Company facts · updated
Lupine Science at a glance
Lupine Science is an early-stage deep-tech company building a verification layer for AI-driven materials discovery. It measures the structured error of machine-learned interatomic potentials (MLIPs), corrects that error at runtime as a physical field beside the model, and seals its quantitative claims as machine-checked Lean 4 theorems over provenance-hashed data, in public.
- Company
- Lupine Science
- What it does
- Lupine Science makes AI predictions for new materials trustworthy before a lab spends a year on them.
- Stage
- Early stage, founder-led
- Funding
- Funded and not raising. Closed to further investors as of October 2026.
- Founder
- Alex Welcing, founder
- Next public appearance
- TMS2027 Annual Meeting & Exhibition, Orlando · 14–18 March 2027 · poster presentation
- Open surfaces
- Articles · Library · LUPI viewer · Source · Feed
How it works
- Measure
- Prediction errors of interatomic potentials are not random. In the committed benchmark they collapse onto a low-dimensional ribbon, and the dominant error directions are largely shared across potentials.
- Correct
- A measured error field runs beside an existing MLIP as an additive energy term with analytic forces. No fine-tuning, no retraining.
- Prove
- Quantitative claims are sealed as Lean 4 theorems over provenance-hashed data. Corrections are licensed only where evidence supports them; unsupported cases abstain.
Where to start
- Analysts and investors following the field. Start with the founding thesis, the Savings Stack booklet, and the environmental-series proof pack. Every number links to the committed data it was computed from. Start here →
- Materials scientists. Start with the live evidence and the benchmark it is computed from, including the corrections that failed. Start here →
- MLIP and simulation teams. Start with the instrument: inspect structures and error fields in the browser. Start here →
Milestones
- — Poster abstract accepted for TMS2027, Orlando
- — Small-cell MLIP elastic benchmark release archived; DOI registration pending
- — Shared DFT anchors proof pack published
- — The Savings Stack: free 16-page booklet with a hashable data pack
- — Four preregistered Round-4 campaigns reported, failures included
- — Environmental series and climate proof pack released
- — The Trust Layer: the founding thesis
- — Live evidence: MLIP rankings hold across 36 model–material cells
- — Why LUPI: the browser-native evidence viewer, explained
Latest notes
- — Learning What an Error Pattern Can Tell Us Published
- — Lupine Science is heading to TMS2027 Poster abstract accepted
- — What the Cap Was Hiding Draft for discussion
- — Sharing DFT anchors reduces projected evaluations by 72.4% Published
- — The Small Cell Held: A 16-Element MLIP Elastic Benchmark In review
- — The Union Verdict: A De-brief on the Z1 Campaign FOR EDITOR REVIEW
All articles → · Subscribe to the feed →
Common questions
What does Lupine Science do?
Lupine Science is an early-stage deep-tech company building a verification layer for AI-driven materials discovery. It measures the structured error of machine-learned interatomic potentials (MLIPs), corrects that error at runtime as a physical field beside the model, and seals its quantitative claims as machine-checked Lean 4 theorems over provenance-hashed data, in public.
What problem does it solve?
Generative models and machine-learned potentials now propose and screen far more candidate materials than any lab can synthesize. The bottleneck has moved from generating candidates to knowing which predictions to trust. A prediction checked before the bench saves the year a lab would otherwise spend finding out.
How is this different from building another interatomic potential?
Lupine Science does not compete with model builders. It sits between model generation and laboratory validation: model builders get external verification, and laboratories get shortlists with characterized failure modes. The correction runs beside an existing model, with no fine-tuning or retraining.
Is the work peer-reviewed?
Not yet. Every article carries its own status label (draft, final, live evidence), and every quantitative claim links to the committed data and the machine-checked Lean theorem it rests on, including the corrections that failed.
What stage is the company at, and is it raising?
Early stage, founder-led. Funded and not raising. Closed to further investors as of October 2026. Talking with materials labs, MLIP builders, and research partners ahead of TMS2027.
Where can I meet the team?
At TMS2027 Annual Meeting & Exhibition in Orlando, 14–18 March 2027: poster presentation in Accelerating Innovation in Materials and Manufacturing. Write ahead to set a time.
Contact
Talking with materials labs, MLIP builders, and research partners ahead of TMS2027. Write to alex [at] lupinesci.com.