EV Batteries Live and Die by Their Weakest Cell, Real-World Data Shows

Somewhere past 170,000 kilometers on the odometer, a quiet divergence begins deep inside an electric vehicle's battery pack. Among the dozens or even hundreds of cells wired together, just a handful start losing capacity and gaining resistance far faster than the rest. Because a battery pack is an electrical circuit with its cells connected in series, the whole system must throttle its output and cut off charging to match whatever the weakest link can handle. Even if most individual cells still have plenty of energy left to give, the moment the frailest cell hits its safety limit, the entire pack is effectively retired.

A paper published in Nature Energy on June 23, 2026 (DOI: 10.1038/s41560-026-02131-5) quantifies just how much cell-to-cell degradation variability constrains overall pack performance, drawing on large-scale telemetry from real-world commercial and private EV fleets. The research team—led by Professor CHEN Zhongwei of the Dalian Institute of Chemical Physics, Chinese Academy of Sciences, and Professor ZOU Changfu of Chalmers University of Technology in Sweden, together with researchers from Beijing Institute of Technology and Zeekr Technology Europe AB—built their analytical framework not on accelerated lab degradation tests, but on massive operational data from vehicles actually driving on the road. Litao Zhou and Xiaolei Bian share first authorship as equal contributors.

The numbers the paper arrives at are substantial. Cell-to-cell nonuniformity cut usable pack state of health (SOH) by 6.2% in passenger cars and 7.5% in buses. Compared to the average lifespan of individual cells, the effective pack lifespan was shortened by 17.7% in passenger cars and 22.8% in buses. Power capability losses tied to variation in internal resistance were estimated at 12.9% and 15.1%, respectively. And when the researchers combined these effects into a single lifetime energy-resource utilization metric, passenger car packs reached only 80.7% and bus packs only 72.9%—meaning that by the time the weakest cell triggers retirement, somewhere between 19.3% and 27.1% of the pack's theoretically extractable potential energy is left stranded, unused.

From lab benches to 500 million rows of fleet telemetry

For decades, research into lithium-ion battery degradation has largely relied on charge-discharge cycling tests performed inside temperature-controlled environmental chambers. Repeatedly measuring a single cell under tightly controlled temperature and current is well-suited to unpacking the fundamental electrochemical reaction steps involved. But this approach has never been able to fully capture the complex phenomena that occur inside a real, on-road battery pack.

In an actual battery pack, subtle material variation, manufacturing tolerances in electrode coating, differences in heat dissipation between the center and edges of the pack, and the tangled influence of individual driving habits and charging environments all pile on top of each other. Because of these compounding factors, even cells drawn from the same manufacturing batch drift apart in degradation rate once they're installed in a vehicle.

The research team analyzed high-precision operational data transmitted wirelessly at a nominal 0.1 Hz (once every 10 seconds) through China's National Monitoring and Management Platform for New Energy Vehicles. The fleet consisted of 116 passenger cars equipped with prismatic NMC (nickel-manganese-cobalt) ternary cells and 17 electric buses equipped with prismatic lithium iron phosphate (LFP) cells. Data collection spanned more than three years, and some individual vehicles had logged total mileage exceeding 300,000 km.

After preprocessing, the dataset fed into the analysis totaled 587.5 million rows, capturing pack-level voltage and current, per-cell estimated state of charge (SOC), cell voltage, and temperature sensor readings. Rather than relying on the uniform degradation patterns produced in a lab, the study's observational foundation is a large population of vehicles that actually traveled real road networks. The fleet under study remains in active operation; this is not a statistical analysis of packs that physically failed and were scrapped. The retirement thresholds and unused-energy figures discussed in the paper are model estimates, calculated by applying a specific retirement criterion to degradation trajectories reconstructed from operational data.

A three-stage estimation pipeline and six diagnostic metrics reveal internal states

The biggest obstacle in working with real-world driving data is that seasonal swings in outside temperature and variation in charging depth from one trip to the next impose heavy noise on measured voltage and resistance. Internal resistance appears smaller in summer and spikes sharply in winter. To strip away environmentally driven fluctuation and isolate the component of degradation that is intrinsic to each cell, the research team designed a three-stage estimation method.

