AI / ML Model Registry

Pixel-mask cracks, finger interruptions, dark areas in EL images.

Grades each cell A/B/C from a 224² EL crop using a fine-tuned ViT.

Detects hairline cracks invisible to YOLO seg head.

Segments grid lines; flags broken fingers > 3 segments.

Locates ribbon-to-busbar joints; classifies cold/over-soldered.

Detects EVA outgassing bubbles, edge delam, and trapped-air blisters from backlit RGB and EL signature (dark core + bright fringe). Drives bubbleRisk per module.

Surface defects on ITO/IWO after sputter.

Generates rare-defect EL images to balance training.

Cell edge / corner chip localisation pre-stringer.

Total-thickness-variation and bow from optical interferometry.

Pairwise similarity between current EL batch and golden lot.

Reads laser-engraved IDs; ties EL → recipe → operator.

Open-set: flags EL features the labelled classifier has not seen.

Detects pre-lamination glass scratches > 30 µm.

Compares expected vs actual cell motion.

Side-cam ribbon offset > 0.3 mm flag.

Reconstruction-error anomaly on photoluminescence images.

IR thermography time-series → uniformity index.

Aligns simulated EL to factory camera domain for training.

Verifies junction-box ribbon soldering geometry.

UV-fluorescence based EVA cross-link estimate.

3D bead width/height vs spec for module framing.

Fuses RF power, gas flows, OES into hidden i-layer thickness state.

Estimates instantaneous deposition rate from V/I and pressure.

Reconstructs true tip T from thermocouple lag.

Maps recipe → IV curve in <1 ms (replaces SCAPS).

Smooths multi-zone TC noise into per-zone peak T.

GP regression Pmax from upstream EL + IV-test subset.

NILM-style disaggregation of plant energy → per-station kW.

Probabilistic position of every cell on the line (latency-aware).

Continuous-time KOH concentration from titration + temp.

Estimates joint torque without torque sensors.

Couples Arrhenius kinetics with measured T to estimate cure %.

DAG over latent process states feeding inverse solver.

Solves transient heat eq. on cell from sparse TC+IR.

Fits leak rate from p(t) rise vs pump curve.

Decomposes OEE into A·P·Q with confidence bands.

Couples LeTID, PID, hydrolysis → power(t) trajectory.

Days-to-failure with 95 % CI from vibration spectra.

Survival probability for PECVD/cure heater filaments.

Discrete-time hazard for sputter-target erosion.

Ranks which defect class is likeliest to spike next hour.

Score >0.7 raises a Medium/High alert.

Top-3 drivers behind the current anomaly.

Posterior over candidate causes given observed KPI drift.

P50/P90 modules-per-hour next 60 min.

Quantile forecast of grid tariff for battery dispatch.

Next-shift module demand and WIP build-up.

Hazard of stock-out for each consumable in next 7 days.

Predicts when KOH bath chemistry leaves spec.

Time-series transformer detects slow drift before SPC limits trigger.

Predicts spectrogram band-energy growth on rotating assets.

Splits FPY loss into station × defect contributions.

GP forecasts module Pmax bin distribution before flash test.

Hazard for potential-induced-degradation under field stress.

Estimates light-elevated-temperature degradation susceptibility.

Graph NN over line topology forecasts cascade delays.

Self-exciting model groups bursty alerts into incidents.

Decomposed linear DL for hourly site kW forecast.

TFT fuses multi-asset signals into joint RUL.

Per-module probability of triggering a warranty claim in 25 yr.

Counterfactual ΔPmax for a proposed recipe change.

Closed-loop MPC steers i-layer thickness to set-point.

Sheet-R control via O₂ partial-pressure MPC.

Auto-tunes 16-zone lamination platen PID gains.

Packs PMs into low-throughput windows to minimise loss.

Assigns operators to stations subject to skills + fatigue.

Nearest-qualified technician for active alert.

RL agent dispatches battery vs grid given tariff forecast.

Joint speed of inter-station conveyors to absorb micro-stops.

Bayesian optimisation over recipe levers to maximise η.

Packs modules into customer SKU bins to maximise revenue.

Co-optimises iron T, dwell, force for ribbon strength.

EOQ/JIT mix per consumable under stock-out hazard.

Schedules module placement on multi-platen press.

Schedules sensor calibrations w/o disturbing FPY.

Cuts HVAC kWh while holding cleanroom RH/T spec.

Robust to leak; minimises pump-down time.

Minimises mismatch loss within each module string.

Stochastic plan for multi-week ramp under demand uncertainty.

Conversational copilot grounded in live snapshot + Bayesian posteriors.

Retrieves similar past events + resolutions for the open alert.

Drafts step-by-step SOPs from asset + symptom.

Transcribes operator radio chatter into structured notes.

Generates handover notes from MES + alerts + chat.

Suggests safe PLC ladder patches for repetitive faults (human-approved).

Generates rare-fault vibration signals to balance RUL training.

Drafts warranty / RMA replies grounded in module dossier.

Localised TTS for safety & andon broadcast.

Explains a Bayesian-Opt suggestion in plain language with trade-offs.

Describes EL anomalies for the operator with cited regions.

Plain-language explanation of any logged AI decision.

Reconstruction error flags dead/drifting sensors.

Federated averaging of yield models across sister fabs.

PSI / KS drift on every model's input + output.

Conformal wrapping yields valid 95 % intervals for any model.

HTE of each lever on η across sub-populations.

Filters RL/MPC actions against safety envelope before commit.

Tracks latency, accuracy, drift SLAs of all 100 models.

Graph linking models, evidence, decisions, outcomes for audit.