Community exchange

A Technical Exchange for Metal AI Builders

Built for metallurgists, process engineers, researchers, data scientists, plant teams, students, and AI builders. Trending discussions, open research questions, dataset requests, and monthly challenges, all organized by metal, process, and topic.

Monthly challenge

July: improve the rare earth recovery baseline

Beat the reference MAE on the open split. Reference and research submissions welcome.

View challenge
  • Open research question
    6123

    Which features most improve stage-loss attribution in SX trains?

    On the synthetic separation set, adding phase-ratio dynamics beat static pH features for attributing loss to a stage. Curious what others have found on real circuits.

    k.almeida · Process data scientist · 2d ago

  • Dataset request
    3411

    Looking for anonymized aluminium scrap assays with cast-quality labels

    Happy to contribute synthetic augmentation if anyone has non-confidential scrap stream assays paired with downstream quality outcomes.

    r.okafor · Digital product lead · 4d ago

  • Benchmark proposal
    4517

    Proposal: add calibration to the lithium impurity-risk benchmark

    ROC-AUC alone rewards ranking, not calibrated probabilities. For a qualification decision, Brier + reliability should be first-class. Draft rules attached.

    p.nakamura · Researcher · 5d ago

  • Discussion
    2914

    Slag chemistry features that generalize across furnace types

    Fe/SiO₂ and magnetite indicators transfer reasonably; absolute temperatures don't. Anyone normalized furnace features across configurations?

    l.zhang · Process engineer · 1w ago

  • Discussion
    269

    Oxygen thresholds for qualification: one global cut or per-alloy?

    The reference classifier uses a configurable threshold. In practice we set it per alloy family, so should the benchmark reflect that?

    d.fischer · Powder R&D · 1w ago

  • Open research question
    7231

    How far do tabular baselines get us before physics-informed models pay off?

    Across the recovery benchmarks, gradient boosting is within a point or two of fancier models. Where does physics-informed structure actually earn its keep?

    s.iyer · ML researcher · 2w ago

Contributor highlights

KA

K. Almeida

Top contributor · Rare earths

148
SI

S. Iyer

Benchmark lead · Methods & benchmarks

132
LZ

L. Zhang

Domain expert · Copper

119
PN

P. Nakamura

Schema author · Lithium

104
RO

R. Okafor

Dataset contributor · Aluminium & recycling

92
HN

H. Novak

Model author · Separation models

87
Contribute to open metal AI

Bring your process knowledge to the exchange

Ask a question, contribute a reference model, request a dataset, or propose a benchmark. Open industry knowledge and private customer advantage are kept separate.