In the first stage, the team selected data from long, low-rate standard charging sessions—typically below 0.2C—to construct the relationship between open-circuit voltage and state of charge (the OCV-SOC curve). At the same time, they identified ohmic internal resistance ($R_0$) using local linear regression applied to instantaneous voltage changes accompanying stepwise current changes. Rather than relying on differences between single data points, using windowed regression helps suppress the effects of sensor noise.

In the second stage, with $R_0$ fixed, the team applied particle swarm optimization (PSO) to a first-order equivalent circuit model, inversely solving for cell capacity ($C$) and relaxation parameters such as polarization resistance and capacitance.

In the third stage, the resulting capacity and resistance estimates were normalized to a standardized reference temperature and operating condition. A feedforward neural network was used, taking current, temperature, and mileage as inputs and passing them through a single hidden layer of 16 neurons. The network used ReLU activation, a linear output layer, a dropout rate of 0.1, the Adam optimizer with a learning rate of 0.001, mean squared error (MSE) loss, and early stopping within 20 epochs. This mathematically smoothed out seasonal variation and operational bias, pulling every cell back to a common reference environment so they could be fairly compared.

From these standardized parameters, the research team defined six diagnostic metrics to isolate the impact of nonuniformity:

• $\sigma \text{-SOH}$: the degree of variation in state of health (SOH) across cells.
• $\beta \text{-SOH}$: the rate at which the most degraded cell constrains the pack's overall effective health.
• $\beta \text{-life}$: the ratio of the pack's overall effective lifespan to the average lifespan of individual cells.
• $\beta \text{-SOC}$: the rate of charge-capacity loss that balancing cannot recover.
• $\beta \text{-power}$: the degree to which pack power output is limited by variation in internal resistance.
• $\beta \text{-energy}$: lifetime total energy-resource utilization efficiency (the product of $\beta \text{-life}$ and average $\beta \text{-SOC}$).

Rather than the SOH 0.80 (80% of initial capacity) threshold widely used as a retirement criterion for EVs, the study adopted a higher threshold of SOH 0.85, because the vehicles under study were still on the road and the fleet as a whole had not yet reached the conventional retirement limit. The lifespan-shortening rates and unused-energy percentages presented in the paper are directly tied to this modeled assumption of "SOH 0.85."

The 170,000-km inflection point and individual variation across the passenger car fleet

The analysis revealed a distinctive shared pattern among the NMC-equipped passenger car fleet. Up until accumulated mileage reached roughly 170,000 km, variation in SOH across cells ($\sigma \text{-SOH}$) remained extremely low, with cells within each pack aging gradually and in near-lockstep.

• NMC Passenger Cars (116 vehicles)
• LFP Electric Buses (17 vehicles)

データを表で見る

	NMC Passenger Cars (116 vehicles) (%)	LFP Electric Buses (17 vehicles) (%)
Usable SOH Reduction	6.2	7.5
Effective Pack Lifespan Reduction	17.7	22.8
Power Capability Reduction	12.9	15.1
Lifetime Potential Energy Unused	19.3	27.1

As the chart shows, under an SOH 0.85 retirement threshold, both the shortening of effective pack lifespan and the underutilization of energy resources emerge as substantial losses for both passenger cars and buses. Notably, compared to the relatively modest loss in usable SOH from simple capacity decline (6.2% and 7.5%), the reduction in overall lifespan (17.7% and 22.8%) is considerably larger.

Past the 170,000-km mark, certain cells in some vehicles reached what the researchers call an "aging knee"—a point of sharply accelerated degradation—causing variability to widen rapidly. The timing and sharpness of this accelerated degradation varied markedly from vehicle to vehicle. In the worst-performing vehicles in the fleet, cell-to-cell divergence spiked sharply around the 170,000-km mark, with the weakest cell rapidly losing capacity. By contrast, in better-performing vehicles, cell-to-cell variation stayed low all the way out to 300,000 km.

