting process lies in charge transfer, ionic transport, contact loss, or heterogeneous utilization^[54]. Even the same numerical SOH can imply very different remaining safe charging windows, thermal sensitivities, and second-life values depending on how that SOH was reached^[72]. A digital twin can only support lifecycle decisions if its health representation preserves some of this mechanism-level structure^[10].

This is why early-cycle information has regained importance. The value of early-cycle features is not merely that they allow earlier forecasting, but that they can encode intrinsic tendencies of a cell before heavy ageing obscures them. Mechanistically guided residual learning supports continual state correction^[18], while chemistry-enabled and early-trajectory studies show how early signals can shape future ageing paths^[43,55]. Generative and foundation-model approaches further suggest that these early signals can be embedded into richer trajectory models^[72,81]. In practice, this would move health management from static diagnosis toward continual prognosis.

Generalization remains the decisive bottleneck. A model that performs well only on a single chemistry, protocol, or fleet is not a digital-twin solution but a local regression model. Real deployment requires transfer across manufacturers, climates, pack topologies, and charging habits. Generative degradation models and ageing/safety analyses reveal why extrapolation is difficult^[27,39,40]. Fleet-level collaborative learning and impedance-aware field diagnosis provide routes to personalization under heterogeneous data^[48,54-56]. Fragmented-capacity estimation and domain-adaptive learning show how partial evidence can still support transfer^[64,77], while inter-cell prediction and additional transfer studies extend this logic across ageing conditions^[80,82]. A mature health twin should not merely survive domain shift; it should indicate when a new operating regime lies outside the reliable envelope of its prior experience.

The health-management literature is therefore converging toward a representation in which observables, latent states, and future trajectories are learned jointly rather than sequentially. Fragmented-charge capacity and partial-window features support health estimation under incomplete field data^[76,83-85]. Multimodel fusion and personalized health prediction connect terminal signals to mechanism-aware health states^[86,87]. Voltage-anomaly and short-circuit diagnostics then extend this state representation toward fault-sensitive health tracking^[88-90]. Early-trajectory prediction and cloud-assisted health updating extend the representation toward prognosis^[91-93], while realistic fault detection and relaxation-voltage diagnosis support fleet deployment^[94-96]. This perspective also clarifies why health estimation, prognostics, and second-life grading should not be developed as isolated subfields, but as connected inference tasks sharing a common state representation.

Figure 5 translates this argument into health-management workflows in the order from observable feature construction to lifelong prognosis. Figure 5A shows multimodal feature engineering from EV field data, and Figure 5B embeds these features into a deep-learning SOH-estimation framework^[12]. Figure 5C and D connect a deployable power-autocorrelation indicator to capacity loss under realistic operating conditions^[67]. Figure 5E addresses available-capacity estimation under fragmented charging, where complete laboratory cycles are unavailable^[82]. Figure 5F and G emphasize that lifetime prediction must be evaluated across diverse ageing factors and capacity-fade trajectories^[86]. Figure 2G-H extends health inference toward mechanistically guided residual learning for lifelong SOC/SOH correction^[18], whereas Figure 5H illustrates early full-lifecycle prediction through BatteryGPT^[27]. The figure therefore supports a vector-health view in which observables, latent mechanisms, transfer regime, and uncertainty are jointly relevant to decisions.

The shift from scalar SOH to vector health also changes how one should think about interpretability. In much of the battery literature, interpretability is treated as an attribute of the algorithm^[5]. In practice, the more consequential notion is interpretability of the state representation: a transparent algorithm is of limited value if it estimates a quantity that is not aligned with the maintenance or charging decision^[13]. Conversely, a moderately complex model may still be useful if it yields a health representation that can be decomposed into physically meaningful limiting factors^[54,72]. For battery twins, interpretability therefore begins with what is estimated, not only with how it is estimated.

This has immediate implications for pack management. A battery pack does not fail gracefully according to the average cell; it is constrained by the least healthy cell, the hottest region, the most resistive connection, or the most uncertain safety margin^[11,12]. A twin that estimates only fleet-average or pack-average health may therefore underestimate operational risk. In contrast, a vector-health formulation can preserve pack dispersion, worst-cell margin, and uncertainty around both^[18,48]. This is essential when health outputs feed into balancing, charging derating, warranty assessment, or second-life grading^[55].

