MOFGEN

MOFGEN is a generative AI system for metal-organic framework (MOF) discovery developed by CuspAI, a British AI-for-materials company headquartered in Cambridge. The system combines a language model that proposes novel MOF compositions, a diffusion model that generates corresponding crystal structures, quantum mechanical agents that optimize and filter candidates, and synthetic-feasibility agents guided by expert rules and machine learning. CuspAI has reported a 49 percent valid-unique-novel (VUN) candidate generation rate for MOFGEN, ahead of peer research programs from Microsoft Research (10 percent) and Meta FAIR Chemistry (16 percent), and the system was validated through experimental synthesis of five "AI-dreamt" MOFs in laboratory partnerships.

At a glance

• Lab: CuspAI
• Released: Capability disclosures and partnership announcements through 2025 and 2026. Underlying System of Agentic AI for MOF Discovery paper posted to arXiv in April 2025 (arXiv:2504.14110).
• Modality: Generative materials AI. Inputs include target chemistry properties and synthesis constraints. Outputs are proposed MOF compositions with predicted crystal structures and synthesis routes.
• Open weights: No. CuspAI develops commercial generative-materials models accessed through partner-customer arrangements rather than as standalone open releases.
• Context window: Not applicable in the language model sense. The system handles MOF design tasks specified by target physical and chemical properties.
• Pricing: Not publicly disclosed. Access is through industrial customer partnerships (Hyundai, Meta, Kemira, and others) rather than self-service developer access or hosted API.
• Distribution channels: Direct industrial-customer partnerships, scientific collaborations (the OpenDAC dataset partnership with Meta and Georgia Institute of Technology), and continued model-iteration cycles through customer engagements.

Origins

CuspAI was founded in 2024 in Cambridge, United Kingdom, by Chad Edwards (a chemistry PhD and former senior executive at the quantum computing company Quantinuum) and Max Welling (Professor of Machine Learning at the University of Amsterdam, former Distinguished Scientist at Microsoft Research, and former vice president of technology at Qualcomm). The founding thesis combined Edwards's deep-technology operating experience with Welling's research direction in geometric deep learning and equivariant neural networks. The premise was that materials discovery, like drug discovery before it, sits in a structural position where laboratory cycles run on multi-year timescales and computational generative methods could compress those cycles by orders of magnitude.

The company's June 2024 seed round of $30 million, led by Hoxton Ventures with Basis Set Ventures and Lightspeed, established the initial capitalization. The September 2025 Series A of $100 million at a $520 million valuation, co-led by New Enterprise Associates and Temasek, supported the expansion of MOFGEN and the broader CuspAI search-engine-for-materials platform across industrial customer engagements.

The MOFGEN system was the company's flagship public capability disclosure. The paper "System of Agentic AI for the Discovery of Metal-Organic Frameworks" (arXiv:2504.14110) documented the system architecture in April 2025, presenting MOFGEN as an integrated multi-agent system rather than as a single generative model. The reported 49 percent VUN generation rate, against benchmarks of 10 percent for Microsoft Research models and 16 percent for Meta research models, was characterized in NEA's investment thesis and in industry coverage as one of the more credible quantitative capability signals among 2024 to 2025-vintage AI-for-materials startups.

CuspAI's scientific advisory board includes Geoffrey Hinton, Yann LeCun, Kristin Persson (the Lawrence Berkeley scientist behind the Materials Project database), and Verity Harding. The MOFGEN approach has been validated through the OpenDAC partnership with Meta and Georgia Institute of Technology (the world's largest direct air capture database with more than 100 million data points) and through the SkyVault project, which demonstrated end-to-end progression from generative MOF design through laboratory synthesis in approximately six months.

Capabilities

MOFGEN is built specifically for the generative discovery and synthesis of metal-organic frameworks. MOFs are a class of porous crystalline materials with applications in carbon capture, gas storage, water purification, and catalysis. Three capability features distinguish MOFGEN from peer materials-AI approaches.

