Lattice Graph × Felicis

Lattice Graph is a computational materials-discovery platform built around a knowledge graph that spans millions of compositions, connecting structure to property to patent to synthesis route in a single governed graph. What makes the platform technically distinct for any investor evaluating materials companies is multi-engine validation: every candidate material is assessed by multiple independent physics engines simultaneously, including machine-learning interatomic potentials such as MACE and CHGNet alongside first-principles density functional theory calculations. Phonon stability and thermodynamic stability must reach consensus across those independent engines before a composition advances. This is not a single-model opinion; it is an adjudicated result, and that distinction matters enormously when you are trying to determine whether a startup's "validated" compounds will survive independent scrutiny. The platform also carries two assets that almost no AI-materials company can replicate in a short pitch cycle. The first is comprehensive freedom-to-operate and patent-whitespace screening across more than 300,000 materials patents. That gives Felicis a direct, evidence-backed answer to the defensibility question: does this company's claimed invention actually sit in open whitespace, or is it already inside a competitor's filed claims? The second is a labeled corpus of more than 23,000 failed-experiment edges — materials and routes that have been tested and do not work — organized and queryable by domain. Because public materials databases are almost entirely composed of positive results, a model trained on public data alone has never been stress-tested against real failure modes. Lattice Graph's negatives corpus is the stress-test that exposes that gap. For a venture firm doing technical due diligence on AI-materials and critical-minerals rounds, these capabilities collapse into three precise questions: Is the science real and independently reproducible? Is the IP position defensible and uncrowded? And does a credible buyer exist at a price that makes the thesis work? The platform is built to answer all three with provenance-backed evidence that traces to source data, not to the company's own deck.

Felicis has built a genuine and differentiated thesis around AI-for-science and frontier deep tech, including an explicit focus on AI-materials and critical-minerals companies. That thesis is intellectually sound and well-timed. The challenge it creates, though, is asymmetry: AI-materials decks are consistently polished, and nearly every one claims a proprietary dataset, a generative model trained on private experimental data, and a roadmap of novel compounds validated by internal DFT. The claims are difficult to falsify quickly, and the failure modes are technical rather than commercial, meaning the conventional diligence toolkit — team quality, cap table, comparable exit multiples — does not reach the right layer of risk. A model that overfits to the Materials Project, a "discovery" already claimed in a 2019 patent filing, or a battery chemistry whose performance numbers rest on a single noisy computational source all look identical to a polished deck. Felicis needs an independent instrument for exactly that read, and it needs one it can deploy on the timeline of a live round. Lattice Graph is purpose-built to be that instrument. We hold no equity position in any company under evaluation, which means our output is independent and usable in front of a full investment committee and co-investors. The knowledge graph spans the same composition space the target companies are working in — we can place their named compounds in context, check provenance, flag cross-source property disagreements, and run the claimed roadmap against the full patent landscape in a single workflow. The result is not a qualitative expert opinion; it is evidence tied to source data, with a clear audit trail. For a firm that co-invests with other leading funds where IC standards are high, that provenance is as valuable as the conclusions. The strategic fit also holds for the firm's broader portfolio construction. When Felicis takes a position in an AI-materials company, ongoing access to the platform gives portfolio companies a validation and IP-screening resource that strengthens their defensibility story for subsequent rounds. And because the platform covers critical-minerals supply chains — element-level concentration risk, deposit geography, conversion-route economics — it supports the energy-storage and critical-minerals threads of the firm's thesis at the same depth as the AI-materials thread.

Felicis has built a genuine and differentiated thesis around AI-for-science and frontier deep tech, including an explicit focus on AI-materials and critical-minerals companies. That thesis is intellectually sound and well-timed. The challenge it creates, though, is asymmetry: AI-materials decks are consistently polished, and nearly every one claims a proprietary dataset, a generative model trained on private experimental data, and a roadmap of novel compounds validated by internal DFT. The claims are difficult to falsify quickly, and the failure modes are technical rather than commercial, meaning the conventional diligence toolkit — team quality, cap table, comparable exit multiples — does not reach the right layer of risk. A model that overfits to the Materials Project, a "discovery" already claimed in a 2019 patent filing, or a battery chemistry whose performance numbers rest on a single noisy computational source all look identical to a polished deck. Felicis needs an independent instrument for exactly that read, and it needs one it can deploy on the timeline of a live round. Lattice Graph is purpose-built to be that instrument. We hold no equity position in any company under evaluation, which means our output is independent and usable in front of a full investment committee and co-investors. The knowledge graph spans the same composition space the target companies are working in — we can place their named compounds in context, check provenance, flag cross-source property disagreements, and run the claimed roadmap against the full patent landscape in a single workflow. The result is not a qualitative expert opinion; it is evidence tied to source data, with a clear audit trail. For a firm that co-invests with other leading funds where IC standards are high, that provenance is as valuable as the conclusions. The strategic fit also holds for the firm's broader portfolio construction. When Felicis takes a position in an AI-materials company, ongoing access to the platform gives portfolio companies a validation and IP-screening resource that strengthens their defensibility story for subsequent rounds. And because the platform covers critical-minerals supply chains — element-level concentration risk, deposit geography, conversion-route economics — it supports the energy-storage and critical-minerals threads of the firm's thesis at the same depth as the AI-materials thread.

For a venture diligence team, the platform's most useful daily surface is the composition-intelligence and batch-screening workflow used as a diligence cockpit. A partner or technical associate drops a target company's named compositions into the composition-intelligence view and immediately sees provenance, cross-source property values with trust-and-disagreement flags, and patent-whitespace context in one place. Running the full claimed roadmap through batch screening takes the same workflow to portfolio scale: how many of the startup's "breakthroughs" are genuinely novel, how many are already claimed in a patent, and how many land in the labeled negatives set? That question, answered in evidence rather than opinion, is the core of a credible technical-read memo. The opportunity and buyer intelligence dashboards and the negatives and eval-data views are the two screens that most directly answer Felicis's deal questions. They convert the ranked opportunity index, funded-buyer affinity scores, and the full kill-edge corpus into exportable evidence that can flow directly into an investment memo or IC presentation. Because every data point in the knowledge graph carries a source trace, every claim in that memo is auditable, which is exactly what a co-investor or a follow-on partner will ask for when the firm takes a conviction position in a technically complex company.

The right structure for Felicis is a scoped per-target diligence sprint rather than a broad portfolio license. Felicis names a company under active evaluation and we deliver an independent technical-read memo covering four dimensions: novelty and opportunity whitespace (where the target's claimed compositions sit on the ranked opportunity index and how much uncrowded headroom is real), defensibility (freedom-to-operate screening across the full patent landscape, with claim-level flagging), data-quality integrity (cross-source property trust scores and negatives-corpus hit rate for the target's key compounds), and commercial pull (funded-buyer affinity and a ranked leaderboard of credible acquirers or offtakers for the target's specific material class). All findings are sourced to the platform, so the memo is independently auditable. Based on the context for this engagement, scoped reads land in approximately the $30,000 to $60,000 range per target, with meaningful discounts structured for a multi-target commitment across a thesis or a quarter's deal flow. For a firm with recurring AI-materials and critical-minerals deal flow, a diligence subscription is the cleaner structure: a fixed quarterly or annual fee covering a set number of target reads plus standing access to the Opportunity and Buyer Intelligence and Negatives and Eval-Data Atlas products, so associates can pre-screen inbound decks and identify obvious disqualifiers before pulling in partner time. In all cases, Lattice Graph takes no equity position in any evaluated company, preserving the independence of the read for use with co-investors and across the IC.