Lattice Graph × Accel

Lattice Graph operates a computational materials-discovery platform built around a knowledge graph that spans millions of compositions, connecting formula to crystal structure to measured and predicted properties, to synthesis routes, and to the patent record. That graph is not a literature scraper — it is a governed, provenance-tracked network where every edge carries a source citation and a trust score derived from agreement across independent calculations. When a property prediction appears in the graph, analysts can see whether it is corroborated by multiple independent sources or whether it sits on a single, potentially overfit model run. Validation inside the platform relies on multiple independent physics engines running in parallel. Machine-learning interatomic potentials, including MACE and CHGNet, provide rapid thermodynamic and phonon stability screening; those results are then checked against density functional theory calculations to build consensus. Candidates that survive all engines are structurally and thermodynamically defensible, not just predicted. Candidates that fail or produce conflicting signals are recorded as such — and that record of disagreement is itself a diligence signal. The third differentiating layer is what Lattice Graph knows about failure. The platform holds a labeled atlas of more than 23,000 failed-experiment and kill edges — negative results that never make it into public datasets or benchmark sets. Most AI-for-materials models are trained and evaluated on positives-heavy corpora that systematically overstate hit rates because the dead ends were never labeled. That asymmetry is exactly the gap Lattice Graph closes for a technical reader who needs to know whether a reported model score was earned on a genuinely hard, negatives-inclusive benchmark or inflated on public data the model has essentially memorized.

Accel's "prepared mind" approach requires getting deeply knowledgeable about a category before a round prices, and that discipline has served the firm well across software, infrastructure, and consumer internet — domains where diligence can be anchored in metrics, references, and comparable cohorts. The growing slice of inbound in AI-for-science, foundation models for chemistry, and critical-minerals processing represents a structurally different challenge. The question on a deep-tech term sheet is not whether the team can ship or whether the market is large; it is whether the underlying science is real, whether the training data is honest, and whether the claimed IP position survives contact with a materials-patent corpus. Those are questions that require a purpose-built technical instrument, not a generalist consultant. The pressure has tightened because valuations in AI-materials and critical-minerals are moving faster than the tooling to verify them. A foundation-model-for-chemistry deck competes on candidate hit-rate claims, benchmark scores, and training-data differentiation — almost all of which are unverifiable by a software-oriented investment committee without a reference dataset that includes failures. Critical-minerals and energy-storage decks bundle electrochemistry, conversion-route economics, feedstock criticality, and freedom-to-operate claims that a generalist IC is not staffed to disentangle. A single misread of technical defensibility is expensive for Accel — financially and reputationally, given that the firm's positioning in deep-tech depends on having actually gotten smart on the category rather than having pattern-matched a narrative. Lattice Graph functions as the independent technical-diligence layer for exactly these rounds, with no competing portfolio overlap. The platform's opportunity-and-buyer intelligence index provides a third-party ranked view of what is genuinely inventable in a target's chemistry lane, independent of the roadmap on the pitch deck. The negatives atlas provides the one instrument that most AI-materials decks cannot reproduce from their own data: a test against 23,000-plus labeled failure edges that surfaces whether a claimed hit rate holds up when the hard negatives are actually included. Together they convert a trust-the-founder memo into a trace-it-in-the-graph technical exhibit.

Accel's "prepared mind" approach requires getting deeply knowledgeable about a category before a round prices, and that discipline has served the firm well across software, infrastructure, and consumer internet — domains where diligence can be anchored in metrics, references, and comparable cohorts. The growing slice of inbound in AI-for-science, foundation models for chemistry, and critical-minerals processing represents a structurally different challenge. The question on a deep-tech term sheet is not whether the team can ship or whether the market is large; it is whether the underlying science is real, whether the training data is honest, and whether the claimed IP position survives contact with a materials-patent corpus. Those are questions that require a purpose-built technical instrument, not a generalist consultant. The pressure has tightened because valuations in AI-materials and critical-minerals are moving faster than the tooling to verify them. A foundation-model-for-chemistry deck competes on candidate hit-rate claims, benchmark scores, and training-data differentiation — almost all of which are unverifiable by a software-oriented investment committee without a reference dataset that includes failures. Critical-minerals and energy-storage decks bundle electrochemistry, conversion-route economics, feedstock criticality, and freedom-to-operate claims that a generalist IC is not staffed to disentangle. A single misread of technical defensibility is expensive for Accel — financially and reputationally, given that the firm's positioning in deep-tech depends on having actually gotten smart on the category rather than having pattern-matched a narrative. Lattice Graph functions as the independent technical-diligence layer for exactly these rounds, with no competing portfolio overlap. The platform's opportunity-and-buyer intelligence index provides a third-party ranked view of what is genuinely inventable in a target's chemistry lane, independent of the roadmap on the pitch deck. The negatives atlas provides the one instrument that most AI-materials decks cannot reproduce from their own data: a test against 23,000-plus labeled failure edges that surfaces whether a claimed hit rate holds up when the hard negatives are actually included. Together they convert a trust-the-founder memo into a trace-it-in-the-graph technical exhibit.

The most useful platform surfaces for Accel's diligence workflows are the knowledge-graph explorer and the opportunity-index dashboards rather than the chemistry-authoring or synthesis-planning tools. The knowledge-graph explorer functions as a diligence cockpit: an analyst pulls a composition-intelligence report on a target's flagship material, walks the chain from formula to structure to property to patent to synthesis route, and examines the cross-source trust and disagreement scoring to see where predictions are well-corroborated and where independent engines conflict. That disagreement signal is directly relevant to gauging whether a target's model is calibrated or systematically overconfident — a question that rarely appears in a pitch deck but matters considerably for a platform company whose value proposition is prediction accuracy. The opportunity-index and buyer-intelligence views render the ranked opportunity analysis into an interactive screen: an analyst selects a chemistry domain, sees the ranked opportunity scores and funded-buyer leaderboard, and compares a target's roadmap against independent rankings in a single session — generating the kind of structured framing that maps directly onto an IC narrative. For deals involving an AI-materials platform rather than a single compound, batch screening lets the team run an entire candidate list through the negatives check and opportunity scoring at once, which is essential for underwriting companies whose value is a pipeline. The freedom-to-operate and patent-whitespace dashboard supports claim-level IP checks on a target's headline patents, grounding novelty assertions against the full materials patent corpus rather than a keyword search.

The natural engagement structure for Accel is a scoped technical read on a single live deal, rather than an asset license or co-development arrangement. A per-deal engagement combines the opportunity-and-buyer index and the negatives kill-edge test applied to the target's own candidates, with knowledge-graph trust scoring and freedom-to-operate layers pulled in where the specific deal warrants. The output is a buyer-grade memo structured to slot into an investment-committee narrative. Per-deal engagements run in the range of roughly $30,000 to $60,000 — a planning estimate for framing purposes, not a quote — consistent with a single-round technical read. Turnaround is scoped to work within live deal timelines. The recommended sequence is a paid pilot read on one live or recently-passed deal in Accel's current pipeline, chosen so the partnership can calibrate Lattice Graph's findings against the team's own judgment before committing to a broader arrangement. On success, the more efficient structure for a firm underwriting multiple AI-materials and critical-minerals deals per year is an annual diligence subscription that bundles API access to the opportunity intelligence and negatives atlas products for Accel's own analysts, alongside a defined number of scoped per-deal reads. The differentiator worth underscoring is that the 23,000-plus-edge negatives moat is largely internal and absent from the public data market, so the independent read it enables is one most AI-materials decks — and most competing diligence providers — cannot reproduce independently.