Fuel-cell twin
Hydrogen-electric systems / Research build

A twin that shows its evidence.

Not physically validated

Explore the fuel-cell stack, its supporting systems and a causal estimation loop. Synthetic rig replay sits alongside a separately locked measured-data challenge. Evidence classes stay explicit.

Predicted stack voltage
—
Prior prediction, before assimilation
Net electrical output
—
Stack less assumed auxiliaries
Membrane water λ
—
Model-based virtual sensor
Hydrogen consumption
—
Faraday consumption, excludes purge

Voltage · prediction and observation

TwinSynthetic measurement
0 s
Shading: nominal 95% model interval, including an assumed discrepancy floor. Coverage is scored; this is not a safety probability.

Observer state

Replay
Predict→Score residual→Assimilate
Voltage residual
Effective resistance
Parameter update
Data / domain gate
An alarm does not uniquely identify degradation. Sensor bias and omitted physics can produce the same residual.

Thermal response

Membrane-water estimate

No direct water measurement is fed to the observer. Hidden-state scoring against synthetic truth is available in the exported evidence.

Frozen protocol. Visible failures.

These provisional gates test synthetic prediction performance. They are not regulator-agreed or derived from an aircraft safety margin.

ScenarioTwin RMS / VPersistence / VInterval coverageEligibleOutcome
Persistence predicts the last observed voltage. No model is credited simply for fitting a smooth signal. Fault challenges are scored over the complete eligible sequence, so a deliberate anomaly may cause a prediction gate to fail.

Acceptance contract

Interval width is checked alongside coverage. A very wide interval cannot pass just by covering everything.

What is — and is not — established

  • Executable physics and causal replay implementation.
  • Separate scalar arithmetic rechecks the primary scorer.
  • Missing data and domain exits withhold claims.
  • Shared simulator equations limit the strength of the synthetic tests.
  • No measured fuel-cell validation, independent scientific review, remaining-life qualification or certification evidence.
Physical validation: not established

Frozen identity

The local manifest is written before replay generation. Its hash establishes identity, not proof of unseen data, independent preregistration or regulatory approval.

Can these parameters be identified?

Range-scaled sensitivity with singular-value analysis; multistart bounded fitting. This local test cannot prove global uniqueness.

Calibration dataNumerical rankCondition number
Concentration and limiting-current parameters should not be updated from low-current data alone. The live loop only adapts an effective resistance, within bounds and excitation gates.

Best next test candidates

Local D-optimal information gain per assumed rig minute. Advisory only; no automatic test execution. See Bayesian design for the new finite-prior information-gain engine with marginalized hydration uncertainty.

j / A cm⁻²Pressure / barRHGain / min

Bounded parameter map

Gaussian process on logit-transformed resistance. Separate assumed observations; candidate map is not used by the replay. Bounds are enforced; outside-domain queries are refused.

Average cell versus weakest cell

Mean cell5th percentile of stack minimum

400 cells × 500 synthetic draws at each load. Shared and cell-specific variation are included. Tail distributions remain assumed; this is not a stack derating recommendation.

A queryable component model

Select a component to see the implemented boundary. The example uses 400 cells and 400 cm² active area; these are illustrative, not ZeroAvia specifications.

Physical relationships

Cell voltage combines reversible voltage with activation, ohmic and concentration losses.

Vcell = E − activation − j × ASR − concentration
ASR = Rref × (14 / λ) × exp[−0.012(T − 70)]
C × dT/dt = Qstack − Qcoolant
Qcoolant = ṁcp × [1 − exp(−UA / ṁcp)] × (T − Tin)
ṁH₂ = N × I × MH₂ / (2F)

Thermal heat uses a 1.48 V thermoneutral (HHV) basis. Purge, recirculation, phase-change detail, spatial hot spots and ageing mechanisms are not resolved.

Controlled learning

Fixed physics

Versioned equations and component interfaces.

State estimation

T, pressure, RH and airflow assimilated causally.

Bounded adaptation

Effective resistance only, with transient and residual gates.

Candidate maps

GP proposals remain separate from the released model.

Effective resistance drift cannot establish a unique degradation mechanism. Lifetime and maintenance dates are withheld.

Parameter register

ParameterValueBoundsEvidence

What the battery twin taught this build

Read from feature/v1-heavy-duty at 823fa96. The main branch is a placeholder. These are method transfers; no battery electrochemistry, restricted datasets or source code are copied.

