The history of human civilisation can be read as a history of materials. The Bronze Age. The Iron Age. The Silicon Age. Every technological leap has rested on a physical foundation — the discovery of a substance that could do something no known material could do before. The lithium-cobalt oxide cathode gave us the smartphone. Silicon gave us the computer. Concrete gave us cities.
What it took to find each of these was decades of trial and error, expensive laboratory work, and a great deal of serendipity. Lithium-cobalt oxide was discovered in the early 1980s. It did not power a commercial battery until 1991. The typical journey from laboratory discovery to real-world material is twenty years or more — and costs hundreds of millions of dollars.
In January 2026, Microsoft Research published a paper in Nature introducing MatterGen: a generative AI that does not search through known materials — it creates new ones from scratch, on demand, to specification. It is, as its lead author described, a paradigm shift. And it arrives at a moment when the pressure to find better materials — for batteries, solar cells, carbon capture, and low-emission construction — has never been more urgent.
The question is no longer whether AI can design new materials. It already has. The question is how fast those materials can move from a computer into the world.
The Technology
How MatterGen Works:
From Noise to Crystal
To understand why MatterGen is different, it helps to understand how materials discovery has traditionally worked. The conventional approach is screening: researchers build or query a database of known compounds, filter by desired properties — stability, conductivity, hardness — and narrow down to a handful of candidates worth synthesising in a lab. This process can take years. It also has a hard ceiling: you can only find materials you already know exist.
The DALL-E Analogy
Traditional materials search is like looking for a picture of a cat by scrolling through the entire internet. MatterGen is like DALL-E — you describe what you want ("a stable crystal with high lithium conductivity and no cobalt"), and it generates one directly. The analogy is Microsoft's own: rather than finding needles in a haystack, you grow the needle.
MatterGen is a diffusion model — the same class of AI architecture behind image generators like DALL-E and Stable Diffusion. But instead of operating on pixels, it operates on the three-dimensional geometry of crystal structures: the positions, element types, and periodic arrangements of atoms. It begins with a random, noisy arrangement of atoms, and iteratively refines it — step by step — into a stable, chemically coherent crystalline material that satisfies the design constraints it was given.
The model was trained on 608,000 stable materials drawn from the Materials Project and Alexandria databases — the largest open repositories of computationally validated crystal structures in existence. From this training, it learned the underlying grammar of what makes a material physically possible: which atomic arrangements are energetically stable, which element combinations make chemical sense, and how to navigate the vast space of possible structures without generating nonsense.
The results, as benchmarked in the Nature paper, are striking. Materials produced by MatterGen are more than twice as likely to be both novel and stable compared to previous generative AI approaches, and more than 15 times closer to the local energy minimum — meaning they are far more likely to be physically realisable, not just theoretically plausible.
"MatterGen enables a new paradigm of generative AI-assisted materials design that allows for efficient exploration of materials, going beyond the limited set of known ones."
— Tian Xie, Principal Research Manager, Microsoft Research AI for Science
Microsoft Research · Nature, January 2026
Batteries: The
Lithium Bottleneck
The battery is the central material problem of the energy transition. Every electric vehicle, grid-scale storage system, and consumer device depends on it. And the dominant chemistry — lithium-ion — is approaching its practical limits. Lithium is geographically concentrated, cobalt is ethically fraught, and the energy density of current cells may be within a factor of two of what the chemistry can physically deliver.
The search for what comes next — solid-state electrolytes, sodium-ion alternatives, multivalent chemistries using magnesium or calcium — is one of the most competitive research races on the planet. In one field alone, solid-state batteries, there are fifty new research papers published every single day. No human team can read them all, let alone synthesise the implications.
This is precisely where generative AI enters with force. MatterGen can be prompted to generate candidate materials with high lithium-ion conductivity as an explicit constraint — producing novel solid electrolyte structures that would never appear in a database of known materials. In August 2025, researchers at New Jersey Institute of Technology used a dual-AI generative system to discover five entirely new porous transition metal oxide structures for magnesium-ion batteries — published in Cell Reports Physical Science. One AI generated candidates; the second validated their stability using quantum mechanical simulation.
