AIsteels – Materials Intelligence Stack

Skip to content

Materials Intelligence Stack

Operators and developers in extreme-environment industries make costly material decisions — run/retire, coating replacement, alloy down-selection — mostly on experimental data alone, slowly and expensively. AIsteels compresses that process: three physics-based tools (Austenitic, Ferritic, HEAs) share a common Bayesian uncertainty-quantified approach so that a single experimental campaign updates all relevant modules simultaneously, and blind predictions are lodged before exposure, not after.

The three tools

Physics-informed decision supports that rank options, quantify uncertainty, and identify the next experiment expected to deliver the greatest information gain. From today (a coating on existing alloys) to tomorrow (better bulk materials with or without coatings) and beyond (new alloys for the most extreme environments). The real asset is the experimental prioritisation and uncertainty-tracking layer: the framework aims to reduce experimental campaigns in a measurable, auditable way.

Austenitic + coating

A physics-first coating lifetime optimiser built originally for FeCrAl laser cladding on 316L in LBE nuclear service at 600°C.

The core engine uses Arrhenius corrosion kinetics, Miner’s rule damage accumulation, parabolic oxide growth, Griffith spallation sigmoid, and Bayesian Sobol UQ. The architecture is clean, modular, and genuinely reusable. The LBE nuclear case has a coherent plan and a named experimental facility.

Three adjacent case studies — Oil & Gas HP/HT, supercritical CO2 Brayton cycles, and CSP thermal barrier coatings — have been added with correct physics structure but no experimental grounding whatsoever. It is one validated-pathway tool plus three physics-complete sketches.

Ferritic, FM and ODS

A physics-based digital twin for 9-12Cr steels that reduces experimental iteration.

1) Forward design: give it a composition + service condition → it returns properties, degradation state, creep life, risk scores.

2) Inverse design: give it a target property profile → it searches composition space and returns the alloys that hit it.

3) Multi-objective optimisation: give it competing targets (maximise creep life, maintain toughness, etc.) → it returns the Pareto front of compositions.

4) Classification/screening: it flags which degradation mechanisms are active, ranks candidates, and scores HAZ risk before experimental work.

Ni-based and HEAs

Already validated as a screening and Pareto tool for conventional Ni-base alloys programmes.

Relative ranking is valid even uncalibrated, down-selecting from a compositional space of millions to tens. Inverse design + Pareto can identify the Pareto front across competing objectives using only the physics model. Bayesian optimisation starts from the physics model as a prior mean (not a flat prior) and interpolates between sparse data points.

For compositions such as, say, a Mo-Nb-Ta-W-Zr refractory HEA, the model will give physically self-consistent but quantitatively unreliable absolutes until doing roughly 5-8 targeted experiments covering the T/σ extremes of intended service.

The framework status

Surface coating degradation, conventional FM steel remaining life, HEA compositional design: physical complexity and experimental cost scale in that order, and so does the framework. All three share a consistent Bayesian UQ architecture with explicit applicability domain boundaries and a targeted experimental ask for each unanchored term. Calibration data from a single campaign can inform multiple tools through sequential, traceable handoffs. Case study #4 demonstrates the integrated output: ODS versus conventional FeCrAl for LBE fast reactor cladding, decomposed by failure mode, with the dominant uncertainty term and its resolving measurement identified before the coupon programme begins.