The first is the multi-agent architecture. MOFGEN is structured as an integrated system of agents rather than as a single end-to-end generative model. The architecture includes a language model that proposes novel MOF compositions, a diffusion model that generates corresponding crystal structures, quantum mechanical agents that optimize candidate structures and filter for thermodynamic stability, and synthetic-feasibility agents guided by expert rules and machine-learned heuristics. The multi-agent structure addresses the multi-stage nature of materials discovery: composition design, structure prediction, property optimization, and synthesis-route planning are distinct subproblems that traditional single-model approaches conflate.

The second is the synthesis-aware design focus. The synthetic-feasibility agents filter generated candidates for synthesizability, addressing a persistent failure mode in computational materials science where promising candidates prove impossible to produce at scale. The synthesis filtering is the structural reason CuspAI claims a higher VUN rate than peer research approaches: candidates that pass MOFGEN's synthesis filter are more likely to be experimentally producible, while peer approaches generate larger raw candidate sets at lower synthesis-success rates.

The third is the experimental validation loop. MOFGEN's outputs have been physically synthesized through CuspAI's laboratory partnerships, with five "AI-dreamt" MOFs successfully produced in laboratory experiments at the time of the April 2025 paper. The end-to-end progression from generative design through experimental synthesis is the principal evidence that the model's outputs translate to manufacturable materials.

The CuspAI search-engine-for-materials platform extends MOFGEN beyond MOFs to broader chemistry applications including semiconductor manufacturing, energy storage, and PFAS removal from water systems. The platform supports customer-specific design tasks where target chemistry properties are specified upfront and the system produces synthesizable candidates approximately 10 times faster than traditional materials-discovery methods (per CuspAI's framing).

Pipeline and evaluation

MOFGEN's evaluation framework focuses on materials-discovery-specific metrics rather than horizontal language model leaderboards.

The principal disclosed metric is the valid-unique-novel (VUN) generation rate, a composite measure of generated structure quality across three dimensions: validity (does the generated structure satisfy basic chemistry constraints), uniqueness (is it distinct from previously generated candidates), and novelty (is it distinct from known structures in the training database). MOFGEN reports a 49 percent VUN rate, against benchmarks of 10 percent for prior Microsoft Research models (MatterGen and adjacent) and 16 percent for prior Meta FAIR Chemistry models. The headline 49 percent figure is the principal capability claim CuspAI has used in commercial positioning and investor materials.

A secondary set of metrics relates to synthesis success rates: the fraction of generated candidates that can be physically produced. The five experimentally synthesized MOFs from the April 2025 paper provide point evidence for synthesis success, though large-scale published statistics on synthesis success rate across the candidate population are not yet available.

A third metric category covers target-property accuracy: how closely synthesized materials match the specified property profile (porosity, gas-affinity selectivity, thermodynamic stability). The OpenDAC and SkyVault demonstrations provide evidence for target-property accuracy at the scale of small experimental cohorts.

The standard horizontal language model benchmarks do not apply to materials design. Industry coverage has consistently characterized the MOFGEN VUN result as one of the more credible quantitative capability signals among 2024 to 2025-vintage AI-for-materials startups, with the caveat that the comparison points are publicly disclosed research models rather than commercial offerings.

Access and partnerships

MOFGEN is not available as a hosted public web tool, an API, or open weights distribution. The system is closed source, and access is structured through commercial customer partnerships rather than self-service developer access.

The principal disclosed partnership channels are:

• Hyundai Motor Group. Multi-year collaboration on sustainable-energy applications, including battery and energy-storage materials. The most prominent industrial-customer partnership at the May 2026 reading.
• Meta and Georgia Institute of Technology (OpenDAC). Scientific collaboration on the OpenDAC direct-air-capture dataset, which CuspAI has characterized as the world's largest such dataset with more than 100 million data points. The collaboration produced the SkyVault end-to-end synthesis demonstration.
• Kemira. Collaboration with the Helsinki-listed chemicals company on materials for removal of PFAS "forever chemicals" from water systems.
• Internal CuspAI design programs. The company runs internal materials-discovery programs targeting carbon capture, semiconductors, water purification, and energy storage, with the search-engine-for-materials platform as the underlying infrastructure.