The practical speed gains are substantial. A 2025 review in Advanced Functional Materials documented that AI closed-loop optimisation reduced fast-charging protocol testing from over 500 days to just 16 days. A battery fire-prevention gel that would have taken a chemical company "many months" was reformulated in two weeks using Citrine Informatics' AI platform.
✦ How AI accelerates battery discovery
- Generate novel solid-state electrolytes with targeted lithium-ion conductivity
- Discover magnesium- and sodium-ion cathode materials free of cobalt and lithium
- Reduce cycle-life testing from months to days via closed-loop AI optimisation
- Predict failure modes, thermal runaway conditions, and degradation pathways
- Identify fire-safe quenching materials and safer electrolyte formulations
→ Remaining challenges
- AI generates candidates; experimental synthesis still requires months per compound
- Scale-up from lab to manufacturing remains a 10–20 year process
- Simulation accuracy (DFT) has known limitations at extreme conditions
- IP complexity when AI generates materials across multiple databases
Nature · Cell Reports Physical Science · Advanced Functional Materials · Citrine Informatics, 2025
Solar Cells:
The Inverse Design Revolution
Silicon solar cells have been the workhorse of the renewable energy transition for decades. But silicon has a ceiling — both in efficiency and in manufacturing flexibility. The next generation of solar technology is built on perovskites: a class of crystal structures that can be deposited as thin films on flexible substrates, manufactured at lower cost, and tuned chemically to absorb different parts of the solar spectrum.
The challenge is finding the right organic molecules to serve as hole-transport layers — the materials that carry electrical charge through the device. From a search space of over one million candidate molecules, researchers at the Karlsruhe Institute of Technology and Helmholtz Institute Erlangen-Nürnberg used an AI-guided inverse design workflow to identify top candidates in just 150 targeted experiments — experiments that would otherwise have required hundreds of thousands of tests.
One of the AI-discovered molecules increased a reference perovskite solar cell's efficiency by approximately two percentage points to 26.2% — a significant jump in a field where fractions of a percent are considered meaningful progress. Published in Science in late 2024, the work demonstrates a broader principle: rather than searching for a good material in a vast space, AI allows researchers to start from the desired outcome — a target band gap, a target charge mobility — and work backwards to the structure. This is what the field calls inverse design.
MatterGen is purpose-built for this mode of thinking. Given a target band gap value — the electronic property that determines which wavelengths of light a material absorbs — it generates crystal structures that satisfy that constraint. In one demonstration in the Nature paper, it generated stable crystalline materials with a target band gap of 3 eV, a property directly relevant to multi-junction solar cell design.
"With only 150 targeted experiments, we were able to achieve a breakthrough that would otherwise have required hundreds of thousands of tests."
— Professor Christoph Brabec, Helmholtz Institute Erlangen-Nürnberg, 2025
Science (2024) · ScienceDaily · KIT / HI ERN · Microsoft Research Nature Paper
Concrete:
The Hidden Emissions Giant
Of all the sectors in which better materials matter, construction is among the least glamorous and among the most consequential. Cement production accounts for roughly 8% of global CO₂ emissions — more than aviation, more than shipping. The process of heating limestone to produce clinker releases carbon dioxide not as a byproduct of burning fuel, but as an inherent chemical reaction. It cannot be decarbonised by simply switching to renewable electricity.
Finding low-carbon alternatives to Portland cement that can match its compressive strength, durability, and cost at scale is one of the materials science field's most urgent and commercially valuable targets. Current options — geopolymers, supplementary cementitious materials, magnesium-based binders — exist but have not displaced the incumbent at scale. The search space for better formulations is enormous.
MatterGen's lead researcher, Tian Xie, cited low-carbon concrete as an explicit target application in the Nature paper and subsequent interviews. The model's ability to generate materials with targeted bulk modulus values — the measure of a material's resistance to compression — is directly relevant. The experimentally validated material produced during the paper's research, TaCr₂O₆, was generated to meet a bulk modulus target of 200 GPa: a mechanical property specification typical of high-performance construction materials.
Concrete is the second-most-consumed substance on Earth after water. Redesigning its chemistry at the atomic level is, without exaggeration, one of the highest-leverage climate interventions available.
The challenge is speed-to-scale. Unlike battery materials, where a new chemistry can be validated in a pouch cell and licensed to a manufacturer relatively quickly, construction materials face decades of building code approval, long-term durability testing, and deeply entrenched supply chains. AI can compress the discovery phase dramatically. It cannot yet compress the adoption phase.
