Investor presentation · July 2026

The trust layer for climate-critical materials.

AI is designing the matter the energy transition runs on. Every prediction rests on interatomic potentials that are wrong in structured ways. Lupine measures, proves, and corrects that wrongness — so climate capital funds validated materials, not guesses.

Addressable climate impact5–12 GtCO₂/yr
Build-locked theorems899
Validation rate today0.2%
01 lupine.science
Sources: IEA Net Zero Roadmap; Merchant et al., Nature 2023 (GNoME); Lupine Science Lean 4 library.
The climate window

Capital must roughly triple — and materials decide whether it works.

The IEA projects clean-energy investment must grow from $1.8 trillion in 2023 to $4.5 trillion a year by the early 2030s. Batteries are directly linked to ~20% of required 2030 CO₂ reductions and indirectly to another 40%. The binding constraint is no longer generating candidates; it is knowing which predicted materials can actually be made, at cost, inside the 2025–2035 deployment window.

Clean-energy investment$1.8T → $4.5T
2030 CO₂ reductions, battery-linked~60%
Deployment window2025–2035
02 lupine.science
Sources: IEA Net Zero Roadmap 2023 update; IEA Global EV Outlook.
The hidden bottleneck

Generation is solved. Validation is the 0.2% problem.

GNoME predicted 2.2 million crystals; 380,000 were computed stable; only 736 had been independently synthesized by late 2023 — a 0.2% validation rate. A-Lab's Author Correction records 36 confirmed of 57 eligible targets, 4 inconclusive, and one compound removed; independent review separately found many “novel” targets were already-known disordered phases. Faster synthesis without better verification does not close the gap.

Investor read

We are not short of predictions. We are short of validated predictions. Every false positive burns synthesis budget and calendar time the climate window does not have.

Synthesis funnel: structures generated to commercially viable
03 lupine.science
Sources: Merchant et al., Nature 2023 (GNoME); Szymanski et al., Nature 2023 (A-Lab); subsequent replication critiques.
The shape of wrongness

Errors are not noise — they collapse onto a manifold.

Across 15 elements and hundreds of potentials, prediction errors concentrate on a low-dimensional surface. Classical and foundation MLIPs fail in the same direction. That is the investment thesis: if every model is wrong the same way, one correction layer transfers across the entire stack.

Participation ratio1.05–2.05
Cross-MLIP cosine>0.8
Geometry preserved14/15
Cross-MLIP error direction alignment by element
04 lupine.science
Sources: Lupine Science MLIP Failure Geometry Audit; Deng et al., npj Computational Materials 2025.
What Lupine does

Measure the field. Correct at runtime. Prove what is supported.

Measure

Three anchor observables fix an error field over local atomic environments. Blind prediction achieves Pearson r = 0.906 (p = 10⁻⁴, 95% CI [0.82, 0.96]) across 36 (model, material) combinations.

Correct

An analytic overlay applies the correction inside LAMMPS with any uMLIP — CHGNet, MACE, Orb, and what comes next. 15.6% overhead today, <1% compiled, still ~10⁵× faster than DFT.

Prove

The machine-generated Lean 4 theorem inventory, with zero sorry proofs, states what is supported. Where correction cannot apply, the system proves impossibility or bounded uncertainty instead of reporting a p-value.

05 lupine.science
Sources: Lupine Science error-field formalization; Lean 4 build locks; Deng et al., npj Computational Materials 2025.
New this month

An assurance spine investors can audit.

Every corrected claim now ships as an executable certificate. The universal correction spine formalizes scope, residual bounds, numeric and runtime contracts, trajectory validation, and scientific attestation in Lean 4 — and mirrors the same fail-closed gates in Rust with CLI validators. Discovery gates and licenses turn correction coverage into a diligence artifact across candidate panels for cathodes, halide electrolytes, MOFs, and perovskites.

Lean 4 formal modules20+
Rust policy runtimefail-closed
Candidate campaign rounds1–3
06 lupine.science
Source: lupine-rhizo main — UniversalCorrection module + atlas-distill policy runtime, July 2026.
Where the value lands

Five targets, 5–12 GtCO₂ per year.

