Which startups use AI to discover new materials with closed-loop experimentation?

The integration of artificial intelligence with automated robotics to create closed-loop (or "self-driving") laboratories is transforming materials science. Instead of relying on serial, human-driven trial and error, these systems use generative AI and physical modeling to propose candidate materials, automate robotic synthesis and testing, and feed the resulting data straight back into the AI to direct the next iteration.

Key startups and platforms spearheading this approach for materials discovery include:

• CuspAI – Co-founded by Max Welling and Chad Edwards, this Cambridge-based startup acts as a "search engine for materials that don't yet exist". Their platform combines generative AI models to propose millions of candidate compounds with downstream stability filters, high-fidelity simulations, and physical laboratory validation in a continuous, iterative loop.
• Kebotix – A pioneer in self-driving labs for chemistry and advanced materials. Kebotix utilizes a proprietary "predict-produce-prove" cycle, where machine learning and physical modeling instruct robotic hardware to synthesize and test new chemical formulations, closing the feedback loop autonomously.
• PhaseTree – A spinout from the Technical University of Denmark, PhaseTree combines physics-led computer simulations, AI, and laboratory automation to accelerate the design and discovery of targeted functional materials (such as those for next-gen batteries and solar cells).
• Mitra Chem – Focused heavily on battery materials, this US-based startup uses an AI-powered R&D platform paired with high-throughput automated experimentation to drastically shorten the timeline for discovering, scaling, and qualifying new iron-based cathode formulations for electric vehicles.
• Material Evolution – Operating in the green-tech and industrial materials space, they use machine learning and closed-loop data feedback to discover and formulate low-carbon, sustainable alternatives to traditional high-emission materials like ordinary Portland cement.

Would you like to explore:

• The underlying architecture of self-driving labs (combining LLMs, active learning, and robotic hardware)?
• Specific material applications like batteries or carbon capture?

Let me know how you'd like to proceed!

Yes. If by “closed-loop” you mean the system proposes a material → physically synthesizes/tests it → feeds the measured result back into the AI → chooses the next experiment, the startup landscape is narrower than the broader “AI materials discovery” category.

Startups that most closely fit

• Atinary — Probably the clearest fit. Its Self-Driving Labs platform combines AI/Bayesian optimization, robotics and lab instrumentation to design, execute, analyze and learn from experiments in a closed loop. It now operates its own self-driving labs in Boston, including autonomous platforms for chemistry and materials R&D.
• Orbital Materials — Focused specifically on discovering new materials, particularly for cleantech. Its Linus model generates and evaluates candidate materials, with the company operating a physical lab to test materials. Its approach is more AI-generated materials + experimental validation than a completely autonomous self-driving lab, so I'd classify it as partial closed-loop.
• CuspAI — AI-native materials discovery company using generative models to search enormous materials spaces. Its current AI Materials Foundry explicitly brings together AI, laboratories, industrial partners and experimental capabilities to move from discovery toward real materials. I'd put it in the closed-loop/physical-validation category, although its model is more ecosystem-oriented than Atinary's autonomous-lab model.
• Doublet Labs — Particularly interesting for organic crystals, superconductors, dielectrics and electronic materials. It describes a pipeline in which AI generates candidates, synthesis/characterization supplies experimental data, and those results retrain the model. Its in-house autonomous lab is described as the next step toward closing the loop end-to-end.
• Solbion — Building infrastructure that connects experimental hardware with active-learning loops, including automated thin-film deposition, inline metrology and orchestration software. This is particularly relevant if you're interested in the lab infrastructure/software layer rather than a vertically integrated materials company.

The distinction I'd make

The important distinction is that “AI materials discovery” doesn't necessarily mean closed-loop experimentation. Many companies use AI to screen or generate candidates and then hand those candidates to human scientists. A true self-driving system automates the experimental decision itself and uses the resulting measurement to select the next experiment. NIST describes exactly this architecture as AI deciding the experiment, executing it, collecting data, and repeating the cycle.

If you're looking at this as an investment/competitive landscape, I'd also include several non-startup reference points—Berkeley's A-Lab, NIST's autonomous labs, and newer NSF cloud-lab programs—because they're useful benchmarks for determining how much of each startup's loop is genuinely autonomous versus AI-assisted.

The clearest startups I found are:

• Yonder Materials — AI + autonomous high-throughput lab (“LeapLab”) running thousands of materials experiments/day and feeding results back into discovery.
• Autonomous Materials Labs (AML) — self-driving labs for alloy and ceramic development, with AI choosing subsequent experiments from real-time results.
• Doublet Labs — AI/physics models propose organic materials, experiments provide feedback, and the model iteratively improves; its autonomous in-house lab is being developed.
• AstraIQ — focuses on inorganic materials, automatically moving AI-screened candidates into synthesis and closing the discovery loop.
• Alloryn AI — vertically integrates simulation, robotic synthesis, characterization, and active learning into one closed loop for batteries, semiconductors and superconductors.

