Lattice Graph × Entalpic

Lattice Graph is a computational materials-discovery platform built around a knowledge graph that spans millions of compositions, connecting structures, properties, synthesis routes, experimental evidence, and patent records into a single queryable network. What distinguishes the platform from a large-scale database or a foundation model is the validation layer: every candidate material of interest can be interrogated against multiple independent physics engines simultaneously, including machine-learning interatomic potentials such as MACE and CHGNet as well as density functional theory calculations, so that thermodynamic and phonon stability assessments represent genuine multi-engine consensus rather than a single model's output. Cross-source disagreement is made explicit rather than averaged away, giving downstream users a calibrated uncertainty picture that reflects where the science is settled and where it is genuinely contested. For a catalysis-focused AI company like Entalpic, the most strategically significant capability is one that does not appear in most materials platforms: a large, governed atlas of labeled negative results. The platform holds more than 23,000 failed-experiment records and kill edges, capturing the compositions, conditions, and routes that were tried and ruled out. This is the evidence that the peer-reviewed literature systematically omits. Generative models trained on publication-positive corpora will confidently re-propose known dead ends; models trained or evaluated against a substantial negative set will not. The Lattice Graph knowledge graph is built to surface this anti-signal with the same provenance rigor applied to positive evidence, and it includes PGM-free catalyst evidence directly relevant to Entalpic's core domain. Freedom to operate screening, supply-chain intelligence, and natural-language graph traversal complete the platform's capability set. A team can query the knowledge graph in plain language to trace the evidence chain behind any predicted property, identify the original DFT calculation or experimental record that anchors a stability claim, and understand how densely or sparsely a region of composition space is actually characterized. That provenance capability converts a model recommendation from a confidence score into a traceable, auditable case, which is what an industrial customer or a synthetic chemistry partner needs before committing resources to a campaign.

Entalpic is building generative and predictive materials models for catalysis and industrial chemistry, targeting the exact problem where AI-for-materials has the most visible credibility gap: proposing candidates that a chemist or process engineer will actually trust enough to test. Its business lines, AI for chemistry and catalysis, generative materials models, and industrial process materials, point to a company whose product is not a molecule but a recommendation pipeline that must survive contact with skeptical industrial customers. The competitive differentiation for Entalpic is therefore less about raw generative capacity, which is now broadly available following the GNoME and foundation-model wave, and more about whether the candidates its models produce are de-risked, provenance-grounded, and benchmarked honestly. The structural fit with Lattice Graph is direct. Catalysis is one of the most hostile domains for positive-only training data: activity, selectivity, and stability are acutely sensitive to synthesis route, support identity, promoter loading, and operating window, and any model that has never seen the failures in that parameter space will systematically overstate candidate quality. Lattice Graph holds the failed-experiment coverage that fills that gap, and it holds it with provenance, so every kill edge can be traced to its source rather than accepted as an opaque label. This matters for Entalpic specifically because its customers will ask how the model knows a candidate is not a known dead end, and the answer has to be grounded in evidence rather than in model confidence alone. The relationship is therefore a data and evaluation license rather than a patent or IP transaction. Entalpic does not need more discovery assets; it needs the missing training and evaluation signal, the uncertainty calibration layer, and the provenance API that lets its platform answer the "where did this come from" question with a traceable chain. Lattice Graph's PGM-free catalyst evidence in the governed graph adds direct domain relevance, making the partnership a tight thematic match rather than a general-purpose data augmentation play.

Entalpic is building generative and predictive materials models for catalysis and industrial chemistry, targeting the exact problem where AI-for-materials has the most visible credibility gap: proposing candidates that a chemist or process engineer will actually trust enough to test. Its business lines, AI for chemistry and catalysis, generative materials models, and industrial process materials, point to a company whose product is not a molecule but a recommendation pipeline that must survive contact with skeptical industrial customers. The competitive differentiation for Entalpic is therefore less about raw generative capacity, which is now broadly available following the GNoME and foundation-model wave, and more about whether the candidates its models produce are de-risked, provenance-grounded, and benchmarked honestly. The structural fit with Lattice Graph is direct. Catalysis is one of the most hostile domains for positive-only training data: activity, selectivity, and stability are acutely sensitive to synthesis route, support identity, promoter loading, and operating window, and any model that has never seen the failures in that parameter space will systematically overstate candidate quality. Lattice Graph holds the failed-experiment coverage that fills that gap, and it holds it with provenance, so every kill edge can be traced to its source rather than accepted as an opaque label. This matters for Entalpic specifically because its customers will ask how the model knows a candidate is not a known dead end, and the answer has to be grounded in evidence rather than in model confidence alone. The relationship is therefore a data and evaluation license rather than a patent or IP transaction. Entalpic does not need more discovery assets; it needs the missing training and evaluation signal, the uncertainty calibration layer, and the provenance API that lets its platform answer the "where did this come from" question with a traceable chain. Lattice Graph's PGM-free catalyst evidence in the governed graph adds direct domain relevance, making the partnership a tight thematic match rather than a general-purpose data augmentation play.

The most heavily used surfaces for Entalpic's team would be the knowledge-graph explorer and the composition-intelligence reports, operated as an evidence workbench alongside Entalpic's own generative models. A researcher can drive the graph explorer from a candidate catalyst composition, traverse the composition-to-structure-to-property-to-patent-to-recipe subgraph, and retrieve a composition-intelligence report that consolidates the cross-source DFT trust view, the disagreement signals, and the negatives status in one place. The practical effect is that a generated candidate lands with an immediate read on whether it is novel and clean, contested across sources, or already a recorded failure, transforming a model output into a recommendation with an auditable evidence trail. Batch screening, the catalyst and synthesis workflow templates, and the ML formation-energy predictor are the day-to-day computational tools. Entalpic can run an entire generation output through the kill-edge filter in a single batch pass rather than candidate by candidate, route survivors through the synthesis workflow templates to see plausible proof gates and route options, and use the formation-energy predictor as an independent cross-check on its own stability estimates. The natural-language graph query interface allows team members who are not graph query specialists to interrogate provenance and evidence neighborhoods directly, and the API endpoints that back the dashboard can be embedded into Entalpic's own evaluation pipeline so that the negatives and trust layers function as automated guardrails rather than manual review steps.

Because Entalpic's engagement archetype is data and evaluation rather than asset licensing, the practical entry point is a scoped negatives audit. Entalpic gets time-bounded access to the Negatives and Eval-Data Atlas, the Trust and Disagreement Signals, and the Knowledge-Graph API, and runs the catalysis-relevant slice against its own held-out evaluation set to measure directly how much the labeled failures and disagreement signals improve calibration and reduce false-positive candidate rates. The audit is designed to produce the internal evidence Entalpic needs to justify a recurring license, and the deliverable is a calibration report comparing Entalpic's baseline against the augmented pipeline, with specific attention to the catalysis-relevant kill edges and PGM-free evidence coverage. That audit converts naturally into an ongoing data and evaluation license that wires the three API products into Entalpic's training and evaluation loop on a recurring basis, with access scope, call volume, and refresh cadence negotiated based on what the audit demonstrates. The pricing context provided positions an initial negatives audit for a focused AI-materials lab in the range of roughly $40,000 to $75,000 to scope and characterize the dataset; a recurring production license would be structured above that level and negotiated after the audit findings are in hand. Optional expansions include supply-chain and feedstock-concentration intelligence and freedom-to-operate filtering on generated candidates if Entalpic later wants to layer those signals into its candidate-delivery workflow.