TargetClimate impactMarket contextKey barrierLupine mechanism
Cobalt-free LMR cathodes2–5 GtCO₂$400B+ Li-ion chainVoltage fade; TM migrationCorrected migration barriers preserve mobility ranking
Halide solid electrolytes1–3 GtCO₂$886M → $24.3B by 2034Li⁺ barriers underestimated 60%+Corrected barriers recover DFT accuracy at screen scale
MOFs for direct air capture0.5–2 GtCO₂/yr$4.3B DAC market by 2034Humidity stability; cost >$50/kgCorrected hydrolysis barriers; impossibility proofs
Electrochemical ammonia~0.45 GtCO₂/yr$221.6B green NH₃ by 2035N≡N activation; HER competitionCorrected N₂ dissociation; flags scaling-relation breakers
Lead-free perovskites0.5–1 GtCO₂/yr$11B PV market by 2033Sn²⁺ oxidation; metastabilityCorrected vacancy formation energies; provable boundaries
Combined climate impact5–12 GtCO₂/yr
Combined market context>$650B
07 lupine.science
Sources: IEA, IPCC AR6, and market analyses for Li-ion, solid-state batteries, DAC, green ammonia, and PV.
The business model

A profitable floor, an uncapped ceiling.

Floor

Simulation-trust software and calibration services for labs, AI-for-science teams, and industrial R&D — a revenue base that pays for the manifold to expand.

Ceiling

Materials-IP discovery and licensing driven by the corrected signal across the five climate targets — the asymmetric upside investors are actually underwriting.

Evidence surface

LUPI, the browser-native WebGPU viewer, makes every claim inspectable. Partners and auditors can open the same atoms, fields, and certificates we do.

08 lupine.science
Sources: Lupine Science strategy; lupi.live.
The moat

Proof-grade coverage compounds. Data moats decay.

Three pillars make the position defensible. The error field is measured, not learned — fixed by observables and transferable without per-system retraining. The claims are machine-checked, so they cannot be hand-waved around. The runtime is compatible with every major uMLIP, so the layer rides the whole field’s progress. Every screening campaign adds validated field measurements; every impossibility proof sharpens the boundary of applicability; every theorem widens the verified domain.

Measured fieldblind transfer
Machine-checked proofzero sorry
Runtime compatibilityany uMLIP
09 lupine.science
Source: Lupine Science formalization library and runtime overlay.
The economic case

The correction layer sits where value leaks.

NIST’s Materials Genome Initiative economic analysis estimates $123B–$270B in annual value from improved materials innovation infrastructure. ARPA-E’s $3.5B portfolio catalyzed $11.8B in private follow-on funding. Lupine sits between generation and synthesis — the point where speed without accuracy turns into wasted experiments, and where a trusted correction converts directly into saved lab budget and faster deployment.

Annual value at stake$123–270B
ARPA-E leverage$3.5B → $11.8B
Addressable markets>$650B
10 lupine.science
Sources: NIST MGI economic analysis; ARPA-E portfolio data.
Go-to-market

We sell certainty to the people who already spend the synthesis budget.

National labs

NREL spans four of five targets and is the highest-value master-CRADA candidate. Argonne and PNNL’s Battery500 anchor cell-level validation.

Universities

UC Berkeley (Ceder, Yaghi, Long), UT Austin (Manthiram), DTU (Chorkendorff), and Stanford (SUNCAT) provide the experimental credibility layer.

Industrial

GM/Ultium, POSCO Future M, BASF, Solid Power, and Climeworks pull validated materials into cells, sorbents, and deployed systems.

11 lupine.science
Source: Lupine Science partnership map.
The ask

$20M to put materials intelligence in every major lab on earth.

Deliver Lupine’s correction and verification layer to all major supercompute sites and university + industry labs within 12 months; build the founding team; expand the manifold across bcc, hcp, and layered structures.

Raise~$20M
Global delivery12 months
Climate impact5–12 GtCO₂/yr

founder@lupinesci.com · lupine.science · lupi.live

12 lupine.science
Confidential · July 2026 · Lupine Science