Another underappreciated issue is that health estimation and prognosis operate on different evidence regimes. Current-state estimation can rely on short windows and local indicators^[55], whereas long-horizon prognosis requires assumptions about future operating conditions. The most credible health twins will therefore usually combine mechanistic current-state estimation with scenario-conditioned prognosis rather than extrapolate a single trend line indefinitely^[18,43]. Generative sequence models and multimodal latent representations may improve this transition^[27,72,81], but only if such models remain anchored to physically interpretable targets and calibrated uncertainty estimates^[97-99].

Second-life management makes these issues even sharper. A battery entering repurposing does not arrive with a clean, uniformly sampled health trajectory; it arrives with missing context, uneven stress history, unknown micro-fault burden, and heterogeneous cell dispersion^[48,51]. A practical twin for second-life decisions, therefore, cannot rely solely on nominal SOH labels. It must infer residual capability, uncertainty, and risk under incomplete information^[100], ideally using representations that remain meaningful across the transition from first-life to second-life applications^[10].

This is also why cloud assistance is likely to remain central to high-quality health management. Fleet-scale data provide the diversity needed to learn which features remain stable across vehicles, climates, and charging infrastructures^[11,12]. Onboard systems provide the immediacy required for real-time adaptation. Neither alone is sufficient. The most useful health-management architectures are therefore likely to remain hybrid not only in the model sense, but also in the deployment sense: lightweight local monitoring^[18], event-triggered uplink^[48], and cloud-side inference that returns updated priors, thresholds, and control suggestions to the edge^[49,51].

There is also a methodological lesson here for benchmark construction. Health models should not be compared exclusively at equal levels of target observability. Some methods assume complete charge curves; some assume only partial windows; some rely on impedance or pulse data; some infer trajectory from early cycles^[7,12]. A review that compares only endpoint RMSE without declaring these observability assumptions can mislead readers about what is actually deployable. In digital twins, the operational question is whether the assumed evidence stream exists in the intended application^[48,54], and whether it can be acquired repeatedly without unacceptable cost or disruption^[55,56].

The deployment challenge is often underestimated. In practice, a battery digital twin is not deployed once; it is sustained across sensing, logging, synchronization, update, and decision layers. Vehicle batteries differ in initial grading, cooling conditions, aging history, and operational patterns, so a deployable twin must explicitly manage cell inconsistency, worst-cell limitation, and pack-level thermal heterogeneity^[11,12]. This favors a layered architecture in which the pack is not collapsed into a single lumped state unless the information loss is acceptable for the intended decision.

Deployment changes the criteria by which a battery digital twin should be judged. In laboratory studies, the implicit objective is often predictive fidelity under controlled conditions. In field operation, the objective is narrower and more consequential: maintain safe operation, preserve useful life, and keep the model synchronized with incomplete, noisy, and partially labeled data^[11,12]. This is why deployment should be viewed as an architecture problem rather than a post-model implementation step. The useful twin is the one that can be maintained under imperfect information, not merely the one that performs well offline^[48,49,51].

This, in turn, reframes the value of cloud-edge-fleet architectures. Their purpose is not simply to centralize data, but to distribute intelligence across time scales and decision types^[12,16]. Local filtering and immediate protection remain naturally onboard. Heavy health diagnosis can be cloud-mediated when standardized pack procedures define comparable labels^[42], while drift detection, personalization, and fleet-level anomaly learning are more naturally handled by collaborative or cloud-based architectures^[48,49,51]. The edge layer becomes important when latency, communication cost, or privacy prevents full cloud dependence^[55,57].

Benchmarking must follow the same realism. Many current datasets and evaluations still emphasize narrow laboratory cycling, whereas digital twins are meant to operate across protocol changes, temperature shifts, pack configurations, and field data with scarce labels^[5,12,80]. Benchmark design should therefore separate interpolation from extrapolation, laboratory from field transfer, cell from pack configuration shift, and health prediction from safety warning. Without this separation, apparently strong headline accuracy can hide weak deployment robustness.

Evaluation metrics should also move beyond point-error reporting. Probabilistic battery reviews argue that calibration-oriented metrics such as negative log-likelihood, expected calibration error, and sparsification-based measures are necessary whenever predictions are used to guide control^[5]. For safety tasks, warning lead time, false-alarm rate, and missed-event rate are as important as classification accuracy^[16]. Open-access resources such as the Battery Failure Databank can play an important role here because they offer a common basis for benchmarking rare but consequential thermal-runaway behaviors^[78].

