Lattice Graph × Orbital Materials

Lattice Graph operates a computational materials-discovery platform built around a knowledge graph that spans millions of compositions, linking each formula to its known structures, predicted and measured properties, synthesis recipes, and a full provenance trail back to the underlying evidence. What makes that graph unusual is not the scale of positive results it holds, but the explicit inclusion of negative results: over 23,000 labeled failed-experiment and kill edges documenting what was tried and did not work. For a domain where the published literature is systematically biased toward successes, that negative corpus is a distinct computational asset. Validation on the platform works through multi-engine consensus. Candidate materials are evaluated across independent physics engines, including machine-learning interatomic potentials and density-functional theory, to assess phonon stability, thermodynamic stability, and formation-energy predictions. Rather than trusting any single engine, the platform flags cases where the engines disagree, producing calibrated uncertainty signals that reflect genuine physical ambiguity rather than model overconfidence. This cross-source disagreement layer is what makes the platform's outputs auditable rather than merely ranked. For teams running generative models, these capabilities translate into something concrete: a generate-then-filter infrastructure where the model proposes candidates and the platform's negatives atlas, trust signals, and knowledge graph act as sequential guardrails before any candidate reaches physical synthesis. The result is a funnel that catches already-failed regions, flags high-uncertainty proposals for targeted validation, and grounds each surviving candidate in traceable provenance — directly addressing the positive-overfit failure that affects every generative materials program trained on the published literature.

Orbital Materials builds generative models that design new materials for data-center decarbonization and pairs those models with wet-lab synthesis and validation to confirm what the models propose. That architecture places Orbital at a specific and demanding seam: the model's quality determines how efficiently the lab's finite throughput is spent. Every false-positive candidate that reaches the bench without grounds for suspicion is wasted synthesis capacity; every overconfident score that is not caught before routing is a misallocation of the most expensive resource in the pipeline. The structural weakness of that architecture is well understood: generative materials models are trained overwhelmingly on positive experimental results, because positive results are what gets published, curated, and made available as training data. The models therefore learn the shape of success very well but have almost no signal about what failure looks like, where the dead-end regions of composition space are, or how to calibrate their own uncertainty in regions where multiple sources actually disagree. For Orbital, that means a meaningful fraction of model-proposed candidates will fall inside already-failed regions that are simply invisible to the model, and the model's own confidence scores will not reliably distinguish those candidates from genuinely novel ones. Lattice Graph's fit with Orbital is not about supplying more positive data — Orbital has access to the same curated positive literature every other AI-materials company does. The fit is about supplying the three layers that positives-only training cannot produce: the labeled negatives that teach a model what fails, the disagreement signals that calibrate a model's uncertainty against physical reality, and the knowledge graph provenance that makes any flagged candidate explainable and auditable. These map directly onto Orbital's bottleneck and can be deployed as API-layer guardrails without requiring any change to Orbital's underlying generative architecture.

Orbital Materials builds generative models that design new materials for data-center decarbonization and pairs those models with wet-lab synthesis and validation to confirm what the models propose. That architecture places Orbital at a specific and demanding seam: the model's quality determines how efficiently the lab's finite throughput is spent. Every false-positive candidate that reaches the bench without grounds for suspicion is wasted synthesis capacity; every overconfident score that is not caught before routing is a misallocation of the most expensive resource in the pipeline. The structural weakness of that architecture is well understood: generative materials models are trained overwhelmingly on positive experimental results, because positive results are what gets published, curated, and made available as training data. The models therefore learn the shape of success very well but have almost no signal about what failure looks like, where the dead-end regions of composition space are, or how to calibrate their own uncertainty in regions where multiple sources actually disagree. For Orbital, that means a meaningful fraction of model-proposed candidates will fall inside already-failed regions that are simply invisible to the model, and the model's own confidence scores will not reliably distinguish those candidates from genuinely novel ones. Lattice Graph's fit with Orbital is not about supplying more positive data — Orbital has access to the same curated positive literature every other AI-materials company does. The fit is about supplying the three layers that positives-only training cannot produce: the labeled negatives that teach a model what fails, the disagreement signals that calibrate a model's uncertainty against physical reality, and the knowledge graph provenance that makes any flagged candidate explainable and auditable. These map directly onto Orbital's bottleneck and can be deployed as API-layer guardrails without requiring any change to Orbital's underlying generative architecture.

The most operationally relevant surfaces for Orbital's team are batch screening, the composition-360 report, and the knowledge-graph explorer — used as an eval and triage console sitting downstream of the generative model's output. Batch screening is the daily workhorse: Orbital pushes a list of model-proposed candidates through the negatives atlas check and the trust and disagreement signals in a single pass and receives back a triage ranking that separates already-failed candidates, high-disagreement candidates that warrant targeted lab investigation, and clean candidates ready for synthesis. This converts the generative model's raw output into a prioritized queue without requiring a researcher to manually audit each proposal. For candidates that surface the negatives screen or are flagged by the trust layer, the knowledge-graph explorer and natural-language query interface are the investigation tool. A researcher can open a flagged candidate, walk its evidence neighborhood, read the specific sources behind each property prediction, and understand whether a disagreement reflects a genuine measurement conflict, a structural ambiguity, or a known failure mode in a related composition family. The composition-360 report provides the same depth in a structured single-view format for candidates approaching synthesis — structure, formation energy, recipes, and patent context in one auditable document. Together these surfaces turn what would otherwise be a black-box generative output into a documented, reproducible decision at each stage of the funnel.

Because Orbital is a data-archetype fit, the engagement is a data and evaluation license rather than an asset transaction. The natural entry point is a scoped negatives audit: a fixed-scope engagement in which Lattice Graph runs a defined slice of Orbital's model-proposed candidates through the Negatives & Eval-Data Atlas and trust signals, then reports the false-positive rate, the dead-end coverage gap, and the disagreement profile across Orbital's active material lanes. The output of that audit is a quantified picture of how much of Orbital's current candidate stream consists of already-failed or uncalibrated proposals — providing a documented basis for scoping the ongoing data license. The estimated starting range for this initial audit is $40,000 to $75,000. Following the audit, the typical path is a standing API and data license covering the negatives atlas, the trust and disagreement signals, and the knowledge-graph API. The key commercial lever at that stage is whether the negatives set is licensed for evaluation use only — hardening benchmarks and screening outputs — or also for model training, which carries a materially higher license bracket because the negatives corpus is the non-reproducible asset. Exact pricing, training versus evaluation scope, call volume tiers, and contract term are established in diligence after the audit results are in hand.