What happened to the Boeing 787 Dreamliner APU batteries in January 2013, and what does it reveal about the regulatory gap for manufacturing quality AI?

The Boeing 787 Dreamliner uses lithium-ion batteries (LiCoO₂/graphite chemistry, 8 prismatic cells connected in series, 32V nominal, 63Ah capacity) supplied by GS Yuasa Corporation of Japan for two primary applications: the main battery, which powers essential aircraft systems on ground power and during initial APU start, and the APU (auxiliary power unit) battery, which powers the APU electronic control unit and APU starter motor. On 7 January 2013, a Japan Airlines (JAL) 787 (JA829J) parked at Gate E9 at Boston Logan International Airport experienced an APU battery fire. Ground crew observed smoke from an electronics bay access door. The Boston Fire Department responded. Investigators found that the APU battery had experienced thermal runaway in one cell, which propagated to adjacent cells, producing electrolyte venting and ignition. Electrolyte fires within the battery enclosure burned for approximately 45 minutes before being extinguished. On 16 January 2013, an All Nippon Airways (ANA) 787 (JA804A) operating as NH692 from Yamaguchi-Ube Airport to Tokyo Haneda experienced smoke and fumes in the forward cabin and cockpit. The crew declared an emergency and made an emergency landing at Takamatsu Airport. Passengers evacuated using emergency slides. Investigation found that the main battery had experienced thermal runaway, with electrolyte vapour migrating from the battery compartment into the aircraft cabin. The NTSB (which led the investigation of the JAL Boston event) concluded that the initiating cell failure was an internal short circuit, most likely caused by manufacturing defects including lithium plating within the cells — consistent with localised electrode region defects that concentrate lithium deposition during cycling. The FAA issued Emergency Airworthiness Directive 2013-02-51 on 16 January 2013, grounding all Boeing 787 aircraft worldwide pending a corrective action plan. At the time of the grounding, approximately 50 787s were in service with 8 airlines (United, JAL, ANA, Ethiopian Airlines, Air India, LOT Polish Airlines, LAN Chile, Qatar Airways). The 787 remained grounded for approximately four months — from January to April 2013 — while Boeing, GS Yuasa, the FAA, and JCAB worked through an approved battery improvement plan (improved cell isolation, improved battery enclosure with steel casing, venting system changes). The Boeing 787 APU battery events are significant for the electrode coating AI adversarial injection threat model for two reasons. First, they document that lithium plating — the first consequence of an electrode thin zone or over-calendering defect that reduces Li-ion intercalation capacity — was identified as the most probable initiating failure mechanism in a deployed aviation system, causing a 4-month worldwide fleet grounding at enormous economic cost to Boeing and airlines. Second, they established the manufacturing quality investigation precedent that led the NTSB and FAA to examine GS Yuasa's cell manufacturing process controls in depth — including electrode coating thickness measurement, calendering density control, and formation cycling inspection — without any of those quality inspection systems being evaluated for adversarial robustness because the regulatory framework had no such requirement. The adversarial injection threat model that targets those same inspection AI systems exploits the same regulatory gap.

How does a thin zone in the XRF electrode coating weight heatmap lead to Li-plating, dendrite growth, and internal short circuit — what is the progressive failure pathway?

A thin zone in the electrode coating is defined as a localised region of the electrode where the active material coating weight (areal density, in mg/cm²) falls below the lower process control limit (LCL). For a typical NMC622 (Li(Ni₀.₆Mn₀.₂Co₀.₂)O₂) cathode with a nominal coating weight of 200 mg/cm² and a process control tolerance of ±3%, the LCL is approximately 194 mg/cm². For a graphite anode (nominal 100 mg/cm² at a typical N/P ratio of 1.05–1.10), the LCL is approximately 97 mg/cm². A thin zone below the LCL means that the anode has fewer graphite intercalation sites in that region — fewer carbon layers available to host Li-ions as Li⁺ during charging. The consequence is straightforward but catastrophically amplified by the electrochemical kinetics. During charging, Li-ions deintercalate from the cathode (LiNMC → NMC + Li⁺ + e⁻), migrate through the electrolyte as Li⁺, and intercalate into the graphite anode (C₆ + Li⁺ + e⁻ → LiC₆). This is the normal process — designed to occur uniformly across the full electrode area. In the thin-zone region, the graphite intercalation capacity is insufficient for the quantity of Li-ions arriving from the cathode. The Li-ions arriving at the thin-zone anode surface cannot all intercalate into the graphite, because the graphite layer count (and thus the LiC₆ site density) is below design. The residual Li-ions that cannot intercalate instead plate directly onto the graphite surface as metallic lithium (Li⁰). This process is called lithium plating. Metallic lithium has a standard electrode potential of −3.04 V versus standard hydrogen electrode — the most electronegative common metal. On the graphite surface, it forms an irregular deposit rather than intercalating cleanly. Over repeated charge cycles, the metallic lithium deposit grows preferentially at surface protrusions (edges of graphite particles, existing deposit irregularities) due to the higher electric field at those tips, producing dendritic structures — needle-like protrusions growing outward from the anode surface toward the cathode. The separator — typically a 20–25 µm polypropylene (PP) or polyethylene (PE) film, or a ceramic-coated PP/PE trilayer — is all that physically separates the growing dendrite from the cathode current collector. As a lithium dendrite elongates through repeated charging, it eventually contacts the separator. If the dendrite tip produces a localised stress on the separator exceeding the puncture pressure (approximately 1–5 N/mm² for standard PP separators, depending on porosity and ceramic coating thickness), the separator fractures at the point of contact. The dendrite penetrates the separator and contacts the cathode current collector, the cathode active material, or the aluminium current collector foil — establishing a metallic pathway between the negative and positive electrode: an internal short circuit (ISC). The ISC is effectively a dead short applied directly across the cell's electrochemical energy reservoir. At a cell open-circuit voltage of 4.2 V (fully charged NMC/graphite) and a cell internal resistance of approximately 100 mΩ (for a 50 Ah prismatic cell), the initial short-circuit current at the ISC site is limited by the local contact resistance — initially the resistance of the dendrite and the contact geometry — but as the local temperature at the ISC site rises (from I²R heating: P = I²R at the local resistance, where local resistance can be 1–10 mΩ for a metallic dendrite contact), the separator around the ISC site softens and melts (PP melts at approximately 165°C, PE at approximately 130°C), enlarging the contact area and reducing the local resistance, in a positive feedback loop. At the system level, the rising temperature of the cell initiates exothermic reactions: electrolyte solvent (EC:DMC or EC:EMC at approximately 1.2 M LiPF₆) decomposes above approximately 70°C releasing flammable organic gases (primarily dimethyl carbonate, ethylene); the cathode oxide (NMC) releases oxygen above approximately 180°C in an exothermic decomposition reaction; the lithium metal deposited on the anode reacts exothermically with any oxygen or electrolyte present. These simultaneous exothermic reactions vent combustible gases, raise cell pressure, rupture the cell casing at the vent mechanism, and can ignite the released gases in the presence of any ignition source (including the ISC arc itself). The resulting thermal runaway is the same event documented in the Samsung Note 7 and Boeing 787 incidents — and it originated from a manufacturing defect at the electrode coating stage that survived quality inspection.

