Battery Formation: A Complete Process Guide for EV Manufacturers

What is formation cycling?

Formation cycling is the first charge–discharge sequence applied to a freshly assembled lithium-ion cell. It is not simply "the first charge." It is an electrochemical activation process that determines the long-term performance, safety, and lifecycle characteristics of the cell.

During formation, the electrode surfaces are exposed to electrolyte for the first time at operating voltage. At the anode (typically graphite), lithium ions intercalate while the electrolyte simultaneously decomposes to form a passive film — the solid electrolyte interphase (SEI). At the cathode (NMC, LFP, or other lithium-metal-oxide), a similar, thinner interphase (cathode electrolyte interphase, CEI) forms.

The formation protocol — charge rate (C-rate), cutoff voltages, number of steps, rest periods, temperature — directly controls the uniformity and composition of these films. A formation protocol optimized for a given cell chemistry and electrolyte formulation produces a stable, dense SEI with low internal resistance and predictable capacity. A poorly controlled formation produces an uneven SEI, elevated coulombic inefficiency, and early-life capacity fade.

Formation typically runs 6–24 hours per cell, occupying the most space and the highest proportion of capital equipment on the manufacturing line. For a mid-size gigafactory producing 5–20 GWh/year, formation racks can occupy 40–60% of total floor space and represent a significant fraction of the total equipment cost.

SEI layer formation

The SEI layer is a nanometer-scale film of lithium-containing compounds (Li₂CO₃, LiF, lithium alkyl carbonates, and others depending on the electrolyte chemistry) that forms at the graphite anode surface when the cell voltage first reaches the electrolyte reduction potential — typically around 0.8 V vs. Li/Li⁺.

A good SEI is:

• Electronically insulating — stops further electrolyte reduction after the initial formation
• Ionically conductive — allows lithium ions to pass during subsequent charge–discharge cycles
• Mechanically stable — resists cracking during the volumetric expansion of graphite during lithiation (up to 10% in crystallographic c-axis)
• Uniform across the anode surface — prevents local hotspots that lead to lithium plating

Defects in SEI formation are the leading cause of early-life capacity fade and safety incidents. The three most common SEI-related defects seen in production are: (1) dendritic lithium plating at the anode surface from excessive current density or insufficient rest time, (2) thick/porous SEI from electrolyte impurities or moisture ingress, and (3) non-uniform SEI from electrolyte fill head misalignment or temperature gradients in the formation chamber.

Telemetry signals and quality encoding

Modern formation cyclers continuously log voltage (V), current (A), and temperature (°C) at sampling rates of 100 ms to 1 s. For a standard 32-channel rack operating a 4-hour formation cycle at 500 ms resolution, this generates approximately 920,000 data points per cell per cycle.

The key insight that drives Moldpathio's approach: cell quality is not measured in a single scalar at the end of formation. It is encoded in the shape of the voltage–capacity curve throughout the formation cycle. The first charge curve of a lithium-ion cell is a waveform with distinct features corresponding to different electrochemical processes:

• The plateau at ~0.8 V corresponds to initial SEI formation on graphite (electrolyte reduction)
• The staging plateaus at 0.21 V, 0.11 V, and 0.08 V correspond to successive lithium staging in graphite (from dilute solid solution to LiC₁₈ to LiC₁₂ to LiC₆)
• The high-voltage tail (above 4.0 V vs. Li/Li⁺ full cell) indicates cathode delithiation and CEI formation

Deviations from the expected waveform shape — peak shifts, missing plateaus, anomalous capacity values at specific voltage steps — are the fingerprints of defect conditions.

dQ/dV analysis

The voltage-capacity differential (dQ/dV) transforms the raw voltage-capacity curve into a spectrum of electrochemical events. Peaks in the dQ/dV curve correspond to phase transitions and intercalation events in the electrode materials. This representation amplifies subtle deviations that are nearly invisible in the raw V vs. Q curve.

For a NMC622/graphite cell undergoing initial formation, the dQ/dV spectrum typically shows:

• A broad feature between 0.7–0.9 V associated with initial SEI formation and electrolyte decomposition
• Peaks at approximately 0.21, 0.10, and 0.08 V (graphite staging transitions)
• Features at 3.7 V and 4.0 V corresponding to NMC phase transitions

The diagnostic value of dQ/dV is in peak positions, peak heights, peak widths, and the ratio between peaks. For example:

• A right-shifted SEI peak (0.9–1.0 V vs. expected 0.75 V) suggests excessive electrolyte moisture
• A suppressed or absent graphite staging peak at 0.10 V indicates lithium plating blocking intercalation sites
• Asymmetric cathode peaks indicate non-uniform cathode utilization, often from electrolyte fill non-uniformity

