Alloy R&D screening

A ranked shortlist for the next experiments.

We sell a decision, not a number. Send a backlog of candidate compositions. We return a make-and-test shortlist with uncertainty, refusal flags and traceable evidence, so the team can decide what to run next.

The deliverable

Choose what to make and test next.

Experimental R&D is usually limited by test capacity. The service helps narrow a large candidate set into a smaller, reviewable shortlist.

Send candidates

A list or design space: compositions and, where known, process routes.

We screen and rank

Each candidate is scored, checked against the evidence, and ordered by make-and-test priority. Unsupported cases are flagged instead of forced into a confident number.

You get the evidence

A shortlist with per-candidate intervals, refusal verdicts, feature attributions and provenance for internal review.

The aim is not to replace experiments. It is to spend the next experiments better.

Ways to work with us

Start small; deepen if the data supports it.

Begin with one scoped screening engagement. If the workflow is useful and the data is strong enough, move to retained screening or a private model.

Start here

DOE Screening Engagement

Fixed fee · per engagement
  • One candidate list or design space
  • Ranked "make and test first" shortlist
  • Evidence pack: uncertainty, OOD verdict, attributions, provenance
Scope an engagement
Ongoing

Retained Screening

Recurring · monthly or per-project
  • For labs that screen continuously
  • A standing arrangement over the same governed core
  • Priority turnaround and a shared evidence history
Talk to us
Private model

Private Single-Family Model

Annual licence
  • Trained on your own process / DOE data
  • Governed model delivered under licence
  • Specific to your alloy family and data history
Discuss a roadmap

Why it is reviewable

Built-in review checks.

The product is built around the questions a technical lead will ask: how uncertain is this, is this input supported, and where did the data come from?

Calibrated uncertainty

Predictions include intervals whose error rate is measured and reported conservatively.

Fail-closed OOD gate

Candidates outside the evidence are refused or given wider estimates instead of confident unsupported numbers.

Governed promotion

Model variants are promoted only when they beat the incumbent on leakage-aware tests.

Human-reviewed curation

The data pipeline is rule-based, reviewed by a person, and keeps provenance on every row.

Honest scope

Screening, not qualification.

On a leakage-aware, grouped hold-out, the high-entropy-alloy hardness screener reaches R² ≈ 0.84 / MAE ≈ 54 HV (53.4 HV under a clean no-peek selection). That supports screening and ranking, not qualification-grade point prediction.

Processing route is used as a conditioning feature, not a learned process model. Capturing processing effects properly requires customer pilot data.

Governance, in practice

We tested an additive-manufacturing-specific challenger model and declined to promote it because it did not beat the baseline on the available data. That decision is kept as part of the model record.

What we do

  • Ship models with measured, conservative coverage
  • Refuse to answer outside the evidence
  • Rank and recommend with an evidence trail
  • Provide record-level provenance and an audit log
  • Keep a human in the loop on curation

What we don't claim

  • A "validated optimiser" or inverse design
  • A qualification or specification tool
  • "Self-learning" or autonomous AI
  • Accuracy leadership from a single black-box number
  • Answers on inputs the evidence doesn't support

Get in touch

Have a candidate backlog to screen?

Tell us the decision you need to make. We will assess whether the data can support a useful screening engagement.