Also worth watching: Discovery Loop, the new Jeff Dean-led startup, reportedly aims to automate scientific discovery through AI-designed experiments, although its materials-specific capabilities are less clear publicly.

If you're mapping the competitive landscape, Yonder, AML, AstraIQ, and Alloryn look closest to the full AI → experiment → measurement → model update → next experiment paradigm.

The convergence of artificial intelligence, multi-agent frameworks, and robotic lab automation has enabled a new paradigm of closed-loop ("self-driving") experimentation for materials discovery. Rather than relying solely on static computer simulations or manual trial-and-error, these platforms use AI to predict properties, instruct physical robots or automated synthesis workflows to run the experiments, and feed the resulting physical data back into the models to guide the next iteration autonomously.

Several prominent startups and commercializing entities operate in this space:

• Discovered Materials : Backed by a $9M seed round led by Lightspeed and Y Combinator, this startup uses swarms of AI agents to shorten the multi-year timeline of finding and fabricating advanced materials for semiconductor chips and data center cooling. Their platform targets the lab-to-fab timeline by pairing computational agent workflows with rapid testing validation.
• Citrine Informatics : A pioneer in the materials informatics space, Citrine provides the Citrine Platform which integrates materials databases with machine learning and automated/closed-loop computational workflows. They have collaborated with institutional and industrial partners on closed-loop frameworks to discover novel catalysts and battery electrolytes.
• Osmium AI : Backed by Y Combinator, Osmium leverages predictive AI to forecast physical properties of novel industrial materials, helping customers rapidly optimize and cycle through material design candidates.
• PhaseTree : A Denmark-based startup that combines AI, computer simulations, and laboratory automation specifically geared toward accelerating materials discovery for clean energy applications like batteries, wind turbines, and solar cells.
• Copernic Catalysts : Focused heavily on industrial decarbonization, this US-based startup uses AI and density functional theory (DFT) paired with automated feedback loops to redesign and discover catalysts for zero-carbon ammonia and e-fuel production.

(Note: While academic and national lab testbeds like Berkeley's autonomous A-Lab , the multi-agent MARS platform, and Google DeepMind's GNoME framework laid the foundational data and algorithmic architectures, the startups above are commercializing these closed-loop loops for industrial and tech supply chains).

Would you like to explore how multi-agent AI frameworks (like MARS or agent swarms) physically integrate with robotic hardware, or dive deeper into a specific sector like battery electrolytes or semiconductors?

Several pioneering startups combine generative AI, large language or multimodal models, and automated robotic labs to execute closed-loop experimentation (where AI generates hypotheses, robots synthesize and test physical samples, and the results automatically feed back into the AI to design the next iteration).

Would you like to explore how these closed-loop platforms compare with traditional computational software tools (like Citrine Informatics or NobleAI), or focus on a specific sector like batteries or semiconductors?

Yes. The most relevant startups are those that go beyond AI prediction and actually connect AI → physical experiment → measurement → model update → next experiment—the closed-loop/self-driving-lab paradigm.

Strongest examples

Atinary is probably the clearest example of a platform company: it explicitly describes its system as autonomously designing, executing and learning from experiments, with AI, robotics and human expertise in a closed loop.

Dunia is particularly interesting if you're looking for industrial-scale autonomous materials discovery. Its platform explicitly integrates AI, robotic experimentation and the complete DMTA loop.

Yonder is closer to a vertically integrated materials-discovery company: its LeapLab reportedly performs thousands of experiments per day and uses those physical results to validate AI-generated materials.

A useful distinction

There are really three layers of companies in this market:

1. AI-only materials discovery — predict properties or generate candidates computationally.
2. AI + experimental optimization — AI decides which experiment to run next, but humans may execute it.
3. Full self-driving lab — AI proposes → robots synthesize → instruments characterize → data automatically updates the model → AI chooses the next experiment.

If by "closed-loop experimentation" you mean #3, I'd put Atinary, Dunia, Yonder, Solbion and Doublet Labs closest to the target. The field is still young, and some companies describe an eventual fully autonomous loop while currently using partial automation or external synthesis partners. Doublet, for example, says its synthesis is currently handled by CRO partners while it works toward an in-house autonomous lab.

There is also a newer 2026 cohort worth watching: Discovered Materials/Matforge, which YC describes as using AI agents across candidate generation, physical synthesis and testing specifically for semiconductor materials.

If you're evaluating this space commercially, I can also give you a 20–30 company landscape categorized by batteries / semiconductors / catalysts / polymers, with funding, location, founders, and whether the physical loop is actually operational today.