The fleet-wide average $\beta \text{-SOH}$ at retirement was 0.938, meaning effective SOH was reduced by 6.2% relative to a scenario of uniform degradation. In the vehicle with the most severe divergence, $\beta \text{-SOH}$ dropped to roughly 0.83, leaving about 17% of the pack's inherent health capacity unusable.

The knock-on effect on lifespan was even greater. The fleet-average $\beta \text{-life}$ came in at 0.823, meaning pack lifespan was shortened by 17.7% relative to the average lifespan of individual cells. $\beta \text{-power}$, the power-capability metric, averaged 0.871 (a 12.9% reduction in output), with the distribution across vehicles falling roughly between 0.80 and 0.90 (a reduction of 10% to 20%).

$\beta \text{-SOC}$, which reflects the impact of daily charging imbalance on state of charge, stayed above 0.98 for the vast majority of vehicles. Onboard passive balancing kept capacity losses from SOC imbalance under 2%. Lifetime energy-resource utilization, $\beta \text{-energy}$, ranged from 0.72 to 0.86 across the fleet, averaging 0.81. This statistically confirms that the main factor blocking energy extraction isn't day-to-day charging imbalance ($\beta \text{-SOC}$), but rather the premature end of the entire pack's life because one particular cell gives out first ($\beta \text{-life}$).

Metric and vehicle specification	NMC prismatic-cell passenger cars (116 vehicles)	LFP prismatic-cell electric buses (17 vehicles)	Notes and measurement/analysis considerations
Cathode material and cell format	NMC (ternary), prismatic cell	LFP (lithium iron phosphate), prismatic cell	Buses serve as a supplementary reference group, not a direct chemistry-to-chemistry comparison
Series-connected cells per pack	95 cells	156 cells	Buses have more series cells, making thermal gradients more likely to widen
Balancing control	Continuous passive balancing during charging	No explicit voltage balancing control	Bus system design omits voltage equalization
Maximum cell-to-cell voltage variation	Generally stays under 100 mV even after aging	About 100 mV early on, exceeding 200 mV later, up to 300 mV maximum near end of life	Balancing, cell count, and layout differences are reflected in voltage gaps
Effective health ratio ($\beta \text{-SOH}$)	0.938 (6.2% effective SOH loss)	0.925 (7.5% effective SOH loss)	Fleet-average model value based on SOH 0.85 retirement threshold
Pack lifespan ratio ($\beta \text{-life}$)	0.823 (17.7% lifespan reduction)	0.772 (22.8% lifespan reduction)	Ratio of effective pack lifespan to average cell lifespan
Charge-state efficiency ($\beta \text{-SOC}$)	Above 0.98 (capacity loss under 2%)	Above 0.98 (capacity loss under 2%)	Impact of daily SOC imbalance is minor for both
Power retention ratio ($\beta \text{-power}$)	0.871 (12.9% power reduction)	0.849 (15.1% power reduction)	Discharge power limited by nonuniform internal resistance increase
Lifetime energy utilization ($\beta \text{-energy}$)	0.807 (about 19.3% unused)	0.729 (about 27.1% unused)	Product of $\beta \text{-life}$ and average $\beta \text{-SOC}$; actual recovery rate relative to potential

A 156-cell megapack and the absence of balancing drive divergence in the bus fleet

As a reference fleet for comparison with the passenger cars, the research team also analyzed data from 17 LFP electric buses. The results from this bus group suggest how a larger cell count combined with simplified control can accelerate nonuniformity.

At retirement, the fleet-average values for the electric buses were $\beta \text{-SOH}$ of 0.925 (7.5% health loss), $\beta \text{-life}$ of 0.772 (22.8% lifespan reduction), and $\beta \text{-power}$ of 0.849 (15.1% power reduction). As a result, lifetime energy-resource utilization ($\beta \text{-energy}$) fell to 0.729, meaning roughly 27.1% of energy went unused, by calculation.

