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

The intersection of artificial intelligence, robotics, and high-throughput physical synthesis has given rise to self-driving or closed-loop laboratories for materials and chemical discovery. In these workflows, AI models design candidate molecules or alloys (in silico ), direct automated robotic hardware to synthesize and test them in the real world, and automatically feed the resulting physical data back into the algorithm to refine the next iteration.

Key startups and platforms leading this space include:

• • CuspAI : Co-founded by machine learning pioneer Max Welling, CuspAI operates an agentic AI platform designed to orchestrate the end-to-end discovery of new materials. Through its expansive AI Materials Foundry network (in partnership with companies like NVIDIA and Meta), its platform integrates generative design, atomistic simulations, and coordinated physical validation loops.
• • Kebotix : A recognized pioneer in combining artificial intelligence with robotic automation. Kebotix utilizes a "lab-in-the-loop" approach where machine learning models predict novel chemicals, polymers, and advanced materials, and automated hardware physically synthesizes and characterizes them to accelerate R&D cycles.
• • Citrine Informatics : Provides an AI and data platform built specifically for materials and chemicals. Citrine’s framework supports closed-loop integration by coupling machine learning predictions and surrogate models with physical experimental data to optimize target properties like battery electrolytes or clean-energy electrocatalysts.
• • Radical AI: Focuses on vertically integrated, closed-loop automated material science laboratories. They build custom robotic hardware and AI systems capable of operating standard laboratory equipment (such as high-temperature oxidation furnaces) to autonomously synthesize and characterize difficult materials like high-entropy alloys.

Would you like to explore the underlying machine learning architectures (such as Graph Neural Networks or Machine Learning Interatomic Potentials) that power these closed-loop predictions, or focus on a specific application like battery tech or catalysts?

Yes. If by “closed-loop” you mean AI proposes a material → a robot/lab synthesizes and tests it → results feed back into the model → the next experiment is selected, the startup landscape is still fairly small.

Most relevant startups

• CuspAI — Probably the strongest example on the AI-designed-materials side. Its system starts from desired properties, generates novel candidates, simulates them, and increasingly connects discovery to physical synthesis through its AI Materials Foundry network. Its current approach explicitly describes the sequence as defining objectives → generating candidates → simulation → lab synthesis/validation → scale-up.
• Lila Sciences — One of the clearest bets on AI-operated autonomous laboratories. The company combines AI scientists with automated experimentation, making it particularly relevant if you're interested in the physical closed-loop rather than merely AI screening. Industry analysis places Lila among the startups connecting AI directly to synthesis, testing and iteration.
• Periodic Labs — Focuses on AI-driven materials research coupled with laboratory experimentation. It's part of the newer wave moving beyond prediction/screening toward autonomous experimentation and owning discoveries/IP.
• Orbital Materials — Uses its materials foundation model, Linus, to generate candidate materials and has its own wet-lab capability. Its initial focus includes batteries, semiconductors, catalysts and carbon-capture materials. It's somewhat more AI-first + targeted experimental validation than a fully autonomous closed-loop lab.
• Chemify — More chemistry than materials in the narrow sense, but highly relevant to the same architecture: AI designs compounds/material precursors and automated chemistry executes the synthesis, creating an experimental feedback loop. It is especially interesting if you're including molecular/materials chemistry.

A useful distinction

There are really three tiers:

The important caveat is that “AI materials discovery” is much broader than “closed-loop autonomous materials discovery.” Companies such as Citrine Informatics, NobleAI and others primarily provide prediction, simulation, data or optimization software; they shouldn't automatically be counted as autonomous-lab companies.

The underlying technology is well established academically: NIST's CAMEO system, for example, has demonstrated the complete loop of AI choosing an experiment, running it, measuring the result and using that information to select the next experiment.

If you're evaluating this as an investment/startup landscape, I'd put CuspAI, Lila Sciences and Periodic Labs in the most interesting “AI + autonomous experimentation” bucket, with Orbital Materials as a particularly interesting AI-model/materials-design play.

