Lattice Graph × Presidio Ventures

Lattice Graph operates a computational materials discovery platform built around a governed knowledge graph that spans millions of compositions and traces each one from formula through crystal structure, thermodynamic and phonon stability, synthesizability, property predictions, patent claims, and experimental recipes. What makes this technically distinctive for a diligence audience is not a single model but a multi-engine validation architecture: candidate materials are independently evaluated by several machine-learning interatomic potentials including MACE and CHGNet, as well as by density functional theory calculations, and stability conclusions are reached only where multiple engines agree. That consensus requirement filters out the overconfident single-model predictions that inflate benchmark scores on positives-heavy public datasets, and it is precisely the kind of calibration signal a diligence team needs when underwriting a company whose value proposition rests on the accuracy of its property predictions. Layered on top of that physics infrastructure is a freedom-to-operate and patent-whitespace screening capability that covers more than 300,000 materials patents. When a startup presents a claimed-novel chemistry, Lattice Graph can map every composition in that claim space against existing patent filings at the level of composition and claim language — not as a simple keyword search but as a structured overlay on the knowledge graph — so that novelty assertions can be grounded in documented IP reality rather than taken from a pitch deck. This is especially consequential for AI-materials companies, where the most commercially valuable outputs are often the candidates nearest to existing patented territory, making an independent freedom-to-operate read a prerequisite before any capital commitment. Perhaps the most structurally important asset in the platform for technical diligence is what we call the negatives atlas: a curated collection of more than 23,000 labeled failed experiments and kill edges, representing the documented record of what does not work across battery chemistries, catalysts, separations materials, and other critical-minerals-adjacent domains. Most foundation-model-for-materials companies are trained and evaluated exclusively on public datasets that are heavily skewed toward positive, successful results, because failed experiments are rarely published. That systematic absence inflates apparent model performance. Lattice Graph's negatives corpus is largely internal and is not present in the public data sources that competing models or diligence providers can access, which means it is one of the few instruments that can produce an honest, hard-set evaluation of a target company's model quality — and flag candidate materials that already appear on the kill list.

Presidio Ventures occupies an unusual position in the deep-tech investing landscape: it is the U.S. corporate venture arm of Sumitomo Corporation, which means every check it writes carries a dual obligation. Presidio underwrites financial returns like any institutional investor, but it also evaluates strategic fit with a parent whose actual business lines span global commodity trading, critical-minerals flows, metals processing, and industrial infrastructure. An increasing share of its deal pipeline now runs through AI-for-materials and critical-minerals theses — battery electrolytes, catalyst discovery platforms, separations and recovery chemistries, and the generative-materials software companies that claim to discover them — and these are precisely the deals where the distance between a compelling deck and a technically grounded asset is hardest to read from the outside. The structural challenge is compounded by how AI-materials companies present themselves. A foundation-model-for-science company can credibly show a candidate hit rate, a benchmark comparison, and a discovery roadmap while leaving the load-bearing technical questions unanswered: Is the model's performance calibrated on a hard, negatives-inclusive dataset or inflated by a positives-heavy public benchmark? Is the claimed-novel chemistry genuinely free to operate, or does it sit on existing patent claims the company has not fully mapped? Are the feedstock and conversion-route assumptions in the critical-minerals supply thesis consistent with actual deposit concentrations and processing economics, or are they asserted at the level of narrative? These questions require independent technical infrastructure to answer, and they typically surface after capital is committed rather than before it prices. Lattice Graph is the independent computational layer designed to answer exactly those questions before a round closes. We do not compete with Presidio's portfolio companies; we produce the technical read that lets Presidio underwrite them with higher confidence. Our knowledge graph, multi-engine validation, negatives corpus, and freedom-to-operate screening capabilities are purpose-built for materials-science diligence, and they map directly onto the supply-chain and critical-minerals intelligence that Sumitomo's trading operations require. The result is a scoped, fast engagement that grounds a target company's core claims — model quality, IP novelty, supply-side feasibility, and commercial buyer identity — in data most AI-materials decks and most other diligence providers simply cannot reproduce.