The convergence of frontier artificial intelligence with high-throughput robotics has enabled the rise of self-driving, closed-loop autonomous laboratories . Instead of relying strictly on isolated computational predictions or manual trial-and-error, these companies build systems where AI models generate hypotheses, direct robotic hardware to synthesize and test physical materials, and automatically feed the real-world results back into the model to plan the next iteration.

Prominent startups and ventures operating in this space include:

1. Periodic Labs

• What they do: Founded by prominent AI and materials veterans (including co-creators of ChatGPT and Google DeepMind's GNoME material-discovery engine), Periodic Labs pairs cutting-edge "AI scientists" with fully autonomous robotic laboratories.
• The Closed-Loop Approach: They focus heavily on physical solid-state synthesis (such as powder processing) to manufacture and test candidates in-house at a scale traditional labs cannot match. By physically running tests and logging data in a structured format for their AI, the system autonomously forms hypotheses, executes verification, and targets breakthrough physics challenges like advanced semiconductors and superconductivity.

2. Discovered Materials

• What they do: Emerging from Y Combinator, this startup deploys swarms of AI agents combined with foundational physics models specifically targeted at solving thermal and materials bottlenecks in hardware (such as cooling materials for chips running intensive AI workloads).
• The Closed-Loop Approach: Their software pipeline harnesses advanced language and reasoning models to propose candidate materials, which are then routed through custom physics simulation and verification loops to rapidly iterate on viable physical properties.

3. NobleAI

• What they do: NobleAI provides an enterprise science platform trusted across the chemicals, materials, and energy sectors to optimize formulations and accelerate R&D.
• The Closed-Loop Approach: By integrating machine learning models with historical and real-time laboratory data streams, their platform predicts physical properties and guides automated or high-throughput ingredient replacement workflows, drastically reducing physical testing cycles.

Broader Industry & Academic Ecosystem

Beyond dedicated venture-backed startups, major corporate research arms and academic testbeds are pioneering this exact architecture:

• Matique: Uses combinatorial synthesis and AI platforms (collaborating with groups like the Toyota Research Institute) to map nanoparticle libraries and accelerate clean-energy catalyst discovery.
• Academic/Institutional Platforms: Initiatives like NC State's SPEED (Self-driving Platforms for Expedited Experimental co-Design) spearheaded by researchers like Milad Abolhasani, and automated synthesis frameworks out of institutions like Tohoku University, are laying the open-architecture blueprints that transition chemistry from open-loop prediction to true closed-loop autonomous execution.

If you'd like, I can dive deeper into:

• The hardware bottlenecks (like powder handling vs. liquid dispensing) these labs face
• How these systems utilize active learning algorithms to pick the next experiment
• Specific target domains like battery electrolytes vs. superconductors

Let me know how you'd like to explore this further.

Yes. If by “closed-loop” you mean AI proposes a material → robots synthesize/test it → measurements feed back into the model → AI chooses the next experiment, the field is still fairly small. The strongest startup examples I’d put on a shortlist are:

The clearest examples

Atinary is probably the closest match to your description. Its Self-Driving Labs platform explicitly combines AI, robotics and automation in a closed loop that designs, executes and learns from experiments. It is now operating its own self-driving labs in Boston as well as deploying the technology with industrial and academic partners.

Citrine Informatics is another particularly important company. Its platform uses sequential learning to decide what to test next, and Citrine reports an AI-guided closed-loop nanoparticle-synthesis experiment conducted with SLAC. Its 2026 technical description explicitly frames the platform as a closed sequential-learning loop spanning data → ML → constrained candidate generation → experimental evaluation.

Orbital Materials takes a somewhat different approach. Its Linus model generates candidate materials computationally, including materials for batteries, catalysts and CO₂ capture, and the company has built an in-house lab around the model. The important distinction is that Orbital has historically emphasized the AI/material-generation side more than a fully autonomous physical experimentation loop.

Alloryn AI is perhaps the most literal startup implementation of the concept: it describes a loop of computational screening, robotic synthesis, automated characterization, and retraining from every physical result. It's much newer, so I'd treat its claims as promising rather than as established at the scale of Citrine or Atinary.

One important distinction

There are actually three different categories being lumped together under “AI materials discovery”:

1. AI screening/design — predicts properties or generates candidate materials, but humans perform the experiments.
2. AI-guided experimentation — AI chooses the next experiment, but a human/partially automated lab executes it.
3. True self-driving lab — AI chooses and executes the experiment, observes the result, updates its model, and repeats.

Your question is really about #3. That's a much smaller group.

For context, the academic benchmark for this paradigm is Berkeley's A-Lab, which combined computation, ML/active learning, robotics and characterization in a closed loop and synthesized 36 compounds from 57 targets during 17 days of continuous operation. It's a research lab rather than a startup, but it illustrates what the startups are trying to commercialize.