The convergence of generative AI, active learning algorithms, and robotic self-driving laboratories (SDLs) has transformed materials science into a closed-loop Design-Make-Test-Analyze (DMTA) process. Rather than relying on human trial-and-error, AI models propose novel chemical structures or crystal lattices, automated hardware synthesizes and characterizes them, and the resulting data automatically refines the next round of AI predictions.

Key startups and ventures leading this closed-loop autonomous materials discovery space include:

• Discovery Loop — Founded by prominent former Google and DeepMind leaders (including AI/infrastructure veterans Jeff Dean and Sanjay Ghemawat), this public benefit corporation focuses on utilizing frontier AI models to design, execute, and learn from massive parallel experiments in a closed loop. Their scope spans complex engineering and scientific domains like fusion magnets and electric vehicle battery materials.
• Dunia — Operating a full-stack platform explicitly designed to bridge the "Sim2Real" (simulation-to-reality) gap in material engineering. Dunia integrates next-generation AI models with high-throughput automated physical laboratory testing to continuously cycle through ideation, synthesis, testing, and validation.
• Kebotix — An established pioneer in the space that combines artificial intelligence, physical modeling, and advanced lab automation. Kebotix's platform accelerates the discovery of new chemicals and advanced materials by turning the physical laboratory into an automated, digitalized feedback loop.
• Citrine Informatics — While functioning heavily as an enterprise software and AI platform provider rather than a pure wet-lab startup, Citrine builds closed-loop computational and sequential learning frameworks. They partner with organizations to tie machine learning surrogate models directly into physical or high-throughput automated task execution for materials like battery electrolytes and catalysts.

Would you like to explore how self-driving labs handle the Sim2Real gap , or are you looking for details on a specific application like battery electrolytes or carbon capture materials?

If by “closed-loop” you mean AI proposes an experiment → robots run it → instruments measure the result → the result updates the model → AI chooses the next experiment, the startup landscape is still fairly small.

The strongest fits

• Atinary — Probably the clearest commercial example. Its Self-Driving Labs integrate AI, robotics, lab automation and human expertise into a closed Design–Make–Test–Analyze–Learn loop. In 2026 it opened its own AI-powered lab in Boston, with autonomous platforms spanning chemistry, materials and pharma R&D.
• Dunia — Particularly focused on materials discovery, with an AI + robotics platform that integrates the DMTA cycle. Its initial emphasis is electrochemical experiments, making it especially relevant to energy materials.
• Plexymer — Building self-driving laboratories that combine intelligent orchestration, robotic experimentation, application-specific testing and ML. Materials are one of its target areas, alongside biologic formulation. It describes the system as a continuous design/build/test/learn loop, with humans supervising the process.
• Citrine Informatics — A more software/data-centric player. Citrine's platform uses sequential learning to select experiments under constraints such as composition, physics, manufacturability and cost. Its 2026 technical paper describes applications including autonomous nanoparticle synthesis and reports substantial reductions in experimental effort versus random search.

One important distinction

CuspAI is highly relevant to AI-driven materials discovery, but I'd put it in a different bucket. CuspAI's core strength is AI/generative computational discovery and its Materials Foundry connects industrial and laboratory partners; I wouldn't characterize it as a startup whose primary product is an autonomous physical experimentation loop in the same way as Atinary or Dunia.

The benchmark isn't actually a startup

The most impressive demonstration of the concept is arguably Berkeley/Lawrence Berkeley National Laboratory's A-Lab. It combines computation, ML, robotics, characterization and active learning into a closed loop for inorganic-material synthesis. In its original 17-day campaign, it ran 353 experiments and synthesized 36 of 57 target materials.

There is also an A-Lab GPSS effort reported in 2026 that extends autonomous experimentation to air-sensitive inorganic materials and incorporates agentic AI reasoning.