Observed lessonFuel-cell implementationRepository evidence
More thermal states did not repair wrong heat generation.Start with constrained physics; test identifiability before adding states.Two-node results
A summer calibration failed on a different thermal regime: 7/17.Explicit domains and varied-condition calibration; OOD scenario withholds claims.Loop v4 results
A zero-power baseline could also pass the thermal test.Always score persistence alongside the causal twin.Learning L69
Frozen CAN signals and repeated timestamps distorted scoring.Strict timestamps, missing-data handling and stale telemetry detection.Learning L70
Mean cell behaviour hid the first limiting group.Show average and weakest-cell distribution; include shared variability.Pack twin b3
Protocols, results and scorers were separated.Frozen manifest, explicit score gates and separate arithmetic check.Protocol registry

ARCHEE scope mapping

  • Implemented prototype: physics model, queryable parameters, bounded estimation, virtual hydration, residual diagnostics, stochastic cell spread.
  • Demonstrated separately: bounded GP parameter map and D-optimal test ranking.
  • Not reproduced: proprietary ZeroAvia models, full aircraft propulsion controls, high-voltage protection, certified diagnostics, predictive-maintenance schedules.

Primary references

The supplied Tom Fletcher transcript is the source for the algorithm sequence. No affiliation or endorsement is implied.

Replay your measured rig data

The accompanying Python package provides a strict CSV adapter and a stateful Twin.step(sample) interface. This offline HTML replays computed results; it does not connect to equipment.

python -m archee.run replay your-rig.csv --out output/measured

Required inputs include time, current, pressure/RH targets, coolant inlet and flow, stoichiometry target and ambient temperature. Measurements: voltage, stack temperature, cathode pressure, RH and airflow. See docs/DATA_CONTRACT.md for exact columns and units.

Measured-data provenance and hardware parameters must be supplied and reviewed before physical validation can be claimed.
Locked measured records
43,303
14,363 system + 28,940 EIS
Synthetic training examples
1,100
Separate source's 11 load points
Conditional model RMS error
Cell voltage; temperature/hydration proxies
Fixed 0.7 V baseline RMS error
Model does not beat the baseline

External challenge · no fitting

The recorded current density and average cell voltage support a conditional static comparison. Gauge pressure is converted using assumed ambient pressure; coolant outlet temperature and an equilibrium hydration relation act as proxies.

Not physically validated

These assumptions block a physical validation claim. Time samples are correlated. EIS is reserved for future model extensions. The data were inspected during preparation and are not blind.

Training without validation leakage

Bisello's beginning-of-test load points define the training input coordinates. Measured voltages are excluded from the generator. Temperature, humidity, pressure and stoichiometry are explicitly assumed; targets come from frozen physics.

RealHyFC's entire acquired campaign stays in the validation partition. SHA256 checks protect the model parameters and locked records. Neither a trained surrogate nor unique degradation estimates are claimed.

Source disposition

Primary sourceIncludedUse
RealHyFC · Fensterle et al.43,303 observations · CC BY 4.0Locked external challenge; no training or fitting
Bisello et al.178 polarisation pairs; 11 selected load points · CC BY 4.0Training stimulus source only; not validation evidence
Cornet / LEPMI105,668 measured pairs · Etalab 2.0Quarantined: pressure/species conflicts and unresolved area
PHM 2014 / FCLAB mirrorMetadata onlyRaw bytes excluded: license and original-source verification unresolved

Sensitivity is not identifiability

Global Sobol indices quantify parameter influence across the prior. Extended local sensitivities then test which parameters can be distinguished over multiple prior points.

Recommendations are conditional on the chosen noise, priors and observed quantities. They are not proof of global uniqueness and do not enable parameter updates.

What Bayesian design adds

Expected information gain averages over electrochemical uncertainty and marginalizes uncertain hydration. Correlated model discrepancy limits the apparent benefit of repeated tests.

Candidate costs and risk bounds are assumptions. Nearby ranks overlap in Monte Carlo uncertainty. This is a greedy batch recommendation, not an optimal adaptive flight or rig policy.

Global sensitivity · total-order Sobol indices

Independent uniform priors; four randomized Sobol scrambles. Raw estimates are not clipped. Hydration bias is an uncertain nuisance, not a measured state.

Current densityActivationResistanceConcentrationLimiting currentHydration bias

Candidate experiments · expected information per minute

jT / °CPressure / barRHEIG / nats ± MC SEMinutesNats / min

MC standard errors are conditional on the finite prior and do not bound continuous-prior approximation bias. An independent 128-particle / 1,024-draw check is included in the downloadable evidence.

Joint-information batch

StepCandidateJoint EIGIncremental EIGHydrogen / kg assumed

Uses joint likelihoods and correlated discrepancy, not the sum of standalone scores. Hydrogen cost uses the illustrative 400-cell configuration; a hardware-specific test review is still required.

Research basis and remaining gaps

Goshtasbi et al., robust subsets and experimental design · Vetter & Schumacher, sensitivity across operating conditions · Huan & Marzouk, Bayesian information-gain design.

This implements these principles for the reduced model. It does not reproduce the papers' full fuel-cell physics or Fletcher's unpublished student work. Design assumptions need independent development evidence; the measured model error remains unresolved.