Cell grading is the last sorting step before cells enter pack assembly or secondary market channels. A-grade cells meet full spec and go into automotive or premium applications. B-grade cells — lower capacity, higher internal resistance, or wider capacity spread — are diverted to energy storage or lower-demand applications. C-grade cells are scrapped. The grade assignment drives revenue per cell and determines warranty liability downstream.
Most production grading systems are rules-based: measure three to five scalar metrics at EOL test, apply threshold tables, assign grade. Simple, auditable, fast. The problem is that the threshold tables were written for an expected defect profile, and formation-derived defects don't always behave the way the tables expect. When a cell type, chemistry revision, or protocol change shifts the defect distribution, the rules-based thresholds either over-grade (B-cells get A-grade assignments and create warranty exposure) or under-grade (A-cells get B-grade assignments and erode yield revenue). Both failure modes cost money.
How Grading Thresholds Get Set — and Why They Drift
The standard approach to setting grade boundaries for a new cell type: run a characterization campaign at slow rates under controlled conditions, establish the capacity and impedance distribution, set A/B cutoffs at chosen percentile thresholds (commonly 5th percentile for the A/B boundary), validate with accelerated aging tests on a sample from each grade bin. Ship.
This works well when the cell chemistry and formation protocol remain stable. In practice, formation protocols are revised more often than grade threshold tables are — new cycler software, C-rate adjustments for throughput, temperature setpoint optimization. Each revision can shift the population distribution without triggering a threshold review. The thresholds were calibrated for the old distribution. They're applied to the new one. The grade accuracy degrades silently.
A growing NMC-811 production line that changed its formation Step 1 C-rate from C/10 to C/8 for throughput reasons will see a shift in the mean first-cycle CE of roughly −0.2 to −0.4% and a widening of the capacity distribution tail. If the A-grade capacity threshold was set at 97.0% of nominal and the distribution mean shifts down, the B-grade fraction increases — not because cells are getting worse, but because the grading calibration is now stale. The opposite problem occurs when a process improvement (better temperature control, tighter electrolyte fill) shifts the distribution upward: B-grade cells that would have been A-grade under the new distribution continue to be downgraded until someone notices the anomaly.
The Chemistry-Specificity Problem
Threshold tables for NMC-622 don't apply to NMC-811 without recalibration, and neither applies to LFP without a fundamentally different grading logic. LFP cells have a very flat discharge voltage plateau between 3.2V and 3.3V — their capacity distribution is inherently tighter than NMC cells at equivalent manufacturing quality, which means the A/B boundary needs to be correspondingly tighter to be meaningful. LFP cells also have substantially lower first-cycle CE (typically 92–95% versus 97–99% for NMC) due to more extensive initial lithium consumption in SEI formation, so CE-based grading thresholds require chemistry-specific references.
When a factory adds a second chemistry — say, adding LFP to an existing NMC line for a stationary storage program — grading systems that don't handle chemistry-specific reference distributions will either apply NMC thresholds to LFP cells (generating incorrect grade assignments) or require manual threshold configuration updates for each chemistry switch. In a high-mix environment, that configuration burden is significant and error-prone.
We're not saying rules-based grading is wrong for a stable, single-chemistry line running a mature protocol — it's appropriate for that context, and its simplicity is a genuine advantage for process documentation. We're saying it degrades gracefully as process complexity increases, and most growing lines are adding complexity faster than their grading configuration management can accommodate.
What Formation Signatures Add to Grade Classification
The insight that formation-aware grading enables is this: the formation telemetry predicts long-term performance better than the EOL scalar measurements that most grading systems are built on. This isn't speculative — it follows from the fact that formation is where the electrochemical structure of the cell is established, and the formation signature contains direct information about that structure.
Specifically, three formation-derived features have strong predictive value for grading precision:
dQ/dV Feature Consistency
The ratio and relative positions of dQ/dV peaks across the formation cycle set are correlated with active material utilization and electrode balance. Cells with anomalous dQ/dV feature ratios — even if their C/5 capacity is within the A-grade threshold — show accelerated capacity fade at high C-rates in the 100–300 cycle range. Including this feature in grade classification reduces false A-grade assignments for cells that will underperform in high-demand applications.
Internal Resistance Evolution Slope
The rate at which cell IR changes between formation cycle 1 and cycle 3 (for a 3-cycle formation protocol) is diagnostic. Cells whose IR decreases sharply between cycles 1 and 3 — indicating rapid SEI stabilization — have better high-rate performance retention than cells whose IR changes minimally, which often indicates an improperly formed SEI that will continue evolving under service conditions.
IR_slope = (IR_cycle3 - IR_cycle1) / IR_cycle1
Nominal range (NMC-622, 25°C): -0.08 to -0.04
High-risk range (underfilled or low-temp formation): -0.02 to +0.01Population-Relative Anomaly Score
A cell that sits at the 4th percentile of its lot CE distribution is a different risk profile than a cell at the 4th percentile of a lot whose mean CE is 0.4% higher than the reference. Absolute thresholds miss this context. A population-relative scoring approach — flagging cells that deviate anomalously from their own formation lot distribution, not just from a global threshold — captures the relative risk more accurately and adapts automatically to normal lot-to-lot variation.
Grade Classification as a Learning System
The practical challenge with formation-aware grading is the training data problem: to learn the relationship between formation features and long-term performance, you need matched pairs of formation telemetry and field cycle data. That matched dataset takes 12–18 months to accumulate on a production line with typical automotive qualification cycles.
There are two strategies that work before the matched dataset is mature. First: use accelerated aging tests (45–60°C, elevated C-rate) on matched formation-feature cohorts to compress the cycle life validation timeline. Second: use anomaly detection on formation features — flagging cells that are statistical outliers in the feature space — as a proxy for risk, without requiring a labeled training set. The second approach doesn't give you precise grade boundaries, but it does identify cells that warrant diversion to secondary applications rather than primary automotive use.
As the field return dataset matures, the model transitions from anomaly-detection-based risk scoring to a trained classifier with validated performance boundaries by chemistry and application type. That transition typically takes 18–24 months from line start on a new cell format — faster if the line has accumulated a large volume of accelerated aging data or if there is a comparable reference dataset from an earlier cell generation.
The line that starts collecting and structuring formation telemetry for this purpose on day one of production ramp has a fundamentally different data asset at 18 months than the line that starts thinking about this when the first warranty cohort comes back from the field.