A second prerequisite for reproducible evaluation is improved semantic and regulatory infrastructure. Vehicle-level health measurement is not yet standardized, and recent work has argued for repeatable onboard procedures for energy- and capacity-based SOH at the pack level^[42]. At the same time, the earlier data-centric lesson from battery research still holds: broad and diverse datasets become much more useful when organized around meaningful early descriptors and clearly defined targets^[55]. A robust benchmark therefore needs both shared data and shared measurement definitions.

These concerns now extend into lifecycle governance. Unified battery data descriptions^[19] and multiscale digital modeling across the full lifecycle^[41] provide the semantic backbone needed to connect manufacturing genealogy, testing history, operational data, and end-of-life decisions. This is increasingly important under the EU Battery Regulation, which requires battery passports for EV batteries, light means of transport batteries, and industrial batteries above 2 kWh from 18 February 2027^[20]. In that context, the digital twin becomes more than an internal engineering asset; it becomes part of the auditable digital thread that links performance, safety, value retention, repurposing, and recycling. Strategic lithium-ion battery recycling studies further indicate that pack-level records and material provenance are necessary to connect digital-twin outputs with resource-risk management and circular value chains^[151].

Benchmark design deserves separate emphasis because the field remains too tolerant of accuracy numbers generated under unrealistically narrow conditions. A battery digital twin should be benchmarked along at least four dimensions: uncertainty quality^[5], deployment realism from cell to field^[12], mechanism-level transfer^[13,14], and system scale. Deployment realism asks whether the data originate from controlled cycling, semi-realistic duty cycles, or field operation^[54,55]. System scale asks whether the model is evaluated at the material/interface, cell, pack, or fleet level^[16,17,42]. Task type separates health estimation, safety warning, fast charging, and lifecycle decision support; uncertainty quality asks whether the prediction is calibrated rather than merely accurate^[77,80].

Data governance is not an administrative afterthought to this problem; it is part of the scientific architecture. Once the digital twin is deployed beyond a single-cell test stand, battery identity, provenance, manufacturing history, beginning-of-life characterization, field usage, maintenance events, second-life grading, and end-of-life processing all become causally relevant^[19,20]. Without a semantic layer linking these descriptors, apparent prediction errors may reflect hidden differences in labels^[41], measurement protocols^[42], or battery genealogy. Governance therefore determines whether data remain comparable^[104], auditable^[105], and reusable across the lifecycle^[139].

The governance problem also has a scientific dimension. Standardization, pack-level measurement procedures, and semantic continuity determine which deployment claims are testable across institutions, fleets, and lifecycle stages^[33]. A benchmark that ignores privacy constraints, owner heterogeneity, measurement protocols, or regulatory fields may still be publishable, but it is not deployable. Equalization, magnetic, and optical monitoring hardware provide one layer of deployability^[152-154]. Flexible microsensors and battery-monitoring front ends define how those signals enter pack electronics^[155-158]. Self-heating and embedded temperature measurements define another layer of safety-relevant observability^[159-161]. Heat-generation, radial-temperature, and thin-film thermocouple measurements extend this observability into spatial thermal characterization^[162-164]. Gas detection and failure sensing expand the evidence base for early warning^[165], while federated estimation under privacy constraints reinforces the need to treat deployment realism, auditability, and semantic interoperability as first-class scientific requirements rather than downstream engineering details.

Figures 8 and 9 close the deployment and benchmarking argument in sequential order. Figure 2D and E show privacy-preserving collaboration for retired-battery sorting without raw-data exchange^[49]. Figure 8A and B expose battery-type heterogeneity across charging-station owners, and Figure 2F shows how a personalized federated fault-warning architecture handles that heterogeneity^[48]. Figure 8C illustrates second-life SOH estimation under random retirement conditions and deployment without extra SOC conditioning^[69]. Figure 8D adds the lifecycle digital-thread layer that connects battery identity, provenance, health, compliance, repurposing, and recycling. Figure 9 then shifts to benchmark realism: Figure 9A and B show real-world direct current (DC) fast-charging capacity distribution and temperature-dependent charging-profile heterogeneity^[166]; Figure 9C and D address standardized pack-level SOH measurement and differential-voltage transfer from cell to vehicle level^[24,42]; Figure 9E and F show that evaluation should include realistic data flow, data-frame heterogeneity, ROC curves, and reconstruction error rather than point accuracy alone^[117]; Figure 9G adds pack-voltage variability under fixed voltage windows^[42]; and Figure 9H summarizes the benchmark matrix and roadmap for deployable twins. Together, Figures 8 and 9 show that deployment and benchmarking are not external to the digital-twin problem; they define its engineering meaning.