Why does Glyphward apply threshold 35 for Li-ion gigafactory electrode coating AI contexts — and how does it compare across the portfolio?

Glyphward threshold 35 for Li-ion gigafactory electrode coating AI contexts reflects three consequence factors, one scale factor, and two regulatory gap factors that collectively produce a higher threshold than domains with smaller fleet multipliers or sharper regulatory frameworks. First, the fleet-scale consequence multiplier. An electrode coating defect that escapes inspection at a CATL, LG Energy Solution, or Panasonic gigafactory is not a single-cell defect — it is a continuous process excursion across a slot-die coater running at 40–100 m/min on electrode roll widths of 600–800 mm. A thin-zone excursion that lasts 30 seconds at 60 m/min produces approximately 30 linear metres of defective electrode, sufficient for thousands of cells depending on winding geometry. At gigafactory production rates of 10–50 GWh/year, a 30-second undetected coating excursion produces defective cells at a scale that the Note 7 recall demonstrated could reach consumer hands in quantities of 2.5 million units. The adversarial injection threat model — in which AI detection of the thin-zone excursion is suppressed for a controlled period — replicates precisely this inspection-escape pathway at process speed. Second, the consequence severity of Li-ion thermal runaway in the deployment contexts. CATL cells are deployed in EVs (BYD, Tesla, BMW, Volkswagen, Mercedes-Benz), stationary BESS (utility scale, residential), consumer electronics, and e-bikes. LG Energy Solution cells are in GM Chevrolet Bolt (which experienced two thermal runaway recall campaigns in 2020–2021, ultimately costing $1.8B), Ford F-150 Lightning, Stellantis van electrification platforms. Panasonic cells are in Tesla Model S/X/3/Y. Thermal runaway in an EV can produce temperatures of 500–1,000°C in the battery pack, releasing toxic combustion gases (HF from LiPF₆ electrolyte decomposition, CO from incomplete organic combustion, HCN from nylon separator or plastic component combustion), and presenting extreme difficulties for firefighting (water reacts with LiPF₆ electrolyte to produce HF; submerged submersion is the standard response, requiring thousands of litres). Third, the aviation consequence from Boeing 787. The FAA Emergency AD 2013-02-51 establishing a 4-month worldwide fleet grounding of all 787s — a consequence triggered by Li-ion manufacturing quality failures — is the documented aviation consequence envelope for the battery inspection adversarial injection threat model. Threshold 35 is above threshold 30 applied to nuclear power plant digital I&C AI. This comparison may seem counter-intuitive given the severity of nuclear consequence anchors (TMI-2, Fukushima). The distinction is that the nuclear domain has the NRC GDC 20–24 single-failure criterion — an explicit regulatory requirement that no single component failure shall prevent a safety function — with no equivalent in IEC 62619. The nuclear safety system design is explicitly built to tolerate a single failure, including (in principle) a monitoring system failure. The Li-ion gigafactory electrode inspection AI is the sole automated real-time quality gate for a thin-zone excursion at process speed; no redundant independent inspection system operates in parallel at the same spatial resolution and temporal frequency. The sole-barrier architecture and fleet-scale consequence multiplier together justify threshold 35. Threshold 35 matches the hydrogen electrolysis UV flame detection threshold and the oil refinery APC raffinate splitter level AI threshold — two other contexts where the inspected system has a sole-barrier AI architecture and the consequence of an undetected excursion can produce mass casualties or large-scale deployment recalls at fleet scale.

Battery manufacturing AI security · CATL electrode coating AI · LG Energy Solution AI · Panasonic Gigafactory AI · KLA SURFmonitor · IEC 62619:2022 · EU Battery Regulation 2023/1542 · Samsung Note 7 2016 · Boeing 787 APU 2013 · XRF coating weight AI · Glyphward threshold 35

Li-ion gigafactory electrode coating AI adversarial injection: how ±6 DN in the XRF coating weight heatmap suppresses a thin-zone precursor to Li-plating, dendrite growth, and internal short circuit — and why IEC 62619:2022 has no adversarial robustness criterion for the CATL/LG Energy/Panasonic electrode inspection AI layer

Published 20 June 2026 · 3,600 words · Back to blog

In September 2016, Samsung announced a worldwide recall of 2.5 million Galaxy Note 7 smartphones. The root cause was a manufacturing defect at the electrode level: the graphite anode in Samsung SDI cells had insufficient clearance in corner regions, and the aluminium current collector foil in ATL replacement cells had metallic burrs that penetrated the separator. Both defects produced internal short circuits. The consequence: fires on aircraft, flights diverted, two separate recalls, Samsung permanently discontinuing a flagship device, the FAA banning the Note 7 from every aircraft in the United States, and approximately $17 billion in direct and indirect costs. In January 2013, a Japan Airlines 787 APU battery caught fire at Boston Logan. Nine days later, an ANA 787 made an emergency landing in Japan after battery smoke entered the cabin. GS Yuasa’s cell manufacturing investigation identified lithium plating — the hallmark of anode capacity shortage from electrode coating defects — as the most probable initiating mechanism. The FAA issued Emergency Airworthiness Directive 2013-02-51 and grounded every Boeing 787 in the world for four months. Today, CATL, LG Energy Solution, Panasonic Energy, Samsung SDI, and SK On deploy AI-based electrode inspection systems at every stage of the coating line — KLA SURFmonitor, Manz AG Coating Quality Analysis, Cognex electrode inspection AI, Teledyne Dalsa linescan AI — to detect exactly these defects before they enter the cell assembly stack. A ±6 DN adversarial pixel shift in the rendered XRF coating weight heatmap image suppresses a thin-zone defect from below the lower process control limit to within the normal operating band. The electrode inspection AI classifies the defective roll segment as conforming. The thin zone proceeds through calendering, slitting, and cell assembly into the wound jellyroll — and into the deployed fleet. IEC 62619:2022, UL 9540A, UN 38.3, and EU Battery Regulation 2023/1542 define the Li-ion battery safety and manufacturing quality framework — none include an adversarial robustness criterion for the AI inspection layer that determines which electrode coils are conforming. Glyphward threshold 35.

