How the models work, and where they stop

The cell model

Cell voltage is assembled as:

V_cell = E_rev + η_anode(j) + η_cathode + j · ASR

E_rev is fixed at 1.18 V, appropriate for liquid water near 80 °C. The anode overpotential is extrapolated from a reported value at 10 mA/cm² to the operating current density using a Tafel slope: η(j) = η₁₀ + b · log₁₀(j / 0.01). Ohmic loss uses a single area-specific resistance covering membrane, contact and interconnect.

Energy per kilogram follows from Faraday's law — two electrons per mole of hydrogen, 496.03 mol per kilogram:

kWh/kg = 26.59 · V_cell / faradaic efficiency

The cost model

Levelised cost is the sum of electricity, amortised capital, stack replacement and fixed operating cost, each divided by the hydrogen produced per kilowatt per year at the chosen capacity factor. Capital is amortised with a standard capital recovery factor from the WACC and project life. Stack replacement is costed at 40 % of system capex, amortised over the stack's service life at the chosen capacity factor. Fixed opex is 3 % of capex per year.

Screening accuracy

Activity and stability predictions come from machine-learned interatomic potentials, not from converged DFT. This is a deliberate trade: it makes screening fast and cheap enough to be worth doing on a whole composition space, at accuracy sufficient to rank and triage.

Use it to decide what deserves a converged calculation or a lab slot. Do not use it as the final number in a paper or a specification.

General-purpose interatomic potentials are trained predominantly on bulk crystals and are measurably least accurate on oxide surfaces with oxygen adsorbates — which is the acidic OER regime. That gap is the reason this platform is specialised rather than general, and closing it is what our own model work is for.

Where we refuse to answer

Outside the validated composition and condition domain, a run is refused with an explanation rather than returning a plausible number. Refusals appear in results as refusals, not as silently dropped rows. At self-serve scale nobody is there to warn you, so the system has to decline instead.

What none of this covers

• Degradation mechanisms beyond a first-order dissolution-risk estimate.
• Balance of plant, grid connection, compression, storage or distribution.
• Mass-transport limitation at high current density.
• Whether a composition can actually be synthesised, or at what cost.
• Anything a supplier would actually quote you.

Versioning

Method changes get a version number and a changelog entry. Results carry the version that produced them and can be re-run pinned to it. Published versions get a permanent link and a DOI.