Dataset design is central to this transition. A dataset suitable for academic model comparison is not automatically suitable for battery-twin deployment^[12,13]. Deployment-oriented datasets must preserve context: battery identity, chemistry, pack topology, environment, sampling cadence, control history, maintenance events, and the provenance of labels^[17,48]. Otherwise, two apparently identical voltage segments may correspond to different degradation modes^[54], thermal histories^[55], or control constraints. Safety datasets also need explicit rare-event metadata^[77,140] and abuse-condition descriptors if they are to support state-of-safety benchmarking^[78].

The notion of benchmark realism should therefore be broadened beyond data source. Realism also concerns the perturbations that the benchmark allows. A useful benchmark should test adaptation to temperature shift^[5], pack scaling^[13,14], partial charging^[42], rare-event imbalance^[55], missing labels^[77], and changing decision cost^[80]. In battery digital twins, a benchmark that does not expose the model to these distortions may be a test of interpolation, but not a test of deployability.

Regulation further sharpens these requirements. Battery passports and traceability mandates make standardized data descriptors part of the deployment problem^[19,20], while lifecycle reporting obligations require the twin to support explainable and auditable decisions^[41,42]. This is particularly relevant for health-based warranty^[12], second-life valuation^[14], and recycling readiness, where the decision must remain traceable to measurement history, model assumptions, and uncertainty estimates^[51]. Recent graphite-anode recycling analysis similarly shows that degradation understanding should be retained into targeted regeneration and recycling decisions^[167].

Another practical issue is that large-scale deployment requires explicit coexistence of model diversity. The same fleet may contain different chemistries, form factors, pack topologies, software versions, and sensor layouts^[12]. A digital-twin architecture that assumes one model per fleet or one metric per chemistry will quickly become brittle. More realistic architectures will likely require shared state semantics^[19,20], model personalization^[51,55], and uncertainty-aware switching among local, cloud, and fleet-level models^[41,42].

A final deployment consideration concerns auditability. If the twin is used only for internal analysis, modest opacity may be tolerable. If it influences warranty decisions, insurance risk, second-life valuation, charging limits, or safety interventions, then the rationale for its outputs must be inspectable^[20,42]. This does not mean that every subsystem must be fully transparent, but it does mean that inputs^[12], model version^[14], uncertainty estimate^[51], and action recommendation should remain traceable. Auditability is therefore a deployment constraint, not only a regulatory preference^[104,105].

Data resources are likely to determine the pace of progress as much as any individual algorithm. The field now has access to open EV datasets^[12], retired-battery datasets^[51], battery-failure databases^[78], and increasingly rich laboratory datasets with impedance, imaging, and operando sensing^[13,14,42]. Yet these resources are rarely aligned semantically. Different datasets define health differently, record metadata unevenly, and provide safety labels at different levels of granularity^[54,55]. This is why benchmark design must be coupled to data standardization, not treated as a separate post-processing exercise^[140].

This is also why future benchmark initiatives should not separate data quality from governance. A benchmark that is impossible to reproduce, impossible to interpret semantically, or incompatible with lifecycle traceability will not support industrial translation even if it produces informative machine-learning leaderboards^[20,42]. Strong benchmark design in this field therefore requires defined test regimes^[54], clear labels^[55], uncertainty metrics, metadata standards, and auditable provenance^[77].

Finally, the maturation of battery digital twins will depend on whether the community can align scientific rigor with industrial traceability. A high-performing health model is insufficient if its labels are not portable across platforms, its uncertainty is uncalibrated, or its outputs cannot be audited at the pack, fleet, or regulatory level. Conversely, a robust digital thread without mechanism-aware inference will still fail to deliver actionable intelligence. The next phase of progress therefore lies at the intersection of electrochemical understanding, reliable data semantics, lifecycle governance, and computational decision theory. Only by treating these as coupled requirements, rather than as separate research tracks, can battery digital twins become a stable technical foundation for fast charging, predictive safety, warranty management, repurposing, and circular battery value chains.