Common defect types and signatures

Based on formation data from a large number of pilot deployments, Moldpathio has identified six formation-stage defect categories by frequency and economic impact:

1. Electrolyte fill non-uniformity — Wetting front non-uniformity causes differential SEI formation and capacity gradient across the cell cross-section. dQ/dV signature: broad, asymmetric staging peaks; elevated capacity spread within a tray.
2. Lithium plating — Metallic lithium deposits at the anode surface when local current density exceeds the lithium intercalation rate. Risk factors: high C-rate in early formation steps, low temperature, poor electrolyte wetting. dQ/dV signature: suppressed stage-1 peak, additional features at 0.0–0.1 V, asymmetric stripping peak in subsequent discharge.
3. Electrolyte moisture contamination — Water reacts with LiPF₆ to form HF, which attacks the cathode surface and creates a thick, resistive SEI. dQ/dV signature: shifted SEI peak position, elevated first-cycle irreversible capacity.
4. Temperature gradient in formation chamber — Cells at the edge of a rack run colder than cells at the center. colder formation produces different SEI chemistry. dQ/dV signature: tray position-correlated capacity spread; characteristic pattern tied to physical position in rack.
5. Calendar self-discharge from pin-to-pin leakage — Micro short circuits or external conduction paths cause voltage drop during rest periods. Signature: anomalous open-circuit voltage decay during formation rest steps.
6. Formation protocol deviation — Controller failures, cycler calibration drift, or incorrect recipe loading cause cells to experience wrong voltages or C-rates. Signature: anomalous plateau structure; voltage steps at wrong positions relative to capacity.

Why SPC has limits here

Statistical process control is designed for processes where output quality can be captured in a scalar measurement (part dimension, coating weight, film thickness). The control structure is: measure the scalar, compute control limits based on historical mean and variance, trigger investigation when readings fall outside limits.

Formation cycling is different in three ways that make standard SPC insufficient:

1. Quality is multi-dimensional. A cell's formation quality is encoded in a waveform with thousands of features, not a single number. SPC on end-of-formation capacity alone misses defects that manifest as shape deviations rather than capacity loss.
2. Normal variation is non-stationary. The "expected" formation waveform changes with electrode chemistry, electrolyte lot, ambient conditions, and process recipe. SPC control limits set on a single chemistry cannot correctly flag deviations when process conditions shift.
3. Defect detection latency. SPC triggers only after the formation cycle is complete — 4–8 hours after formation began. At that point, defective cells have been through a full formation run, and the cell may need to be re-formed (if possible) or scrapped. ML inference during formation (at T+40 minutes) can trigger intervention before the full protocol completes.

ML approaches to formation quality

Machine learning models for formation quality prediction take the voltage-time-current time series as input and output either a quality score (regression) or a defect class prediction (classification). The feature engineering problem is central: raw 100ms-resolution time series are high-dimensional and chemistry-specific. Effective approaches extract physically meaningful features aligned with the electrochemical processes described above.

Moldpathio's inference pipeline uses a combination of:

• Sliding window dQ/dV feature extraction (computed online during formation)
• Temporal convolutional network (TCN) for sequence-level pattern recognition
• Physics-informed feature vectors derived from staging transition positions
• Cross-channel correlation features (tray-level analysis to detect rack-position effects)

Models are calibrated per customer per cell chemistry. A model trained on NMC622/graphite cells is not directly applicable to NMC811/SiOx cells without recalibration — the staging transition positions and SEI formation characteristics are different. Moldpathio's onboarding process includes a 2-week model calibration phase to adapt the base model to the customer's specific electrode chemistry and formation protocol.

Deployment considerations

Deploying an ML-based formation monitoring system in a production environment requires integration at the data layer, the inference layer, and the alerting layer.

Data layer: Formation cycler telemetry is typically accessible via the cycler's data export API or historian interface. Moldpathio's edge connector software integrates with major cycler brands (Maccor, BioLogic, Arbin, Neware) and the MES layer. Data is streamed in real time from the channel controller to the edge compute node. No modifications to the cycler hardware or firmware are required.

Inference layer: The ML inference engine runs on an edge compute node (Moldpathio-supplied or customer-supplied) co-located with the formation racks. This keeps raw formation data on-site and avoids round-trip latency to a cloud inference endpoint. Inference is performed continuously during formation cycling, with quality scores updated every 5 minutes per active channel.

Alerting layer: Quality alerts are surfaced via the Moldpathio dashboard, webhook to the MES, or direct SMS/email notification. Alerts include tray ID, channel position, defect classification, confidence score, and the specific waveform features that triggered the flag. The alert taxonomy is designed to translate directly to line operator actions — not "anomaly detected" but "lithium plating risk on TRAY-07, CHANNEL-03, formation step 2, confidence 87%."