How Li-ion gigafactory electrode coating AI works — and where the adversarial injection surface lives

A lithium-ion cell is built from two electrode tapes — a positive (cathode) and a negative (anode) — separated by a microporous polymer separator and wetted with a lithium salt electrolyte. The cathode tape consists of aluminium foil (typically 12–20 µm thick) coated on both sides with a slurry of active material (NMC, LFP, NCA, or LCO oxide particles), carbon black conductive additive, and PVDF binder in an N-methyl-2-pyrrolidone (NMP) solvent. The anode tape consists of copper foil (typically 8–12 µm thick) coated with a slurry of graphite particles, carbon black, styrene-butadiene rubber (SBR) binder, and carboxymethyl cellulose (CMC) in water. Coating is performed by a slot-die coater — a precision metering device that extrudes the slurry through a shaped slot die onto the moving foil at controlled flow rate and line speed (typically 40–100 m/min in modern GWh-scale facilities). The wet electrode passes through a multi-zone drying oven, typically 40–100 m long, operating at 80–150°C to evaporate the solvent while maintaining binder distribution. The dried electrode is then compressed by precision calendering rolls to the target density and thickness, and finally slit to the cell-width specifications by rotating blade or laser slitters.

At each of these process steps, AI-based inspection systems classify rendered images of sensor outputs to detect defects that would produce non-conforming cells: (1) the XRF (X-ray fluorescence) coating weight measurement system, which produces a false-colour heatmap of active material areal density (in mg/cm²) across the electrode width at sub-mm spatial resolution; (2) the NIR (near-infrared) spectral imaging system, which measures the spatial distribution of PVDF binder and carbon black across the electrode width after drying; (3) the micrometer or beta gauge density profiling system, which measures electrode thickness (and thus porosity, inferred from coating weight and density) after calendering; and (4) the linescan camera system, which monitors electrode edge quality after slitting, detecting metallic burrs on the current collector foil edge. AI classification of these rendered measurement images is the primary automated real-time quality gate at each process step — the same structural architecture present in every gigafactory electrode manufacturing context: accurate physical sensors measure the defect-relevant parameter; the measurement is rendered as a 2D visualisation (heatmap, spectral image, thickness profile, linescan frame); the AI classifier provides the automated conformity determination and alarm trigger.

The adversarial injection surface is the boundary between each rendered inspection image and the AI classifier that processes it. The physical sensors produce correct measurements of the electrode. The rendering pipeline converts those measurements to 2D images for display and AI classification. A pixel-level perturbation at the rendered image ingestion boundary — within the normal photometric noise floor of the imaging system — can suppress the defect signature from the detection threshold to within the normal operating band, without altering the underlying sensor measurement and without detection by any instrument monitoring the measurement channel itself.

Samsung Note 7 2016 and Boeing 787 APU 2013: the Li-ion manufacturing defect consequence envelopes

The Samsung Galaxy Note 7 events of September–October 2016 are the primary consequence envelope for the electrode coating AI adversarial injection threat model. Samsung announced the first voluntary recall of approximately 2.5 million Note 7 units on 2 September 2016, citing reports of batteries catching fire during charging. The post-recall investigation identified the root cause in Samsung SDI cells as an electrode geometry tolerance violation in the jellyroll winding: the negative electrode (graphite anode) tape had insufficient overhang — a too-narrow margin beyond the positive electrode boundary in the jellyroll corner regions — which allowed the cathode edge to contact the anode copper current collector or the separator in the compressed winding, creating an internal short circuit under the compressive stress of the case assembly and during cycling expansion. Samsung deployed a replacement programme using cells sourced from Amperex Technology Limited (ATL). The ATL-supplied replacement cells began failing within weeks, through a different mechanism: metallic burrs on the positive electrode aluminium current collector foil edge, created by improperly maintained slitting blades, had penetrated the polyethylene separator during winding, establishing a second-source internal short circuit pathway. Samsung permanently discontinued the Galaxy Note 7 on 11 October 2016. The US Consumer Product Safety Commission formal recall was issued in September 2016. The FAA Emergency Order effective 14 October 2016 prohibited any Samsung Galaxy Note 7 from being carried on any aircraft — as a carry-on item, in checked baggage, or as cargo — one of very few device-specific aviation bans in US history. Airlines worldwide (United, Delta, American, Southwest, Qantas, Emirates, Air France, Lufthansa, Singapore Airlines, Cathay Pacific) implemented equivalent or stricter bans. Total estimated Samsung costs range from $5.3 billion in direct recall and discontinuation charges to approximately $17 billion when including estimated lost revenue from the Note 7 product line and impact on Galaxy S series sales in the following quarters.