According to the paper, the drop in energy efficiency for buses is likely rooted in the greater difficulty of maintaining uniformity in a large pack. While each passenger car pack contained 95 cells, the large buses had 156 cells connected in series. As cell count and pack volume increase, spatial temperature gradients along cooling airflow or coolant paths become more pronounced. Cells positioned in hotter zones undergo electrochemical reactions faster than cells kept cool, and their degradation pace drifts out of sync with the rest.

Differences in operation widened the gap further. The passenger cars' battery management systems (BMS) featured continuous passive balancing, bypassing a small current during standard charging, which kept cell-to-cell voltage variation from exceeding 100 mV even as the vehicles aged. The buses analyzed, however, did not incorporate explicit voltage balancing control due to operational constraints. As a result, cell-to-cell voltage differences that started out around 100 mV in early operation surpassed 200 mV over time, eventually widening to as much as 300 mV by the end of the observed period.

The research team cautions against interpreting this contrast as evidence that NMC chemistry is inherently superior to LFP. The passenger car and bus fleets differ not only in battery chemistry but simultaneously in series cell count, geometric layout, cooling structure, driving cycles, and the presence or absence of charge balancing. It's more appropriate to read the presented data not as a controlled materials comparison, but as a reflection of how entire real-world systems—including differences in control strategy and thermal design—behave differently as a whole.

Limitations of the observational data and remaining engineering challenges

While the figures presented offer valuable insight into the real-world state of EV batteries, they also carry limitations inherent to large-scale observational data.

First, because this is remote telemetry gathered during actual operation, it is impossible to fully separate the intrinsic electrochemical degradation mechanisms of battery cells from the transient operational effects of driving conditions and temperature changes. Obtaining direct electrochemical ground-truth measurements—of the kind possible in a lab—for individual cells degrading while actually being driven on the road is fundamentally not achievable. As the authors themselves state plainly in the paper, the calculated metrics should be interpreted not as "intrinsic electrochemical degradation itself," but as "long-term aging indicators standardized under real-world operating conditions."

Second, there is statistical variability inherent to the machine-learning-based standardization model. The research team retrained the neural network 100 times with different random seeds and examined the 5% to 95% confidence interval. They confirmed that in mileage ranges with abundant data, the confidence interval was extremely narrow and stable, but at extreme mileages exceeding 300,000 km, uncertainty in the estimates widened due to the smaller number of training samples available.

Third, the economic loss figure presented by the authors is not a measured financial loss but a rough estimate premised on projected future market size. The paper cites a prior forecast that the global lithium-ion battery market will reach approximately $400 billion by 2030, and by simply multiplying that figure by the roughly 19% of unused energy in passenger cars, arrives at an estimate of roughly $76 billion in potential value left unrecovered. This figure does not account for value recovered through battery recycling or second-life use as stationary storage, and should be read as a rough gauge intended to give readers an intuitive sense of the scale of the engineering problem, rather than a precise figure.

As countermeasures against the energy losses caused by nonuniformity, the authors point to improving manufacturing precision, optimizing initial cell selection and grouping during pack assembly, introducing high-precision active balancing circuitry, advancing thermal management, and even adopting reconfigurable battery systems that use power electronics to dynamically switch the connection topology among cells. If degraded cells could be bypassed while the remaining healthy cells continue supplying power, it may be possible to avoid discarding an entire pack simply because it was dragged down by its weakest cell.

However, all of the proposed countermeasures remain theoretical and design-stage proposals; the study did not demonstrate their improvement effects on actual vehicles. In particular, adding dynamic switching circuitry introduces its own trade-offs: added hardware complexity from extra power transistors, increased weight, reduced volumetric efficiency, power loss from parasitic resistance, and higher upfront cost.

From idealized charge-discharge cycles in the lab to the harsh reality of the open road: the research team has published a GitHub repository (Rico-dicp/Ev-battery-cell-to-cell-inconsistency) containing the full code for preprocessing, OCV-SOC curve estimation, PSO parameter identification, and calculation of each metric, opening the door for outside researchers to verify and build on this work. In an era when millions of EVs travel the streets, the challenge now is shifting our perspective on batteries—from viewing them through the lens of "averages" to grappling, as a matter of systems engineering, with how to control and tame the variability among individual cells.