ML-Guided Materials Selection

Using machine learning to rank candidate materials against many coupled requirements before slow, costly testing.

The selection problem

A plasma-facing or structural material must satisfy many requirements at once: high-temperature strength, tolerance to neutron damage, low activation, compatibility with coolant and tritium, and manufacturability. No single material wins everything. Selection is multi-objective ranking over a large candidate space with sparse, expensive experimental data.

Where ML helps

• Screening: predict properties from composition and processing to shortlist candidates
• Property prediction: fill gaps where measurements are missing
• Active learning: propose the next material to test to reduce uncertainty fastest
• Trade-off mapping: build Pareto fronts across competing properties

The data reality

Fusion-relevant irradiation data are scarce because test facilities and exposure times are limited. Models trained on thin data must report uncertainty honestly and avoid extrapolating confidently past their training range. A prediction outside the data envelope is a hypothesis for testing, not a conclusion.

def shortlist(candidates, predict, need):
    ranked = []
    for c in candidates:
        p, sigma = predict(c)          # property + uncertainty
        if p - sigma >= need:          # conservative: use lower bound
            ranked.append((c, p, sigma))
    return sorted(ranked, key=lambda r: -(r[1]-r[2]))

Kronos context

Reduced-activation steels and high-temperature superconductors (REBCO tape for the magnets) sit at the center of material choices for both the Hyperion breeder and the burner. ML narrows the search and orders the test queue; measured data and physics-based degradation models make the final call. The pattern is decision support, not decision replacement.

Failure modes to watch

Correlations in a materials database can be spurious; a model can learn the lab that made the sample rather than the physics. Guarding against this needs held-out tests, physical sanity checks, and skepticism toward any ranking that lacks a mechanistic reason.