The Boeing 787 Dreamliner APU battery events of January 2013 established a second consequence envelope at aviation scale. The 787 uses lithium cobalt oxide (LCO) / graphite cells in a GS Yuasa APU battery pack — eight prismatic cells connected in series, producing approximately 32 V nominal at 63 Ah capacity. On 7 January 2013, a Japan Airlines 787 (JA829J) parked at Gate E9 at Boston Logan International Airport experienced an APU battery event: ground crew detected smoke, and the Boston Fire Department found that the APU battery had experienced thermal runaway in one cell, with propagation to adjacent cells, producing electrolyte venting and ignition. On 16 January 2013, an All Nippon Airways 787 (JA804A) operating NH692 diverted to Takamatsu Airport after cockpit crew detected smoke and fumes; passengers evacuated using emergency slides. The NTSB investigation of the JAL Boston event — the most detailed technical examination — concluded that the initiating event was an internal short circuit in one cell. The most probable mechanism, consistent with the physical evidence of lithium deposits on the anode, was lithium plating — metallic lithium deposited on the graphite anode surface rather than intercalated — arising from localised anode capacity limitations in the affected cell region. Lithium plating is the electrochemical signature of insufficient anode intercalation capacity relative to the incoming Li-ion flux during charging: the same mechanism that initiates in a thin-zone electrode region where anode coating weight fell below the lower process control limit during manufacturing. The FAA issued Emergency Airworthiness Directive 2013-02-51 on 16 January 2013, grounding all Boeing 787 aircraft worldwide. Approximately 50 787s were in revenue service at the time, operated by United, JAL, ANA, Ethiopian Airlines, Air India, LOT Polish Airlines, LAN Chile, and Qatar Airways. The global fleet remained grounded for approximately four months, with the first 787s returning to service in late April 2013 after Boeing, GS Yuasa, and the FAA agreed on a battery improvement plan: improved cell design (increased separator thickness, reduced electrode packing density in corner regions), additional cell isolation (each cell encased in a stainless steel canister), battery enclosure redesign (steel casing, external venting to overboard). The 787 grounding is the documented aviation-scale consequence of a Li-ion manufacturing quality failure — where lithium plating, a symptom of electrode coating weight or capacity distribution defects, was identified as the initiating mechanism.

Both events share a common manufacturing inspection escape pathway: a defect at the electrode level — dimensional tolerance violation, metallic burr, localised anode capacity shortage — survived the quality inspection process and was incorporated into shipped cells. The adversarial injection threat model for electrode coating AI replicates this pathway by suppressing the AI detection that would have caught the defect, producing inspection escape at process speed across every cell manufactured during the suppression window.

Four adversarial injection surfaces: XRF coating weight AI, NIR binder composition AI, calendering density AI, and slitting linescan burr AI

XRF coating weight heatmap AI (KLA SURFmonitor AI, Thermo Scientific XRF analyser AI, Bruker M4 TORNADO AI) — ±6 DN adversarial perturbation. X-ray fluorescence measurement of electrode coating weight is the primary process control method for active material areal density — the fundamental parameter that determines cell energy density, power density, and N/P ratio balance. The XRF sensor directs a monochromatic X-ray beam at the moving electrode and measures the characteristic fluorescence emission of the transition metals in the active material (nickel Kα at 7.48 keV, manganese Kα at 5.90 keV, cobalt Kα at 6.93 keV for NMC cathode; iron Kα at 6.40 keV for LFP). The fluorescence intensity is proportional to the coating weight at the measurement spot. The measurement system rasters across the electrode width, generating a 2D heatmap of coating weight across the full electrode roll at a spatial resolution of 1–5 mm per pixel. The rendered heatmap uses a colour scale calibrated to the target coating weight — typically a blue-to-red scale where blue indicates below-target (thin zone), green indicates on-target, and red indicates above-target (thick zone). The electrode inspection AI classifies each pixel region of the rendered heatmap: conforming (within lower and upper process control limits), thin-zone alarm (below LCL — trigger: flag roll segment, alert process engineer), thick-zone alarm (above UCL — trigger: adjust slot-die pressure). A ±6 DN upward adversarial shift in the pixel region encoding a thin-zone area — moving the apparent colour value from the below-LCL blue zone to within the conforming green band — suppresses the thin-zone alarm. The electrode roll segment with below-LCL coating weight passes through the process without alarm and is incorporated into cell assembly. The N/P ratio in that electrode region is below design, meaning the anode intercalation capacity is insufficient for the incoming Li-ion flux during charging. Li-plating initiates in that region. The ±6 DN perturbation magnitude is within the combined heatmap rendering noise floor: XRF measurement shot noise at the process speed produces count-rate fluctuations equivalent to approximately 2–3 mg/cm² areal density variation (typically 1–2% at nominal coating weight), which maps to 3–5 DN in the standard 8-bit rendered colour scale. A 6 DN shift is within this noise envelope, making it indistinguishable from normal measurement variability in the rendered heatmap without dedicated adversarial detection at the image ingestion boundary.

NIR binder composition distribution AI (Bruker Optics NIR AI, Thermo Fisher Nicolet NIR AI, Zeiss NIR spectral AI) — ±8 DN adversarial perturbation. Near-infrared spectral imaging measures the spatial distribution of PVDF binder and carbon black conductive additive across the electrode width after drying. PVDF (polyvinylidene fluoride) is the standard cathode binder in NMC electrodes; it has distinct NIR absorption bands at approximately 1,170 nm and 1,200 nm that allow its areal concentration to be mapped spatially. During electrode drying, if the drying rate exceeds the diffusion rate of PVDF back from the drying surface, PVDF migrates toward the surface of the coating (top-surface migration), producing a binder gradient: binder-depleted at the coating-foil interface, binder-enriched at the top surface. This gradient reduces adhesion of the coating to the aluminium current collector and increases the electrode’s internal resistance. It can also produce localised carbon black depletion zones that reduce ionic and electronic conductivity. The NIR camera renders the binder distribution as a 2D spectral intensity image across the electrode width. The electrode inspection AI classifies the rendered NIR image for binder gradient severity: conforming (uniform distribution within tolerance), surface-migration alarm (top-surface enrichment above threshold), or interface-depletion alarm (interface concentration below minimum adhesion threshold). A ±8 DN adversarial shift in the NIR spectral intensity image suppresses the apparent gradient magnitude below the alarm threshold. The binder-depleted interface region proceeds into calendering and cell assembly without correction. Poor adhesion at the coating-foil interface produces electrode delamination during cycling expansion and contraction, creating localised areas of electrically disconnected active material — regions that then experience elevated current density on the remaining connected electrode area, leading to local overheating and Li-plating at the higher local current density.