Microsoft Research · Nature Paper · World Economic Forum AI Report, 2025 · IPCC Construction Sector Data
The Competitive Landscape
The Broader Race:
GNoME, MatterSim & the Flywheel
MatterGen does not exist in isolation. It is the most prominent publicly released generative materials AI, but it is one part of a rapidly accelerating field that involves Google DeepMind, Lawrence Berkeley National Laboratory, and a growing ecosystem of startups.
| System / Initiative |
Organisation |
Approach |
Scale / Result |
| MatterGen |
Microsoft Research |
Generative diffusion model — creates novel crystals from property constraints |
Trained on 608K materials; open-source MIT licence; deployed in Azure AI Foundry (Nov 2025) |
| MatterSim |
Microsoft Research |
AI emulator — accelerates property simulation (complements MatterGen) |
Tens of thousands of downloads; "fifth paradigm of scientific discovery" |
| GNoME |
Google DeepMind |
Active learning — iterative prediction of stable crystal structures |
2.2 million new crystal predictions; 381,000 predicted stable; 52,000 lithium-ion conductor candidates |
| A-Lab |
Lawrence Berkeley Natl. Lab |
Autonomous robotic synthesis lab guided by AI — closes the lab-to-synthesis loop |
Demonstrated synthesis of AI-predicted materials without human intervention |
| ReactGen / XtalPi |
Deep Principle / XtalPi |
AI synthesis route discovery + integrated automated laboratories |
XtalPi went public at $2.5B valuation; integrated data-generation and inference flywheels |
| Orbital Materials / DPA-2 |
Orbital Materials / DP Technology |
Large pre-trained machine learning potentials for molecular dynamics |
Publicly released, accelerating simulations across chemistry and materials science |
The most important structural development is what Microsoft calls the "flywheel" between MatterGen and MatterSim. MatterGen generates novel material candidates. MatterSim rapidly simulates their properties — dramatically faster than traditional Density Functional Theory calculations. The results feed back into MatterGen's next generation. Each cycle explores further into unknown material space, at a pace no human team could match.
DeepMind's GNoME takes a different approach — it predicts which crystal structures are stable, expanding the known universe of viable materials from roughly 20,000 known inorganic compounds to over 2.2 million predicted candidates. Among them: 52,000 lithium-ion conductors, compared to the 1,000 previously known. This is not a marginal improvement. It is a categorical expansion of the search space.
Microsoft Research · Google DeepMind · Lawrence Berkeley National Lab · World Economic Forum, June 2025
The Real-World Test:
TaCr₂O₆
Scientific claims about generative AI's power are easy to make and hard to verify. Microsoft's Nature paper included an attempt at honest ground-truthing: a collaborative experiment with researchers at the Shenzhen Institutes of Advanced Technology, part of the Chinese Academy of Sciences, to actually synthesise a material that MatterGen had designed.
The target: a novel compound with a bulk modulus of 200 GPa — a mechanically hard material — using no rare or critical elements. MatterGen generated TaCr₂O₆, a tantalum-chromium oxide structure with no precedent in any known materials database.
The lab synthesised it. The measured bulk modulus was 169 GPa — approximately 20% below the target. The researchers noted that the final material displayed compositional disorder between tantalum and chromium atoms, but that its overall crystal structure aligned closely with MatterGen's prediction. A 20% discrepancy is, in the context of novel materials synthesis, a remarkably small error. Most materials that look promising on a computer never survive contact with a laboratory at all.
Why the error matters — and why it doesn't
A 20% gap between AI prediction and experimental result sounds like a failure. In materials science, it is a promising beginning. Computational screening tools routinely produce larger errors. The point is not that MatterGen predicts perfectly — it doesn't. The point is that it generates physically realisable structures in regions of chemical space that no previous method could reach, and that those structures survive experimental synthesis. The errors can be corrected with iteration.
Early 2022
MatterGen development begins
Tian Xie joins Microsoft Research AI for Science initiative under Chris Bishop. Work on generative materials AI starts as a two-person project.
June 2024
MatterGen presented at NeurIPS
Early version of MatterGen demonstrated at NeurIPS 2023; Microsoft Research blog post announces property-guided materials design capability.