An equally important open issue concerns the coupling between management decisions and information acquisition. Most battery studies still assume that sensing is passive and fixed, but practical digital twins should also decide when to probe the system more aggressively. A short rest period, a diagnostic pulse, a low-amplitude impedance measurement, or a temporary charging-rate modulation may all be interpreted as information-seeking actions, because they improve the observability of hidden states such as plating tendency, resistance growth, or pack imbalance. In this sense, battery management is not only an optimization problem over energy and power, but also an optimization problem over uncertainty reduction. Future twins will likely combine passive monitoring with selective active interrogation, thereby coupling estimation, control, and diagnostic experiment design within the same closed-loop framework. This shift from passive prediction to active state clarification is particularly relevant for safety-critical operation, where uncertainty itself is often the variable that must be managed.

A further conceptual challenge is that the twin must remain identifiable under changing evidence regimes. In early life, electrochemical signatures are information-rich, yet the damage state is small. In mid-life, degradation modes become separable, however, operating histories diverge; late in life, pack constraints, thermal asymmetry, and cell dispersion dominate the practical risk picture. A deployable twin should therefore not be calibrated once and then only updated numerically. It should adapt its own reliance on model classes, latent states, and observables as the battery ages. For example, a newly commissioned pack may benefit from impedance-informed initialization and formation-aware priors, whereas an aged field pack may rely more on uncertainty-aware residual learning, abnormality scoring, and conservative control thresholds. This age-dependent reweighting of evidence is rarely made explicit in current literature, yet it is essential if digital twins are to remain scientifically interpretable and operationally useful across the full battery lifecycle.

The preceding sections suggest that progress in battery digital twins will be limited less by isolated algorithmic accuracy than by whether the field can construct state representations that are observable, identifiable, calibrated, and useful for intervention. Three questions should therefore be addressed together: what the twin can see, what physical state it represents, and how uncertainty is converted into action. Treating sensing, modeling, control, and governance as separate topics is no longer sufficient for deployable battery intelligence.

The first priority is active and economical observability. Future twins will not reliably infer plating, thermal-runaway precursors, degradation-mode evolution, or pack imbalance from voltage, current, and surface temperature alone. Pressure sensing^[7], thermal-wave measurements^[8], ultrasound imaging^[9], and MHz-band electromagnetic diagnostics^[10] each make different hidden states more identifiable. Interfacial stability and fault-mechanism studies define which safety signals are physically meaningful^[21,25], while side-reaction and online impedance analyses refine that mapping^[46,47,53]. Industrial-scale twin architectures and field-learning frameworks then determine whether these signals can be deployed at scale^[113,114]. Current-sensor and field-health electronics define one part of the hardware cost of that deployment^[133-135], and rapid pack health/degradation models define another^[136,137]. Shunt and fluxgate devices define additional pack-monitoring options^[138-141], while magnetometer-based sensing further extends current-measurement capability^[142,143]. The key problem is therefore not maximal sensing but value-of-information design: what is the cheapest evidence stream that makes a decision-relevant hidden state identifiable?

The second priority is state-sufficient multiphysics modeling. A useful twin does not need to reproduce every microscopic detail of a cell; it must retain the internal variables needed for the intended decision. Physics-plus-machine-learning frameworks define the general management architecture^[4], while fast-charging degradation diagnosis provides a decision-specific example^[15]. Physics-informed neural networks and mechanistically guided residual learners show how learnable models can remain state-aware^[17,18]. Cross-scale electrochemical-mechanical modeling and lifecycle digital modeling extend this state representation across physical scales^[37,41], and impedance-based forecasting illustrates how reduced evidence streams can still support probabilistic state prediction^[54]. The open problem is to define, for each task, which latent coordinates are sufficient: degradation modes for health management, plating and thermal margins for charging, state of safety for hazard mitigation, and traceable descriptors for lifecycle governance.