Presidio Ventures occupies an unusual position in the deep-tech investing landscape: it is the U.S. corporate venture arm of Sumitomo Corporation, which means every check it writes carries a dual obligation. Presidio underwrites financial returns like any institutional investor, but it also evaluates strategic fit with a parent whose actual business lines span global commodity trading, critical-minerals flows, metals processing, and industrial infrastructure. An increasing share of its deal pipeline now runs through AI-for-materials and critical-minerals theses — battery electrolytes, catalyst discovery platforms, separations and recovery chemistries, and the generative-materials software companies that claim to discover them — and these are precisely the deals where the distance between a compelling deck and a technically grounded asset is hardest to read from the outside. The structural challenge is compounded by how AI-materials companies present themselves. A foundation-model-for-science company can credibly show a candidate hit rate, a benchmark comparison, and a discovery roadmap while leaving the load-bearing technical questions unanswered: Is the model's performance calibrated on a hard, negatives-inclusive dataset or inflated by a positives-heavy public benchmark? Is the claimed-novel chemistry genuinely free to operate, or does it sit on existing patent claims the company has not fully mapped? Are the feedstock and conversion-route assumptions in the critical-minerals supply thesis consistent with actual deposit concentrations and processing economics, or are they asserted at the level of narrative? These questions require independent technical infrastructure to answer, and they typically surface after capital is committed rather than before it prices. Lattice Graph is the independent computational layer designed to answer exactly those questions before a round closes. We do not compete with Presidio's portfolio companies; we produce the technical read that lets Presidio underwrite them with higher confidence. Our knowledge graph, multi-engine validation, negatives corpus, and freedom-to-operate screening capabilities are purpose-built for materials-science diligence, and they map directly onto the supply-chain and critical-minerals intelligence that Sumitomo's trading operations require. The result is a scoped, fast engagement that grounds a target company's core claims — model quality, IP novelty, supply-side feasibility, and commercial buyer identity — in data most AI-materials decks and most other diligence providers simply cannot reproduce.

For a diligence-focused team, the most useful surfaces in the Lattice Graph platform are the investigative and screening dashboards rather than the chemistry-authoring workflows. The opportunity-index and buyer-intelligence views translate the ranked opportunity data into an interactive analyst interface: a user selects a chemistry lane relevant to a deal under review, sees ranked opportunity scores alongside a funded-buyer leaderboard and a shortlist of highest-conviction candidate spaces, and can immediately check whether Sumitomo-adjacent buyers appear in that commercial landscape. The negatives-atlas dashboard allows analysts to submit a target company's claimed candidate materials in batch and receive a structured kill-edge report — a concrete, auditable artifact that shows which candidates are already documented as failures and how much negative-result coverage exists in the surrounding chemistry space. These two surfaces together convert what is currently a judgment call in AI-materials diligence into a data-grounded deliverable. The knowledge-graph explorer provides provenance and evidence-neighborhood views that let an analyst trace any material or property claim from a deal memo back to its underlying sources — DFT calculation provenance, cross-source agreement and disagreement flags, recipe and experimental coverage — so that calibration questions can be answered at the level of individual claims rather than evaluated impressionistically. The freedom-to-operate and patent-whitespace dashboard supports composition- and claim-level IP screening against the full 300,000-plus patent corpus, providing a structured overlay on novelty assertions that is far more rigorous than keyword search. For deals with a critical-minerals or feedstock angle, the supply-chain intelligence views render deposit, concentration, and conversion-route data in a form that Sumitomo's trading teams would recognize — allowing Presidio to stress-test supply-side assumptions against real-world mineral economics before those assumptions are baked into a deal structure.

The natural engagement structure for Presidio is a scoped, fixed-scope technical diligence read on a single active deal — chosen to test the platform's value against a deal where Presidio has already formed a view, so findings can be calibrated against internal judgment. A single-deal engagement covers a kill-edge check on the target's candidate materials using the negatives atlas, an opportunity-index and buyer-affinity read on the target's claimed commercial lane, a freedom-to-operate screen on the target's headline IP, a cross-source trust and calibration assessment on the target's reported property predictions, and a supply-side stress test where the deal involves critical-minerals feedstock or conversion claims. The deliverable is a buyer-grade technical memo that addresses each of these dimensions with documented evidence rather than narrative assertion. Per-deal diligence engagements of this type are estimated in the range of approximately $30,000 to $60,000, consistent with a single-round technical read. For a corporate venture fund underwriting multiple AI-materials and critical-minerals deals per year, the more efficient structure is an annual diligence subscription that provides API access to both the opportunity-and-buyer-intelligence and negatives-atlas data products, along with a defined number of scoped per-deal reads with knowledge-graph, freedom-to-operate, and supply-intelligence layers available across engagements. The recommended entry point is one paid pilot on a live or recently-passed deal — ideally one with a visible Sumitomo strategic-fit angle — converting to the subscription structure on successful calibration. All figures cited here are planning-level estimates and not binding quotes; actual scope and terms are set per engagement.