What to put in, what comes back, and how to read it. Written for the engineer who will actually use the result — including the parts where the platform will tell you it does not know.
Steel (Fe) or aluminium (Al). Titanium and nickel superalloys arrive in V3.0 and are not covered today — the platform declines them rather than extrapolating into chemistry it has never been trained on.
Weight percent, element by element. Leave unknowns blank rather than guessing. A blank field is filled from a stated prior and the result tells you it did so; a guessed number is treated as a measurement and silently believed. That distinction matters more than most people expect.
The forming route first — forged, rolled, cast, extruded — then the heat treatment stages in order, each with its temperature and time. The route carries as much of the answer as the chemistry does: the same steel normalised and quenched-and-tempered are different materials.
Every property arrives with a ±1σ spread, an indication of how much measured data sits near your input, and any physics warnings that fired. The spread is part of the answer, not a disclaimer attached to it.
Yield, tensile, hardness and elongation from composition and route — predicted together, so they stay consistent with each other instead of one being derived from another through a multiplier.
Endurance limit and life, through the Basquin S–N chain with porosity and surface knockdowns, Murakami defect threshold and Goodman mean-stress screening.
The question run backwards: give a target property window, get compositions and process routes that land inside it.
Carbon equivalent CE(IIW) and Pcm on the parent metal, hydrogen-cracking risk and a preheat indication. Screening only — it is not a WPS, and it says so on every run.
The same arithmetic is free and ungated at the carbon equivalent calculator.
A plain-language way in: describe the problem, and METALL calls the same engineering tools the forms do — then proposes the run and waits for you to confirm before anything executes. It is on the paid plans because every message costs real money at a commercial language-model provider.
| What you see | What it means | What to do |
|---|---|---|
| ±1σ on a property | One standard deviation of the model's own spread. Not a guarantee, not a statistical confidence interval. | Use the band, not the midpoint, when the decision is close. |
| P10 design life | The fatigue life that 90 % of measured specimens exceeded. | This is the number to design against, not the median life. |
| Data coverage / distance | How far your input sits from anything in the training corpus. | A large distance means the number is an extrapolation. Test it. |
| “Beyond model range” | The question is outside where the model was fitted. | Treat any accompanying figure as informational, never as a design value. |
| A warning | A physics check fired — SCC risk, an impossible temper, a route with thin data. | Read the condition it names. It is specific, not boilerplate. |
| A refusal | The platform will not answer — unsupported family, infeasible schedule. | Take it at face value. A refusal is cheaper than a confident wrong answer. |
A composition and a route is enough. If the platform cannot answer well, it will tell you that instead of guessing.
Open the app — free tier Free calculators, no account