Every major industry built a digital brain.
Metals never did. We're changing that.

Worked Examples

The honest way to judge a prediction platform is to read one of its reports end to end — including the places where it hedges, and the places where it tells you not to use the number at all.

Our worked example is a real material-substitution study, anonymised: a component turning at about 1,700 rpm for twenty hours a day, where the question was whether an aluminium alloy could replace precipitation-hardening stainless without losing fatigue margin.

• Four candidate alloys compared on yield, tensile, hardness and elongation
• Fatigue assessed at the real duty cycle — roughly 625 million cycles a year
• Every warning the platform raised, reproduced verbatim
• The two points where it refused to give a design answer, and why that is the useful result

Inverse Design & Process Optimization

Engineers rarely ask “what are the properties?” They ask “how do I reach them?” AutoForge inverts the question: give it a target window and it returns compositions and process routes that land inside it — ranked by whether your shop can actually run them.

• Target-driven recipes: composition, solution and ageing schedule, quench medium, cooling rate
• Pareto-ranked candidates, so a conservative route and an aggressive one sit side by side
• Manufacturability screened before ranking — carbon equivalent, hydrogen-cracking risk, fatigue
• Three searches in parallel: differential evolution, Bayesian optimisation, quantum-hybrid annealing — the same Pareto front in 92 s against 1,611 s classically

AutoForge proposes; it does not certify. Every candidate carries the same uncertainty as a forward prediction, and a route with thin training coverage is labelled thin rather than ranked as though it were proven. Validate on a trial melt before committing tooling.

R&D Acceleration

A screening prediction takes seconds. That changes which questions are worth asking — a design space that would have cost a year of trial melts can be narrowed before the first one is poured.

• Thousands of virtual conditions overnight, on your own compositions and routes
• Physics-aware search: chemistries that cannot exist are rejected, not returned
• The trade-off between strength, ductility, weldability and fatigue, visible at once
• Traceable results — which physics fired, what data supported it, and how far the input sits from anything in training

Where this sits: screening and process comparison, ahead of the laboratory rather than in place of it. Design allowables remain MMPDS and AMS. The longer goal is alloys that do not exist yet — the same search pointed at a blank slate instead of a known grade.

Who We Are

METALLAI sits between a metallurgist’s judgement and a laboratory schedule. We compute a physics baseline that states its own assumptions, correct it with machine learning trained only on measured data, and return the number together with how much of it to trust.

We are metallurgists and machine-learning engineers working from the KWORKS incubator at Koç University in Istanbul. The platform began with an observation any plant manager will recognise: the knowledge that decides a heat treatment usually lives in a few senior engineers’ heads, and it leaves when they do.

• Physics first, machine learning second. Strengthening mechanisms, transformation temperatures and defect-initiated fatigue are computed before any model runs; the model corrects what is left over.
• Open, citable training data. Roughly 830 steel and 466 aluminium mechanical records, each traceable to its source.
• Accuracy quoted on alloys held out of training. ±19 % on yield for an alloy the model has never seen. The in-sample figure is 6–7 % and we do not quote it — it measures fit, not accuracy.
• It declines rather than extrapolates. Outside the range it was fitted on, the platform says so instead of returning a confident number.

Request a Demo

See METALLAI V2.0 run on one of your own alloys, with your composition and your process route. Leave an email and company below and we normally reply within one working day with a session built around your part.