By Hyun-Soo Lee, Head of Software
Under 800 microstrain of bending, our carbon-fiber composite cell shows an 11 millivolt OCV drop at 50 percent state of charge. That sounds minor, but the LFP OCV plateau near 50 percent is nearly flat: a 10 mV window covers roughly 8 to 12 percent of usable capacity. An estimator that ignores the shift reports 50 percent when the cell is actually at 42 to 44 percent. The error recurs whenever the platform flexes.
Why Bending Changes the OCV Curve
In a conventional prismatic or pouch cell, the can or stack keeps the laminated electrodes flat under uniform compression. At a given SoC, open-circuit voltage reflects lithium intercalation stoichiometry at the cathode and anode surfaces, which remain chemically stable at that charge state. Modelers therefore treat temperature as the main environmental variable.
A structural cell bonds its electrode stack to the fiber plies within a carbon-fiber composite layer. As the member bends, layers on the tension side of the neutral axis stretch. The graphite anode experiences a slight increase in interlayer spacing, shifting the equilibrium intercalation potential by a measurable amount. The shift depends on strain level, fiber layup orientation, and the electrode thickness-to-composite-thickness ratio.
We measured the OCV shift in our standard 2-mm CFC layup under three four-point bending conditions. Across the mid-SoC plateau, 400 microstrain produces a 4 to 6 mV shift, 800 microstrain produces 9 to 13 mV, and 1200 microstrain reaches 18 to 22 mV. A vehicle floor panel commonly sees 600 to 900 microstrain during cornering and uneven-surface driving. Thus, 800 microstrain is an expected floor-panel operating condition, not an edge case.
How Conventional SoC Estimators Work
Production BMS designs generally use one of two methods: coulomb counting with periodic OCV recalibration during rest, or an extended Kalman filter, EKF, that combines current integration with an OCV model to limit drift. Both rely on the same premise: the OCV-SoC lookup table is fixed by cell chemistry and varies only with temperature.
That premise works for prismatic and cylindrical cells in conventional enclosures. It breaks down when the cell deforms as part of a load-bearing structure. During a rest period, the EKF compares terminal voltage, including the bending offset, with a lookup table that has no bending variable. The filter then converges on the wrong SoC.
Coulomb counting fails less visibly. During active operation it does not consult the OCV model, so cumulative current tracking remains accurate. Its rest-period recalibration still uses the flat-cell OCV table, however. If a vehicle parks while floor-panel bending remains under the parked weight distribution, the shifted resting OCV anchors recalibration incorrectly. Each event adds bias instead of removing accumulated drift.
SoC Error Over Operating Cycles
With flat cells, EKF error remains bounded because the OCV model is correct and the filter can self-correct. Under repeated structural loading, rest recalibration adds a systematic bias. When the bending shift is negative, cells on the tension side of a floor panel carrying vehicle weight make the estimator report higher SoC than the actual value.
Across 50 to 100 operating cycles, unmodified EKF implementations trained on flat-cell data show accumulated SoC error of 3 to 6 percent. That can understate usable range or, with conservative tuning, end charging and discharging too early. In a 60 kWh structural battery floor panel, 4 percent SoC error equals approximately 2.4 kWh of effective capacity lost to estimation uncertainty. For a 400 km range vehicle, that is roughly 16 km of phantom range loss.
A Strain-Conditioned OCV Model
We use a strain-conditioned OCV model. The BMS derives a strain proxy from differential voltage across adjacent cell groups: during constant-current charging under bending, compression-side cells charge slightly faster than tension-side cells. The difference is only 2 to 4 mV during 0.5C charge, but it is consistent enough to estimate bending without dedicated strain gauge sensors in the assembly.
We create a two-dimensional OCV lookup table with SoC on one axis and estimated bending index on the other. During its innovation step, the EKF uses this strain-conditioned table rather than the flat-cell table. Against cells fitted with fiber Bragg grating strain sensors as ground truth, the estimator keeps SoC error below 1.5 percent across the full operating envelope, including Korean highway and mixed urban driving cycles with continuously varying bending load.
The strain proxy adds no hardware. Per-cell voltage monitoring already provides the differential signal at 1000 samples per second. Two-dimensional interpolation in the EKF update raises computational cost by approximately 15 percent over the standard implementation, within the capacity of embedded BMS processors running at 240 MHz.
Where Conventional Estimators Still Work
The point is not that standard EKF implementations fail in every structural-cell application. In members exposed mainly to compression, such as a vertical post or compression strut, electrode deformation is more symmetric and the OCV shift is small enough for flat-cell calibration to remain within 1 percent. A thick composite skin relative to electrode stack thickness also attenuates transmitted strain, reducing the effect proportionally.
The issue is concentrated in thin structural members with a high aspect ratio, length-to-thickness greater than roughly 20:1, and bending-dominated load paths. EV floor panels, roof panels, and long beams in robotic limbs match this geometry. In those cases, strain-conditioned estimation is a correctness requirement rather than an optimization.
Integrating With OEM Platforms
When a new partner enters our early-access program, we begin with a combined characterization drive cycle and stress-OCV correlation test. Cells are mounted in the target structural geometry, fitted with reference strain gauges, and taken through a full SoC sweep under flat and loaded conditions. The result is a bending-index correction surface for that geometry and cell layup.
We load that surface into the BMS firmware as the platform's baseline calibration. A significant geometry change requires a new surface built from loaded OCV measurements. That is why we connect BMS software to the physical cell and structural design instead of providing geometry-agnostic firmware. The estimation algorithm and the cell's mechanical context cannot be separated cleanly.
Engineers porting an existing BMS to a structural-cell platform can carry over the EKF state equations and covariance tuning almost directly. The OCV table is different. Treat it as a new characterization task, not as a lookup-table replacement copied from a prismatic data sheet.