Fuel Cells

PEM fuel cell manufacturing is bottlenecked by catalyst-coated-membrane defect rates. A 4-8% missing-catalyst defect area can cause a cell to reach end-of-life within 50 hours of accelerated stress testing. The detection gap is not sensor capability, it is data plumbing between IR thermography, optical line-scan, slot-die telemetry, and stack performance systems that have no shared cell identity.

The CCM Manufacturing Chain

Catalyst ink, platinum on carbon (Pt/C) dispersed with Nafion ionomer and solvent, is coated onto either the membrane directly (CCM route) or a decal substrate for subsequent transfer. Wet thickness targets 10-20 μm. Platinum loading ranges from 0.05 mg/cm² at the anode to 0.4 mg/cm² at the cathode, with tighter loading tolerance at the cathode because Pt loading directly controls activity and durability. Drying at 60-80 °C removes solvent while preserving ionomer morphology.

Hot pressing at 130-150 °C under 2-4 MPa laminates the gas diffusion layer (GDL) to the CCM. GDL fiber penetration into the catalyst layer, caused by excessive hot-press pressure or GDL misregistration, is a common defect that creates local current-density hotspots during operation. Gasket framing, bipolar plate assembly, and stack assembly complete the manufacturing chain.

CCM Defect Classes and Their Detection Requirements

Catalyst-layer pinholes and missing-catalyst defects (MCLD) are the primary performance killers. A missing-catalyst zone as small as 4-8% of electrode area creates a region of locally high membrane resistance that generates heat under current load, accelerating membrane thinning. IR thermography detects platinum-loading variations at millimeter scale with approximately 1 second temporal resolution, it sees the loading deficit but cannot resolve geometric defects below the thermal diffusion length. Optical line-scan detects pinholes, cracks, and edge defects at geometric resolution but cannot directly measure catalyst loading. Neither modality alone provides complete coverage.

Reactive-flow IR, where the cell is operated under controlled hydrogen and air flow while IR images are captured, detects electrode cracks down to 200 μm by imaging the local heat generation pattern. Roll-to-roll inline integration of reactive-flow IR has been demonstrated, enabling configurable web coverage at line speed for crack detection in addition to the loading measurement provided by standard IR thermography.

Ionomer-rich top layers form from over-fast drying, the same mechanism as binder migration in Li-ion electrodes, creating an ionically resistive surface film. GDL fiber penetration at the hot-press stage creates the local resistance hotspots mentioned above. Membrane thinning from either catalyst-layer defects or GDL penetration propagates to premature cross-over failure during operation.

Where Fuel Cell Programs Get This Wrong

The core failure mode is the six-database problem. IR thermography images live in the thermography system. Optical line-scan images live in the vision system. Slot-die telemetry lives in the coater controller. Hot-press parameters live in the press controller. Stack performance data lives in the testing system. Accelerated stress test results live in yet another system. No cell-level genealogy connects them.

The lag between catalyst-coating drift and its appearance in polarisation data is 12-72 hours depending on the inspection schedule and the stack assembly queue. An engineer who wants to understand why a stack lot underperformed must manually reconstruct the timeline across all six systems, identifying which CCM lots fed which stacks, what the IR loading maps looked like for those lots, and whether the slot-die telemetry showed a pressure anomaly that day. Manual RCA at that scale takes days to weeks and relies on individual engineers remembering to check each system. The correlation that would have been obvious in a unified view is invisible in six separate databases.

What Dual-Modality AI Changes

Fusing IR thermography and optical line-scan into a single inspection pass means each defect is characterized by both its loading signature (from IR) and its geometric signature (from optical), and neither modality's blind spots apply to the fused output. A missing-catalyst zone that appears as a loading void in IR and shows no geometric anomaly in optical is still caught. A mechanical pinhole that appears in optical and shows no loading variation in IR is still caught.

The deeper value is cell genealogy. When slot-die telemetry, IR loading maps, optical inspection frames, hot-press logs, and stack polarisation data are unified under a cell-level ID, the question, which coating-process parameters predict polarisation underperformance? , becomes answerable in minutes. New defect types are catalogued automatically with process context and outcome, building the institutional dataset that fuel cell manufacturing programs need to compress the yield learning curve from years to months.

The same correlation architecture applies to SOFC manufacturing, where the dimensional tolerances on electrolyte tape-cast sheets and the sintering profiles are analogous process control problems to PEM CCM coating.

References

1. 1. Aieta, N.V., et al. (2012). Applying infrared thermography to study the effects of membrane electrode assembly defects on proton exchange membrane fuel cell performance. Journal of Power Sources, 211, 4-11. https://doi.org/10.1016/j.jpowsour.2012.02.030
2. 2. Das, P.K., et al. (2014). Dominant role of electrode-related factors in the performance of proton exchange membrane fuel cells. Journal of Power Sources, 261, 401-411. https://doi.org/10.1016/j.jpowsour.2013.12.075
3. 3. Ulsh, M., et al. (2016). Roll-to-roll inline defect detection for fuel cell manufacturing. Fuel Cells, 16(2), 170-178. https://doi.org/10.1002/fuce.201500137
4. 4. White, R.T., et al. (2024). Bridging the gap: in-line quality control for battery manufacturing. Frontiers in Manufacturing Technology. https://doi.org/10.3389/fmtec.2024.1392038