My rough competitive map

If you're looking specifically for startups that own/control the physical experimentation loop rather than merely selling materials-AI software, I'd narrow the list to Atinary, Dunia, and Plexymer, with Atinary currently looking like the most mature commercial example.

Yes. If by “closed-loop” you mean the AI proposes an experiment → robots/materials equipment run it → measurements come back → the model updates → it chooses the next experiment, rather than merely predicting materials in silico, the startup landscape is fairly concentrated.

The strongest fits

Adjacent companies worth watching

Orbital Materials is a major AI-for-materials startup, particularly around batteries and carbon capture. Its core proposition is AI-driven materials discovery, but I would not put it in the same bucket as Atinary/Kebotix if your criterion is an autonomous physical experiment loop.

Periodic Labs, Lila Sciences, Dunia, Radical AI and others are appearing in the broader “AI + physical experimentation” ecosystem. A 2026 industry landscape specifically identifies these as players moving toward synthesis, testing and iteration, but their degrees of actual closed-loop autonomy differ substantially.

A useful way to segment the market

I would divide the startups into three tiers:

1. AI + autonomous physical experimentation

• Atinary
• Kebotix
• Autonomous Materials Labs
• Solbion

2. AI-native materials discovery with an emerging experimental loop

• Doublet Labs
• Auronex AI
• Orbital Materials

3. Broader “AI scientist / physical AI” companies

• Lila Sciences
• Periodic Labs
• Radical AI
• Dunia

The distinction matters. Generating 10 million candidate materials with a model isn't a closed loop. The really interesting companies are the ones where the AI has control over the next physical experiment and receives experimental data back into its decision process. That's the “self-driving lab” paradigm; Atinary explicitly describes it as integrating AI, robotics and human expertise into a closed loop.

For context, this isn't just startup hype: NIST describes its own autonomous materials system as an AI that selects an experiment, runs it, collects the data and repeats the cycle.

If you're looking at this from an investment/competitive-landscape perspective, I can also map ~15–25 companies by funding, founding team, material domain, degree of autonomy, proprietary lab/robotics, customers, and whether they actually own the discovered IP.

The intersection of artificial intelligence and robotic automation has given rise to self-driving laboratories (SDLs) that execute a Design-Make-Test-Analyze (DMTA) closed-loop cycle. Instead of just running static virtual simulations, these startups physically synthesize, characterize, and iterate on materials autonomously.

Prominent startups leveraging AI-driven closed-loop experimentation for materials discovery include:

• Lila Sciences (Backed by major heavy-hitters, this Cambridge, MA-based startup builds miniaturized robotic instruments and "AI Science Factories" that integrate AI agents with physical synthesis hardware—such as vacuum chambers and sputtering setups—to test and refine novel materials and catalysts in weeks rather than years).
• Radical AI (Integrates AI modeling engines with fully robotic self-driving labs. Their "materials flywheel" takes target performance goals as prompts, autonomously runs high-throughput physical tests on thousands of candidate materials per year, and feeds real-world data straight back into the AI).
• CuspAI (Co-founded by machine learning pioneers like Max Welling, CuspAI acts as an "AI matchmaker" and search engine for advanced materials. Through its AI Materials Foundry initiative, it bridges the gap between atomic-level generation and real-world validation partners to design everything from sustainable chemicals to electronics).
• Atinary Technologies (Based in Switzerland, Atinary deploys an active-learning and no-code AI optimization platform integrated with automated lab robotics (such as Chemspeed) to run autonomous, closed-loop experimentation for chemistry, polymers, and materials formulation).
• Materials Nexus (Uses machine learning and quantum-level property predictions paired with rapid automated feedback loops to design industrial alloys and novel material compositions, drastically shortening the loop from computational discovery to physical qualification).
• Discovered Materials (Focuses heavily on agentic AI workflows connected to physical labs to synthesize, test, and qualify specialized thermal and semiconductor materials for next-generation microelectronics and chips).