If you're mapping the competitive landscape, I'd expand this into a list of ~15–25 startups and rank them by (a) true physical closed-loop capability, (b) proprietary AI/models, (c) robotics ownership, (d) materials categories, (e) funding, and (f) whether they sell to customers vs. discover materials themselves. That produces a much more useful picture of who is actually competing in “self-driving materials discovery.”

Orbital Materials takes a somewhat different approach. Its Linus model generates candidate materials computationally, including materials for batteries, catalysts and CO₂ capture, and the company has built an in-house lab around the model. The important distinction is that Orbital has historically emphasized the AI/material-generation side more than a fully autonomous physical experimentation loop.

The intersection of artificial intelligence and robotics has given rise to self-driving or autonomous laboratories that execute closed-loop Design-Make-Test-Analyze (DMTA) cycles. Rather than relying solely on computer simulations, these startups physically synthesize, characterize, and iterate on new materials without human intervention in the experimental loop.

Prominent startups leveraging AI with closed-loop physical experimentation for materials and molecular discovery include:

• Radical AI – Combines frontier multimodal AI models with fully robotic, self-driving laboratories. Users input target performance properties, and the AI engine designs the experiments, sends instructions straight to the automated lab for synthesis and characterization, and uses real-time feedback data to refine subsequent hypotheses in a continuous flywheel.
• Dunia – A Berlin-based deep-tech startup building AI- and robotics-powered systems (such as their autonomous industrial lab platform, IRIS) to bridge the "sim-to-real gap". They use closed-loop automated experimentation to synthesize, test, and scale industrial and climate breakthrough materials, turning physical outputs into validated operational datasets.
• Discovery Loop – Founded by former Google/DeepMind leaders (including chief scientist Jeff Dean), this public-benefit corporation focuses on using AI to design and execute massive parallel experiments in a closed loop. They target complex engineering and deep-tech challenges spanning battery materials for EVs, novel magnets, and broader scientific discovery.
• Kebotix – An AI platform and pioneer in chemical and material discovery that blends machine learning, physical modeling, and high-throughput robotic automation. Their platform drives closed-loop experimentation to accelerate the identification of novel molecules and advanced materials.
• Citrine Informatics – Provides AI and data infrastructure that helps materials and chemicals enterprises implement sequential learning and automated closed-loop frameworks, dramatically accelerating design time and research productivity.

Would you like to explore how these closed-loop labs overcome the 'sim-to-real' gap , or are you looking for startups focusing on a specific application like battery electrolytes or catalysts?

Yes. If by “closed-loop experimentation” you mean the system uses AI to propose candidates → physically synthesize/test them → feed the measurements back into the model → choose the next experiments, there are several startups worth watching.

Strongest matches

• citrine.io — Probably the clearest established example. Its platform uses sequential learning to select experiments, incorporates the experimental results, retrains/improves the model, and repeats. Citrine has reported applications spanning materials, chemicals, batteries, catalysts, and autonomous nanoparticle synthesis.
• cusp.ai — A particularly interesting newer player. Its AI Materials Foundry connects AI-driven materials design with simulation, synthesis planning, experimental laboratories, and characterization. Its 2026 network includes industrial partners and autonomous-lab partners, explicitly targeting the computational-to-experimental loop.
• orbitalindustries.com — Builds AI models for discovering materials and combines them with its own laboratory/hardware capabilities. Its current CurieOS platform is positioned as an AI operating system connected to labs and instruments, with the company describing an end-to-end materials discovery and engineering workflow.
• Alloryn AI — A very direct match conceptually: its platform claims a fully autonomous loop of simulation → robotic synthesis → automated characterization → model retraining, with each experimental result determining the next simulation batch. It's much earlier-stage than Citrine/CuspAI, so I'd treat its claims accordingly.

Adjacent players

• matlantis.com — Strong AI/atomistic simulation technology for predicting material behavior and guiding experimental discovery, but its public positioning is more AI simulation preceding experiments than a fully autonomous physical closed loop.
• citrine.io is worth distinguishing from generic “AI materials discovery”: its system explicitly describes the loop as model → select experiments → run experiments → learn from results → select the next experiments.

The landscape

I'd divide the market into roughly three tiers:

The important distinction is that “AI materials discovery” alone isn't necessarily a self-driving lab. Companies such as Matlantis can generate very powerful computational predictions, whereas the truly closed-loop companies connect those predictions to robots/instruments and real experimental feedback.

For context, this architecture is now sufficiently established that NIST describes autonomous materials systems in essentially exactly these terms: AI chooses an experiment, the experiment is performed, data are collected, and the next experiment is selected in a loop.

If you're looking at this from an investment/competitive-landscape perspective, I can also give you a list of 20–30 startups worldwide, categorized by materials focus (batteries, semiconductors, catalysts, polymers, carbon capture, etc.), degree of lab automation, funding, founders, and whether the loop is actually autonomous versus marketing language.