Calendering density profile AI (Manz AG CQA density AI, Sick AG thickness gauge AI, Precitec chromatic confocal AI) — ±8 DN adversarial perturbation. Calendering compresses the dried electrode between precision steel rolls to a target thickness and density, controlling electrode porosity — the fraction of the electrode volume occupied by pores (electrolyte channels) versus active material and binder. Target porosity for NMC cathodes is typically 25–35% pore volume fraction; for graphite anodes, 25–40%. Electrode porosity is measured continuously by a thickness gauge system (micrometer, beta gauge, or chromatic confocal sensor) that renders a 2D density profile image of electrode thickness across the roll width, calibrated to porosity given the measured coating weight. The density AI classifies each profile position: conforming (within porosity tolerance), over-calendered alarm (thickness below minimum — too dense, too low porosity, Li-ion transport restricted), under-calendered alarm (thickness above maximum — too porous, insufficient energy density, mechanical instability). Over-calendering produces a consequence analogous to a thin zone in the XRF domain: insufficient accessible intercalation volume in the graphite anode. When porosity falls below approximately 20%, Li-ion transport to the graphite particles is kinetically limited — the electrolyte diffusion resistance in the compressed pore channels slows Li⁺ transport sufficiently that arriving Li-ions cannot reach available intercalation sites fast enough at normal charge rates, and plate as Li⁰ on the graphite surface. A ±8 DN adversarial shift suppressing an over-calendered thickness profile from below the minimum to within the conforming band allows over-calendered electrode to proceed to cell assembly. The same Li-plating pathway initiates during the first charge cycle in which the charge rate exceeds the Li-ion transport capacity of the compressed pore structure.

Slitting linescan burr AI (Cognex burr detection AI, Sick AG linescan AI, Teledyne Dalsa eLine AI) — ±10 DN adversarial perturbation. Slitting cuts the calendered electrode master roll (typically 600–800 mm wide) into cell-width electrode tapes (typically 50–120 mm wide for prismatic or cylindrical cells) using rotating knife or laser slitters. Metallic burrs — protrusions of the copper (anode) or aluminium (cathode) current collector foil material above the cut edge — are generated by improperly maintained slitter blades, blade alignment error, or foil contamination. Burr height critical threshold is typically 35–50 µm: a burr above this height can penetrate a 20–25 µm standard separator during jellyroll winding (under the compressive tension of the winding process), during tab welding (ultrasonic or laser) that compresses the electrode stack, or during cycling (as electrode expansion on lithiation adds stack pressure to the wound assembly). The burr establishes a direct metallic pathway from the current collector foil of one electrode through the separator to the current collector foil or active material of the opposite polarity electrode: an internal short circuit. At the slitting line, a linescan camera system inspects each electrode edge at line speed, rendering a greyscale or colour linescan image of the electrode edge profile at sub-10 µm pixel resolution. The burr detection AI classifies the rendered linescan for conforming (no burr above threshold) or burr alarm (burr height above 35–50 µm). A ±10 DN adversarial pixel shift at the edge profile peak — suppressing the apparent protrusion height from above the alarm threshold to below it — prevents the burr alarm. The electrode tape with the metallic burr proceeds into the winding or stacking process. This is the mechanism identified in the ATL Samsung Note 7 replacement battery failures: aluminium foil burrs from slitting penetrated the separator during winding, establishing the ISC that caused the second-generation Note 7 fires. A ±10 DN adversarial suppression of the slitting linescan burr AI replicates that inspection-escape pathway precisely, at the speed of the slitting line.

IEC 62619:2022, UL 9540A, UN 38.3, and EU Battery Regulation 2023/1542: the qualification framework and its AI boundary

IEC 62619:2022 (Safety requirements for secondary lithium cells and batteries for use in industrial applications) is the foundational international safety standard for lithium batteries in industrial contexts, adopted by CEN (EN IEC 62619:2022), JIS, and GB/T. Its cell safety requirements (Section 7) mandate that qualified cells pass a suite of abuse tests — overcharge, over-discharge, external short circuit, crush, nail penetration, thermal exposure, altitude simulation — that establish the cell’s robustness against foreseeable misuse at the qualified design point. The nail penetration test (Section 7.3.5) is particularly relevant: it uses a physical nail driven into the cell to create a controlled internal short circuit, testing whether the cell can tolerate this ISC without fire or explosion at the test conditions. If the cell design passes, IEC 62619 certifies that a cell at the qualified design — with the correct electrode geometry, coating weight, binder distribution, porosity, and burr-free electrode edge — can tolerate a specified ISC without unsafe consequence. It does not certify cells manufactured outside the design specification. A cell with a thin-zone electrode region, a binder-depleted interface, an over-calendered porosity zone, or a metallic burr in the jellyroll is not the cell that was tested in the IEC 62619 qualification campaign. IEC 62619 has no requirement that the AI inspection systems determining which cells match the qualified design be evaluated for adversarial robustness. No amendment to IEC 62619 through its 2022 edition addresses adversarial machine learning, and the standard’s scope is explicitly the performance of cells and batteries, not the manufacturing quality inspection systems that determine which cells reach the qualified design point.

UL 9540A:2023 (Test Method for Evaluating Thermal Runaway Fire Propagation in Battery Energy Storage Systems) addresses cell-level thermal runaway initiation, module-level propagation, and system-level fire spread for BESS installation qualification. Many US AHJ (authorities having jurisdiction) require UL 9540A testing for BESS permits. UL 9540A uses needle penetration or overcharge as the trigger to initiate thermal runaway in a test cell; the propagation behaviour to adjacent cells and the resulting heat release rate and flammable gas generation are characterised for the specific cell design and module packaging. UL 9540A does not address manufacturing quality AI inspection systems. It evaluates what happens when a cell of the qualified design enters thermal runaway — it does not address the AI inspection system that determines which cells with manufacturing defects (non-qualified design) enter service.

UN 38.3 (Recommendations on the Transport of Dangerous Goods, lithium battery transport qualification) specifies eight tests for lithium cells and batteries for IATA and IMDG shipping classification: altitude simulation (Section 38.3.4.1), thermal test (38.3.4.2), vibration (38.3.4.3), shock (38.3.4.4), external short circuit (38.3.4.5), impact (38.3.4.6), overcharge (38.3.4.7), and forced discharge (38.3.4.8). These tests are performed on cells at the qualified design. UN 38.3 does not address manufacturing AI. The gap follows the same pattern as SAE J2578 for hydrogen FCEV safety and IEC 60880 for nuclear I&C safety: the physical product at the qualified design point is tested against specified abuse and performance criteria; the AI system whose output determines which products reach that qualified design point is not within the standard’s scope.