Would you like to explore the underlying machine learning architectures (such as graph neural networks or active learning algorithms) these labs use, or are you looking into a specific application like battery electrolytes or catalysts?

Yes. If by “closed-loop experimentation” you mean the system can propose a material → physically synthesize/test it → ingest the experimental result → choose the next experiment, rather than merely using AI to predict materials, the field is still fairly small.

The startups I’d put on the shortlist

A few worth distinguishing

Atinary is probably the best example if you're looking specifically for a commercial “self-driving lab” platform. Its SDLabs system uses ML/Bayesian optimization to select informative experiments and can connect to robotics and lab instrumentation. In 2026 it opened its own Boston lab running continuous Design–Make–Test–Analyze–Learn cycles, with experimental results automatically feeding back into the models.

Kebotix is an earlier pioneer. Its platform explicitly combines AI, physical modeling, chemical informatics and laboratory automation in a closed loop. It has reported using its AI/robotics self-driving lab to search millions of molecules and experimentally identify novel electrochromic materials.

CuspAI is slightly different. Its core strength has been AI-driven generation and simulation of materials, but in 2026 it launched an AI Materials Foundry connecting its models with experimental laboratories. Cambridge describes the network as spanning AI design/simulation through synthesis planning and experimental validation, specifically creating continuous closed-loop workflows.

Orbital Industries (formerly Orbital Materials) takes a more vertically integrated approach: its stated goal is to put materials discovery, hardware design and manufacturing into one AI-driven loop. Its original Linus system was built for materials discovery across batteries, semiconductors, catalysts and carbon capture.

Dunia is particularly interesting if you're looking for electrochemical materials. It explicitly describes a closed-loop Design–Make–Test–Analyze process combining physics-informed AI with robotic electrochemical experimentation.

Periodic Labs is the newer, high-profile entrant. It is building AI scientists plus autonomous robotic laboratories, with the explicit goal of creating a closed loop between hypotheses, experiments and real-world results. Its initial target is high-temperature superconductors.

The important distinction

I'd divide the market into three tiers:

1. True physical closed loop

AI → robotic experiment → measurement → AI → next experiment

Atinary, Kebotix, Dunia, Periodic Labs, Alloryn

2. AI materials discovery + experimental validation, but not necessarily fully autonomous

AI generation/simulation → partner lab → results

CuspAI, Orbital Industries

3. AI materials design/simulation platforms

AI → candidate materials → human-led experiments

Companies such as Matlantis fit more here: its AI simulator is designed to guide experimental discovery, but isn't itself a complete autonomous experimental loop.

The distinction matters because “AI materials discovery” is now a crowded category, while genuinely autonomous materials discovery is much rarer. The defining technology isn't just the model; it's the integration of active learning + robotics + automated characterization + machine-readable experimental data.

If you're evaluating this as a startup/market, I'd consider Atinary, Kebotix, CuspAI, Orbital Industries, and Periodic Labs the five most important companies to investigate first.

Several pioneering startups combine generative artificial intelligence, quantum/atomistic simulations, and robotic self-driving laboratories to execute closed-loop experimentation—meaning the AI generates a candidate molecule or formula, directs a robotic lab to synthesize and test it, and automatically feeds the physical results back into the model to plan the next iteration.

Prominent startups leading this space include:

• Altrove : Designs new materials via AI, generates synthesis recipes, and tests them in automated labs to eliminate dependence on rare earths and critical minerals.
• Radical AI : Builds an integrated "materials flywheel" combining frontier AI architecture with fully robotic, self-driving hardware laboratories to run high-throughput physical tests.
• Kebotix : An early pioneer combining robotic automation with machine learning to navigate end-to-end chemical and material discovery.
• CuspAI : Focuses on generative AI models for inverse material design, aiming to synthesize and screen novel functional compounds for energy and carbon capture.
• Orbital Materials : Develops foundational generative models trained on massive atomic structures to discover advanced materials for clean tech.
• Entalpic : Deploys an Atomistic Discovery Engine that maps quantum chemical space and pairs computational insights with validation workflows.
• Mattiq : Fuses generative algorithms with high-throughput physical screening to accelerate the discovery of functional inorganic materials and catalysts.