The third priority is calibrated uncertainty under distribution shift. Field operations continuously change the evidence regime. Early-life data are information-rich but weakly degraded^[27], mid-life trajectories are mechanism-separating, but user-dependent^[11,12], and late-life operation is dominated by pack dispersion, thermal asymmetry, and risk concentration^[18]. A deployable twin should therefore adapt not only its state estimate but also its confidence estimate. This is the difference between a predictor that extrapolates and a twin that knows when it is leaving its validated domain^[5,80].

Figure 10 illustrates this age-dependent rebalancing. In beginning-of-life operation, the physical model can be weighted more strongly because geometry, material properties, formation history, and initial parameters remain close to the calibrated design space. During mid-life operation, the evidence regime becomes mixed: degradation modes become more separable, but user-dependent histories and partial observations increase the need for data-driven residual correction. Near end-of-life, cell-to-cell dispersion, thermal asymmetry, and localized ageing increasingly reduce the reliability of a purely nominal physics model; the twin should therefore increase the contribution of data-driven anomaly scores, residual learners, and uncertainty-calibrated thresholds while retaining physical constraints to prevent nonphysical extrapolation. The practical rule is not to replace physics with ML but to update the fusion weights according to model residuals, uncertainty calibration, and the information content of incoming measurements.

The fourth priority is benchmark realism. Battery twins should be tested across distinct regimes: interpolation within known laboratory distributions^[5], extrapolation across controlled but unseen operating conditions^[13,14], transfer across chemistry or configuration^[42], and in-the-wild operation under fragmented data and uncertain labels^[78,80]. Reporting a single point-error metric across these regimes is not informative enough. Future benchmark reports should separate mean accuracy, calibration, robustness, warning lead time, latency, and auditability.

The fifth priority is semantic and regulatory continuity. Battery identity, manufacturing genealogy, beginning-of-life characterization, field operation, maintenance events, safety incidents, second-life grading, and recycling decisions are not peripheral metadata once a twin is deployed beyond the laboratory^[19,20]. They determine whether health labels are portable, whether uncertainty is interpretable, and whether lifecycle decisions can be audited^[41,42].

A final direction is cautious autonomy. Closed-loop optimization can move battery management from monitoring to supervisory action^[6], but autonomy is credible only when the action policy remains constrained by physical state, calibrated uncertainty, and traceable decision logic. The objective is not to replace electrochemical understanding with end-to-end automation. It is to use learning to determine when measurements, mechanisms, and uncertainty are sufficient for a safe action^[168], when additional probing is required, and when conservative intervention is necessary under reinforcement-learning or safe-control settings^[169,170]. Abnormal-thermal solid-state battery analysis illustrates why autonomous decisions must remain constrained by failure-mechanism awareness^[26]. Recycling-oriented studies likewise show that digital-twin outputs should remain connected to material recovery and resource-risk decisions rather than end at first-life operation^[151,167].

From this perspective, the long-term value of digital twins lies in reorganizing battery management around state sufficiency, uncertainty calibration, and lifecycle accountability. If this reorganization succeeds, apparently separate functions - state estimation, degradation diagnosis, fast-charging supervision, fault warning, second-life grading, and regulatory traceability - become different queries to the same evolving representation of the battery system.

This Review argues that lithium-ion battery health management, safety diagnosis, and fast-charging control should not be treated as independent software functions. They are different decision aspects of the same coupled electrochemical, thermal, mechanical, and interfacial degradation system. A digital twin that only predicts scalar SOH or reproduces terminal voltage is therefore insufficient for modern battery management because it does not preserve the mechanism-to-action link required for safe operation.

A deployable battery digital twin must be physics-grounded, uncertainty-calibrated, and decision-oriented. It should fuse multimodal observability with task-matched mechanistic models, infer hidden degradation and hazard states with calibrated uncertainty, and translate those states into control-relevant quantities such as charging margins, safety warnings, maintenance priorities, and lifecycle decisions. This is a stricter requirement than high prediction accuracy under laboratory cycling; it requires robustness under fragmented field data, pack heterogeneity, privacy constraints, and evolving evidence regimes.

The most consequential advances will therefore stem from the explicit coupling of observability, mechanism, uncertainty, and action across the battery lifecycle. When benchmarking, data semantics, and governance mature alongside model development, the digital twin can move from an attractive concept to an auditable battery-intelligence layer for safe fast charging, predictive safety, durable operation, repurposing, and circular battery value management.