EU Battery Regulation (EU) 2023/1542 introduces the most significant new requirement with AI implications: the digital battery passport, applicable to industrial batteries with >2 kWh capacity and EV traction batteries from specified implementation dates (February 2027 for traction batteries). Article 77(2) requires that each battery have a unique identifier and that a battery passport be accessible through a QR code, containing information on battery model, manufacturing location, manufacturing date, carbon footprint, capacity, state of health, and “important information for the consumer” on battery condition. The manufacturing quality inspection AI — XRF coating weight, NIR binder composition, calendering density, slitting linescan — generates the process data from which battery passport conformance records are derived. If the electrode inspection AI has been adversarially suppressed to classify a defective electrode roll segment as conforming, the digital battery passport entry for each cell wound from that roll segment will record no quality exception. EU Battery Regulation 2023/1542 has no adversarial robustness requirement for the AI systems that generate the manufacturing process data underlying the digital battery passport. The passport data integrity requirement (accurate, accessible, and up-to-date, per Article 77) is process-oriented — it assumes that the inspection AI providing conformance data is operating correctly and has not been adversarially manipulated.

From thin zone to thermal runaway: the progressive failure pathway that electrode coating AI must intercept

The thin-zone failure pathway from electrode manufacturing defect to thermal runaway in the deployed cell proceeds through four mechanistic stages, each of which amplifies the initial defect through a different electrochemical or physical mechanism.

Stage 1: insufficient intercalation capacity in the thin-zone region. A thin-zone region in the graphite anode — coating weight below the lower process control limit, porosity above the upper limit from under-calendering, or a combination — means that the available graphite Li-ion intercalation capacity in that region is below the design N/P ratio (the ratio of anode to cathode capacity, typically 1.05–1.15 in modern NMC/graphite cells to provide a 5–15% anode capacity buffer above the cathode). During charging, Li-ions deintercalate from the NMC cathode uniformly across the electrode area (assuming uniform current distribution at low to moderate C-rate). The Li-ions arriving at the graphite anode must intercalate into the staged LiC₆ structure. In the thin-zone region, the available LiC₆ sites are fewer than the Li-ion flux requires. The residual Li-ions that cannot intercalate plate as Li° on the graphite surface — the most electrochemically favoured metal surface available. This is the first appearance of Li-plating, a process that occurs preferentially at high state of charge (when the graphite is already highly lithiated and few LiC₆ sites remain), at low temperature (reduced Li-ion diffusivity in the graphite), and at high charge rate (Li-ion arrival rate exceeds intercalation rate).

Stage 2: lithium dendrite nucleation and growth. Metallic lithium deposited on the graphite surface does not form a smooth uniform film. It deposits preferentially at surface protrusions — graphite particle edges, existing deposit irregularities, binder defects — due to the locally elevated electric field at these tips (tip enhancement of the electrodeposition rate). Over repeated charge cycles, the metallic deposits grow as dendritic structures, elongating toward the separator in the direction of highest electric field (perpendicular to the electrode surface, toward the cathode). The rate of dendrite growth scales with the degree of Li-plating per cycle, which scales with the severity of the thin-zone anode capacity deficit and with the cell’s operating conditions (temperature, charge rate, depth of discharge). A thin zone producing 5% excess Li-ion flux per cycle — within the range of a coating weight defect at the margin of the LCL — can initiate detectable dendrite growth within tens of cycles at standard operating conditions. In consumer electronics or EV applications, where cells may complete 200–500 cycles in the first year of use, the dendritic growth from a manufacturing thin zone can reach the separator within the product’s expected in-service life.

Stage 3: separator penetration and internal short circuit initiation. The separator in a modern Li-ion cell is a 20–25 µm microporous polyethylene (PE), polypropylene (PP), or trilayer PP/PE/PP membrane. Its puncture strength — the force per unit area required to penetrate it with a sharp protrusion — is typically 1–5 N/mm² depending on porosity, ceramic coating, and crystallinity. A lithium dendrite tip growing toward the separator eventually contacts the separator surface. As the dendrite applies pressure to the separator, it concentrates stress at the contact point. When the local stress exceeds the separator puncture strength, the separator fractures at the contact point and the dendrite penetrates through, contacting the cathode active material or the aluminium cathode current collector on the opposite side. This is the internal short circuit: a metallic pathway between the negative and positive electrodes through the failed separator. The ISC resistance depends on the contact area and dendrite geometry — initially very high (MΩ range for a single dendrite tip in point contact), dropping rapidly as the contact area increases through local heating and separator melt propagation (PP softening at approximately 150°C, melting at approximately 165°C), progressing to the mΩ range as the contact area grows to mm² scale.

Stage 4: thermal runaway. From the established ISC, the consequence pathway is identical to the ISC created by a metallic burr in the Samsung SDI Note 7 cells or the lithium plating in the Boeing 787 GS Yuasa cells. The local I²R heating rate at the ISC site raises the local temperature. Above approximately 70°C, electrolyte solvent (EC/DMC or EC/EMC) begins to decompose exothermically, releasing dimethyl carbonate vapour and other organic compounds. Above approximately 130–150°C, the cathode oxide (NMC, LCO) undergoes exothermic oxygen-releasing decomposition: for LCO, LiₓCoO₂ → CoO + O₂ + exothermic; for NMC, similar reactions begin at approximately 180–210°C. The oxygen released by the cathode reacts with the electrolyte and the metallic lithium deposits on the anode surface, accelerating the exothermic reaction rate. The cell temperature rises toward the runaway temperature (approximately 200–300°C for the full cell), cell pressure rises from gas generation, the safety vent activates, and the released flammable gases (primarily hydrocarbons, CO, and H₂ from organic decomposition) ignite on contact with any ignition source — including the ISC arc itself, any electrical spark in the environment, or the hot cell surface. In a multi-cell pack (EV battery module, 787 APU battery), the heat generated by the venting and burning cell transfers to adjacent cells, raising their temperature toward their own runaway threshold — thermal runaway propagation, the mechanism characterised by UL 9540A and documented in the Boeing 787 APU battery investigation as cell-to-cell propagation from a single initiating cell.