The pharmaceutical-industry-style partnership economics (research access fees, milestone payments, royalty structures on resulting materials) are the implied commercial model. Specific partnership terms are negotiated bilaterally and are not publicly disclosed.

Comparison

Direct competitors and adjacent materials-AI systems:

• GNoME (Google DeepMind). The Graph Networks for Materials Exploration program. Frontier-lab peer; positioned around large-scale materials enumeration rather than synthesis-aware generative design.
• MatterGen and MatterSim (Microsoft Research). Microsoft generative-and-simulation materials research programs. CuspAI cites a 10 percent VUN rate for MatterGen-class research models against MOFGEN's 49 percent.
• Meta FAIR Chemistry (Meta AI / FAIR). Meta's chemistry research program, including the OpenDAC partnership with CuspAI. CuspAI cites a 16 percent VUN rate for prior Meta research models.
• MOFGPT and ChatMOF (academic). Language-model-based academic MOF design and prediction lines. Different architectural approach than MOFGEN's multi-agent system.
• Periodic Labs, Lila Sciences. Autonomous-laboratory AI-for-science peers; Periodic Labs has materials focus, Lila Sciences has broader biology focus.
• Citrine Informatics, Schrödinger, Materials Project. Established materials informatics peers with longer commercial track records but different approach (prediction and database curation rather than generative design).

MOFGEN's distinctive position among 2024 to 2025 vintage materials-AI systems: the multi-agent architecture with synthesis-aware filtering, the reported 49 percent VUN generation rate against peer research benchmarks of 10 percent and 16 percent, the experimental validation through five physically synthesized AI-designed MOFs, and the industrial-partnership distribution channels (Hyundai, Meta, Kemira) that anchor commercial reference customers.

Outlook

Open questions for MOFGEN over the next 6 to 18 months:

• VUN benchmark independent reproduction. The 49 percent VUN figure is a CuspAI published claim against publicly disclosed peer research models. Independent third-party reproduction of the metric on shared benchmark conditions would strengthen the result's standing in the materials-AI research community.
• Scale of experimental synthesis success. Five experimentally synthesized MOFs at the April 2025 paper provided point validation. Whether the synthesis success rate generalizes across larger candidate populations and to the more challenging target classes pursued in the Hyundai and Kemira partnerships remains to be demonstrated at scale.
• Pipeline expansion beyond MOFs. CuspAI has positioned the search-engine-for-materials platform as extensible to broader chemistry applications including semiconductors, energy storage, and PFAS removal. Whether MOFGEN-style architecture transfers to these domains will affect the company's commercial scope.
• Series B and follow-on financing. Industry coverage in early 2026 has reported follow-on discussions targeting a unicorn-class valuation. The progression of those discussions and the use of proceeds will shape the model program's trajectory.
• Competitive dynamics with GNoME and MatterGen. Frontier-lab materials-AI programs at Google DeepMind and Microsoft Research operate at greater compute scale. Whether MOFGEN's synthesis-aware focus produces durable commercial differentiation against the larger research programs is the central commercial test.
• OpenDAC and SkyVault scaling. The carbon-capture demonstration program's progression from research validation to commercial-scale carbon-capture deployments will be a watchable signal for the model's translation from laboratory results to industrial deployment.

Sources

• System of Agentic AI for the Discovery of Metal-Organic Frameworks (arXiv). Technical paper covering the MOFGEN architecture and benchmark results.
• CuspAI official site. Company and platform reference.
• NEA: From Algorithms to Atoms - Investment in CuspAI. Lead investor thesis with VUN benchmark detail.
• Fortune: CuspAI raises $100 million Series A at $520 million valuation. Series A funding context.
• SiliconANGLE: CuspAI raises $100M to build AI search engine to transform materials science. Capability and partnership context.
• Hyundai Motor Group partnership announcement. Sustainable-energy collaboration.
• Sourcery VC: CuspAI investment thesis. MOFGEN benchmarks and competitive positioning.
• HumanX: CuspAI startup analysis. Company background and advisory-board composition.