January 2026
Nature paper published; Satya Nadella announces
MatterGen published in Nature. TaCr₂O₆ experimental validation included. Microsoft CEO announces on X: "a paradigm shift in materials design." Source code released under MIT licence.
November 2025
MatterGen deployed in Azure AI Foundry
MatterGen made available to researchers and companies through Azure AI Foundry Labs. Trained on Materials Project and Alexandria datasets. Commercial access begins.
Microsoft Research · Nature, January 2026 · Shenzhen Institutes of Advanced Technology, CAS
The Lab-to-Market Gap:
What AI Cannot Yet Solve
The history of materials science is littered with promising discoveries that never left the laboratory. Graphene won a Nobel Prize in 2010. It has found commercial use in some specialised applications, but has not yet delivered the transformative products predicted for it. Metal-organic frameworks — the subject of the 2024 Nobel Prize in Chemistry — are among the most scientifically remarkable materials ever devised, with extraordinary adsorption properties. Their commercial impact remains limited.
One veteran science journalist, writing in MIT Technology Review in December 2025, put the challenge plainly: in nearly 40 years of covering materials discovery, they could count the truly transformative commercial breakthroughs on one hand. AI compresses the discovery phase. It does not compress the 20 years and hundreds of millions of dollars required to move from laboratory synthesis to manufactured product at industrial scale.
There are also hard technical limits. MatterGen's predictions are currently validated using Density Functional Theory — a computational method that is highly accurate but has known limitations at extreme temperatures, pressures, and in strongly correlated electron systems. Experimental synthesis introduces additional complexity: atomic disorder, defects, surface chemistry, and processing variables that no current AI fully models. As one researcher framed it: "The AI revolution is about finally gathering all the scientific data we have. Turning that data into products is a different problem entirely."
The most honest assessment of where the field stands is this: AI has dramatically shortened the distance from idea to candidate material. It has not yet dramatically shortened the distance from candidate to product. The bottleneck has moved — from discovery to validation, synthesis, scale-up, regulation, and market adoption. Those are human problems as much as scientific ones.
"I don't think we get there — the energy transition — by a long shot, without big advances in materials. It's not just in one area. It's not just in concrete."
— Greg Mulholland, CEO, Citrine Informatics, The Carbon Copy Podcast, 2025
MIT Technology Review, December 2025 · Citrine Informatics · Latitude Media / The Carbon Copy
What Comes Next
The Road Ahead:
Five Paths That Matter
-
01
Close the synthesis loop with autonomous labs
Lawrence Berkeley's A-Lab demonstrated that AI-predicted materials can be synthesised by robotic systems without human intervention. The next phase is coupling MatterGen's generation capability directly to autonomous synthesis pipelines. When the computer can design a material and the robot can make it — overnight, unattended — the pace of discovery accelerates by another order of magnitude.
-
02
Prioritise equitable deployment over discovery speed
The technology to model, predict, and optimise materials exists. The challenge for energy and food systems is deployment in the places that need it most — AI-optimised irrigation in drought-stressed sub-Saharan Africa, better solar cells in tropical developing nations. Discovery without equitable access produces asymmetric benefits.
-
03
Build experimental feedback into training data
MatterGen's current training dataset is overwhelmingly computational — theoretical structures, not experimentally measured ones. As more AI-designed materials are synthesised, the experimental results — including failures — should flow back into training data. Models that learn from what does not work in a lab will be dramatically more useful than those trained only on theoretically stable structures.
-
04
Massively fund the translation gap, not just discovery
The discovery phase of materials science is becoming well-funded and increasingly AI-augmented. The 10-to-20-year translation phase — characterisation, process engineering, manufacturing scale-up, regulatory approval — is not. If the bottleneck is not funded, faster discovery simply creates a larger queue of promising materials that never reach the world.
-
05
Govern dual-use materials AI with the same urgency as biology
A model that can design materials with targeted mechanical, electronic, or chemical properties is, in principle, a model that can be prompted for harmful applications — high-performance energetic materials, compounds with specific toxicological profiles. The AI materials field has not had the same biosecurity governance conversation that AI biology tools have triggered. It should begin now, before capabilities outpace oversight.
Buy me a coffee