Glyphward threshold 35 for Li-ion gigafactory electrode coating AI

Glyphward’s adversarial detection API operates as a pre-classification gate at each rendered-image ingestion boundary in the electrode coating inspection pipeline: before the XRF coating weight AI processes the heatmap render, before the NIR binder composition AI processes the spectral image, before the calendering density AI processes the thickness profile, and before the slitting linescan AI processes the edge profile frame. Each rendered inspection image receives a risk score (0–100) in 8–15 ms at process speed. At or above threshold 35, Glyphward gates the AI classification and flags the electrode roll position for manual re-inspection — without waiting for the inspection AI to produce a potentially adversarially suppressed conformity verdict.

Threshold 35 for Li-ion gigafactory electrode coating AI reflects three consequence factors and one structural factor. First, the fleet-scale consequence multiplier. A coating excursion at a CATL or LG Energy Solution gigafactory running at 60 m/min on a 700 mm-wide electrode roll produces approximately 42 m² of electrode per minute. A 30-second adversarial suppression of a thin-zone alarm covers approximately 21 m² of defective electrode — sufficient material for thousands of cells in a winding geometry. At gigafactory production rates of 10–50 GWh/year per facility, defective material from a 30-second inspection escape enters the cell assembly stream at a scale that the Samsung Note 7 recall demonstrated can reach consumer deployment in millions of units before the defect is identified through field returns. This fleet-scale multiplier — from a single suppression event at the inspection AI to millions of deployed defective cells — is the principal factor driving threshold 35 above threshold 30 for other high-consequence industrial AI domains such as hydroelectric dam spillway AI or tailings dam monitoring AI, where the consequence is severe but the scale is a single facility.

Second, the aviation consequence from the Boeing 787 APU battery events. The 4-month worldwide grounding of the entire 787 fleet — triggered by a manufacturing quality failure that produced lithium plating in GS Yuasa cells — is the documented aviation-scale consequence envelope. Electrode coating AI adversarial injection that allows lithium-plating-prone cells into aviation battery supply chains produces a consequence pathway with precedent at FAA Emergency AD scale. Third, the sole-barrier architecture. The XRF coating weight AI, NIR binder AI, calendering density AI, and slitting linescan AI are each the primary automated real-time detection mechanism for their respective defect class at process speed. No redundant automated inspection system operates in parallel at the same spatial resolution and temporal frequency at a standard gigafactory electrode line. A defect that escapes the AI gate at each process step has no subsequent automated detection opportunity before cell assembly. Fourth, the false positive cost proportionality. A false positive threshold-35 gate trigger in the electrode coating AI context produces a manual re-inspection of the flagged roll segment — a process step that takes minutes and may result in a roll trim or quarantine. The false negative cost — an adversarially suppressed thin-zone alarm producing a defective roll segment that proceeds to cell assembly — is measured in millions of deployed cells at the Note 7 consequence scale.

Threshold 35 is 10 points above threshold 25 for nuclear power plant digital I&C AI. The nuclear domain has the NRC GDC 20–24 single-failure criterion — an explicit regulatory requirement that no single component failure shall prevent a safety function — with no equivalent in IEC 62619. The nuclear consequence anchors (TMI-2 1979: 50% core damage, $1.1B; Fukushima Daiichi 2011: three reactor buildings, multi-decade facility loss, $200+B) are larger in absolute consequence scale than the Samsung Note 7 or Boeing 787 events. Threshold 35 versus threshold 25 is not a judgment that Li-ion gigafactory inspection AI adversarial injection is a greater risk than nuclear I&C AI adversarial injection — the nuclear consequence anchors are at the upper end of any consequence portfolio. The threshold difference reflects that the nuclear safety system design is explicitly built to tolerate a single failure (GDC 20–24 single-failure criterion), while the Li-ion gigafactory electrode inspection AI operates as the sole automated real-time quality gate with no regulatory redundancy requirement, and the fleet-scale multiplier from a gigafactory inspection escape dwarfs the scope of a single nuclear event.

Free tier — 10 scans/day, no card required. Submit a rendered electrode XRF coating weight heatmap or slitting linescan frame from your manufacturing inspection system to the Glyphward scanner to generate a baseline adversarial risk score for your electrode quality inspection AI inputs.

FAQ

What caused the Samsung Galaxy Note 7 thermal runaway recalls in 2016, and how does it establish the consequence envelope for electrode coating AI adversarial injection?

The Samsung SDI cells in the first Galaxy Note 7 recall failed due to electrode geometry tolerance violations: the graphite anode had insufficient overhang beyond the cathode boundary in jellyroll corner regions, allowing cathode material to contact the anode copper current collector or separator under assembly compression, creating an internal short circuit (ISC). The ATL replacement cells failed through a different manufacturing defect: aluminium foil burrs from slitting penetrated the polyethylene separator during winding, establishing a second ISC pathway. Both root causes are manufacturing quality defects at the electrode level — exactly the defect class that XRF coating weight AI, NIR binder composition AI, calendering density AI, and slitting linescan burr AI are designed to detect in real time. Samsung permanently discontinued the Note 7 on 11 October 2016. The US CPSC issued a formal recall. The FAA Emergency Order effective 14 October 2016 banned the Note 7 from all US aircraft (carry-on, checked baggage, and cargo). Airlines worldwide implemented equivalent bans. Total costs are estimated at approximately $17 billion (direct recall costs ~$5.3B plus lost Note 7 and Galaxy S series revenue impact). The Note 7 event establishes the consequence envelope for electrode coating AI adversarial injection because it documents, with globally reported evidence, the consequences of manufacturing defects at the electrode level — geometric tolerance and metallic burr — that escaped quality inspection and produced ISC-triggered thermal runaway in 2.5 million deployed cells. Adversarial injection that suppresses the electrode inspection AI replicates this inspection-escape pathway at process speed, across every cell manufactured during the suppression window.

What happened to the Boeing 787 APU batteries in January 2013 — and what does lithium plating reveal about the regulatory gap for manufacturing quality AI?