If you'd like, I can dive deeper into how self-driving labs handle synthesis bottlenecks or look into the specific funding and industry partnerships for one of these companies.

Yes. If by “closed-loop experimentation” you mean the system actually proposes experiments → runs them with automation/robotics → measures the result → feeds the data back into the AI → chooses the next experiment, rather than merely using AI to screen materials, the startup list is fairly short.

The strongest examples

Kebotix · Dunia · Atinary ·

A useful distinction

There is a much larger group of AI-materials startups that aren't truly closed-loop. For example, companies such as Materials Nexus, CuspAI, Orbital Materials, and MatNex primarily focus on computational discovery, generative models, or candidate screening. Those can be extremely valuable, but they're different from a company that owns or operates the physical experiment → AI → next experiment feedback cycle.

That's why Kebotix and Dunia stand out. Kebotix explicitly describes its system as a closed-loop self-driving lab, while Dunia is building the physical AI/robotics infrastructure around a continuous materials-discovery flywheel.

The really interesting frontier

There is also a distinction between closed-loop materials startups and the broader self-driving-lab movement. Much of the cutting-edge work is still coming from research institutions rather than startups. Berkeley/LBNL's A-Lab, for example, integrates computation, robotics, ML and active learning to autonomously synthesize inorganic materials; it produced 36 compounds from 57 targets during a 17-day autonomous run.

The field is now moving toward what researchers call “Self-Driving Lab 2.0”: more general-purpose agents, interoperable robotics, autonomous characterization, and AI decision-making rather than a single narrowly optimized experiment.

If you're looking at this as an investment/startup landscape, I'd put Kebotix and Dunia in the clearest “commercial closed-loop materials discovery” bucket, Atinary in the “AI/software layer for self-driving labs” bucket, and Periodic Labs in the “new, potentially much broader AI scientist + autonomous lab” bucket.

If you'd like, I can also map 20–30 startups in this space by company, funding, founders, materials focus, autonomous-lab capability, and investors.

The materials science sector is experiencing a major shift from purely computational prediction to closed-loop, autonomous experimentation —where AI models propose novel chemical structures, direct robotic or automated hardware to synthesize and characterize them, and feed real-world physical data back into the AI to refine the next iteration.

Several prominent startups and platforms are leading this space:

• Radical AI — Combines a property-driven generative AI engine with fully robotic self-driving labs. Their "materials flywheel" takes target performance metrics, designs experiments, executes them via physical automation, and uses active learning in real time to bridge synthesis, structure, and property relationships.
• CuspAI — Operates an "AI Materials Foundry" network that combines large-scale curated datasets and generative simulation search engines with actual physical synthesis and characterization capabilities to validate predicted materials on demand.
• Discovered Materials — Focuses specifically on solving thermal management and power-efficiency bottlenecks for next-generation semiconductor chips and GPUs. They deploy AI agents that orchestrate the cycle from multi-modal simulation down to physical synthesis and lab validation.
• Lila Sciences — Utilizes miniaturized, automated hardware modules (such as specialized vacuum deposition and sputtering setups) directed by AI agents to autonomously run physical thin-film and nanomaterial experiments and interpret the results.
• Dunia — Implements a closed-loop Design-Make-Test-Analyze (DMTA) platform. It leverages artificial intelligence paired with robotic precision to accelerate ideation, simulate processing conditions, and capture validated physical data to predict material failure modes early.
• Entalpic — Focuses on molecules and materials for surface-driven industrial and sustainable processes by coupling generative AI and atomistic simulations with tight experimental feedback loops.

Would you like to explore the underlying machine learning architectures (such as graph neural networks or active learning frameworks) used by these platforms, or dive deeper into a specific application like battery electrolytes or semiconductor thermal management?