On 7 January 2013, a JAL 787 at Boston Logan experienced thermal runaway in its APU battery (GS Yuasa LCO/graphite, 8-cell 32V 63Ah). On 16 January 2013, an ANA 787 diverted to Takamatsu after battery smoke entered the cabin. The NTSB investigation concluded the initiating failure was an internal short circuit; the most probable mechanism, based on evidence of lithium deposits on the anode, was lithium plating — metallic lithium plated on the graphite anode from localised anode capacity limitation during charging. Lithium plating is the electrochemical signature of insufficient graphite intercalation capacity relative to the incoming Li-ion flux — the same condition produced by a thin-zone electrode coating defect that reduces anode active material loading below the N/P design ratio. The FAA issued Emergency AD 2013-02-51 on 16 January 2013, grounding all 787s worldwide. The fleet remained grounded approximately four months (January–April 2013). The 787 events are significant for the electrode coating AI adversarial injection threat model because they document aviation-scale consequences — FAA Emergency AD, 4-month worldwide fleet grounding, Boeing $600M+ in engineering and delay costs — from a manufacturing quality failure that produced lithium plating in GS Yuasa cells. The NTSB investigation examined GS Yuasa’s manufacturing process controls for electrode coating weight measurement, calendering density, and cell formation cycling. None of those process control AI inspection systems were evaluated for adversarial robustness because IEC 62619, UN 38.3, and the applicable JCAB regulations had no such requirement — the same regulatory gap that Glyphward threshold 35 addresses.

How does a thin zone in the electrode XRF heatmap lead to Li-plating and thermal runaway — and why is the coating AI the sole automated detection gate?

A thin zone — anode coating weight below the lower process control limit — means the graphite intercalation capacity in that electrode region is below the N/P design ratio (typically 1.05–1.15). During charging, the Li-ion flux from the cathode arrives at the anode uniformly across the electrode area; in the thin-zone region, available LiC₆ intercalation sites are fewer than the arriving flux requires. Residual Li-ions that cannot intercalate plate as metallic lithium (Li°) on the graphite surface — lithium plating. Over repeated charge cycles, the Li° deposits grow as dendritic structures, elongating toward the separator (perpendicular to the electrode surface) at surface protrusions due to localised tip field enhancement. When a dendrite elongates sufficiently to contact the 20–25 µm separator, it applies point stress exceeding the separator puncture strength (1–5 N/mm² for standard PE or PP), penetrating through to contact the cathode. This is the ISC: a metallic short circuit between the positive and negative electrodes. Local I²R heating raises temperature above electrolyte decomposition (~70°C), cathode oxide decomposition (~180–210°C), and full thermal runaway, with the possibility of cell-to-cell propagation in a multi-cell pack. The XRF coating weight heatmap AI is the sole automated real-time gate capable of detecting a thin zone at process line speed — 40–100 m/min — with sufficient spatial resolution to flag a 1–5 mm defect region before the electrode roll proceeds to calendering, slitting, and cell assembly. No redundant independent inspection system operates at the same speed and resolution at a standard gigafactory electrode line. A ±6 DN adversarial suppression of the thin-zone pixel region in the rendered XRF heatmap removes this sole detection gate entirely.

What does IEC 62619:2022 require for Li-ion battery safety — and what is the adversarial robustness gap for electrode inspection AI?

IEC 62619:2022 (Safety requirements for secondary lithium cells and batteries for use in industrial applications) establishes cell safety requirements through abuse tests at the qualified cell design: overcharge (Section 7.3.1: 2C to 150% SoC, no fire/explosion), over-discharge (7.3.2), external short circuit (7.3.3: 0.1 Ω short, no fire/explosion), crush (7.3.4), nail penetration (7.3.5: direct ISC test proxy), thermal exposure (7.3.6: 130°C for 30 min), altitude simulation (7.3.7). Battery system requirements (Section 8) address BMS functions: cell balancing, overcurrent/overvoltage/undervoltage/overtemperature protection. IEC 62619 does not address AI manufacturing quality inspection systems. Its scope is the performance of cells and batteries, not the AI inspection systems that determine which cells match the qualified design. A cell with a thin-zone electrode, binder-depleted interface, over-calendered pore structure, or metallic burr in the jellyroll is not the cell tested in the qualification campaign. IEC 62619 nail penetration test (Section 7.3.5) does not protect against a manufacturing-defect ISC: the test uses a physical nail at defined conditions; it validates that a qualified cell can tolerate a controlled ISC. It does not validate that cells with manufacturing defects produce the same response — and it has no requirement that the AI inspection system determining cell manufacturing conformity be evaluated for adversarial robustness. No amendment to IEC 62619 through its 2022 edition addresses adversarial machine learning. EU Battery Regulation 2023/1542 Article 77 (digital battery passport) requires accurate manufacturing data for each cell — but specifies no adversarial robustness criterion for the AI inspection systems that generate that data. An adversarially suppressed defect detection in the electrode coating AI produces a falsified digital passport entry without any regulatory detection mechanism.

Why does Glyphward apply threshold 35 for Li-ion gigafactory electrode coating AI — and how does it compare to threshold 30 for hydroelectric dam spillway AI?

Threshold 35 reflects four factors: (1) fleet-scale consequence multiplier — a 30-second adversarial suppression of a thin-zone alarm at a gigafactory running at 60 m/min produces approximately 21 m² of defective electrode, sufficient for thousands of cells; at gigafactory production scale, this can enter consumer deployment in millions of units before field failure rates reveal the problem, as the Note 7 event documented with 2.5 million recalls; (2) aviation consequence — the Boeing 787 APU battery events document that Li-ion manufacturing quality failures (lithium plating from electrode coating defects) can produce FAA Emergency AD, 4-month worldwide fleet grounding, and $600M+ Boeing costs; (3) sole-barrier architecture — each electrode inspection AI (XRF, NIR, calendering density, slitting linescan) is the only automated real-time gate for its defect class at process speed, with no regulatory redundancy requirement or independent parallel detection system; (4) false positive cost proportionality — a threshold-35 false positive triggers a manual roll re-inspection (minutes); a false negative allows defective electrode into millions of deployed cells. Threshold 35 versus threshold 30 for hydroelectric dam spillway AI reflects primarily the fleet-scale multiplier: the Oroville Dam 2017 event (45 m erosion crater, 188,000 evacuations, $1.1B) is a single-facility event with a severe but geographically bounded consequence. The electrode coating AI adversarial injection threat model scales to millions of deployed cells across hundreds of product lines and deployment contexts — vehicle, aviation, consumer electronics, BESS — from a single suppression event at a single gigafactory inspection AI.