Explore the research

Materials

Can AI find and test materials that work outside the laboratory, from solar cells to new devices?

26%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Synthesis success30%
Scale and durability30%
Identity and novelty15%
Functional improvement40%
Discovery efficiency25%
Deployment value15%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Deployment value

15%estimated progress

What we track

Cost per qualified unit and lifetime performance against the existing material in a representative device.

Why this estimate?

A mini-module connects laboratory discovery with a device, but commercial unit cost, field lifetime and self-managed deployment economics are unknown.

Advancements

Microsoft and PNNL build a working battery with an AI-found low-lithium electrolyteJan. 10, 2024. A solid-state electrolyte identified by AI screening uses up to 70% less lithium than traditional lithium-ion batteries; discovery through a working battery took under nine months. Researchers cautioned that it may not work at scale.
Meta and Amrize place AI-optimized low-carbon concrete in a data center slabJuly 16, 2025. Meta reports that concrete mixes designed with multi-objective Bayesian optimization (BoTorch/Ax), from a pipeline that generated over a hundred unique mixes, were deployed in a site support section of a building slab at its Rosemount, Minnesota data center. Amrize's project page reports an estimated 35% lower concrete carbon footprint and 43% faster early strength development.
MIT CRESt catalyst gives 9.3 times the power density per dollar of palladiumSeptember 30, 2025. A multielement catalyst found by the CRESt platform achieved a 9.3-fold improvement in power density per dollar over pure palladium in a fuel cell, using about one-fourth of the precious metals of previous devices. Published in Nature.
AI-optimized perovskite is extended to a mini-moduleApril 14, 2026. The perovskite study extends its laboratory approach to a mini-module.

Read the dated findings and sources

Status: Device performance demonstrated; commercial value unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

April 14, 2026: The perovskite study extends its laboratory approach to a mini-module. Autonomous perovskite optimization, Nature.

Interpretation and tracking value

No fully loaded cost per qualified commercial unit or field-lifetime comparison is established by that demonstration.

Next comparable measurement

Compare qualified modules against incumbent technology using yield-adjusted manufacturing costs, installation conditions, warranty-relevant degradation and lifetime delivered energy. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Discovery efficiency

25%estimated progress

What we track

Days, human hours, experiments, and total dollars from target specification to independently reproduced sample meeting all property requirements.

Why this estimate?

Autonomous laboratory workflows support early efficiency capability, but corrected outcome counts and missing all-in independent-reproduction costs constrain the estimate.

Advancements

USTC robotic AI chemist finds a Mars-meteorite oxygen catalyst within two months13 November 2023. A robotic AI chemist made an oxygen-evolution catalyst from five types of Martian meteorites without human intervention, using a model built from about 30,000 simulated high-entropy hydroxides; the release says the search took within two months and estimates 2,000 years for a human chemist. Published in Nature Synthesis.
Microsoft and PNNL screen 32 million battery candidates in under 80 hours8 January 2024. An AI and cloud high-performance-computing workflow screened more than 32 million candidates, predicted around half a million potentially stable materials and narrowed them to 18 promising candidates in less than 80 hours on about one thousand virtual machines; the top candidates (the NaxLi3-xYCl6 series) were synthesized and characterized.
Microsoft Discovery finds a coolant prototype candidate in about 200 hoursMay 19, 2025. Microsoft reports that its agentic Discovery platform identified a new data-center immersion coolant candidate in about 200 hours of digital work; the prototype was synthesized in under four months and initial property tests matched the AI predictions. It remains under analysis and refinement.
NC State dynamic-flow self-driving lab collects at least 10 times more dataJuly 14, 2025. A self-driving fluidic lab using dynamic flow experiments generated at least 10 times more data than steady-state flow labs over the same period, and the release says fewer experiments reduce chemical use and waste (no figure given). Published in Nature Chemical Engineering.
MIT CRESt platform runs 3,500 electrochemical tests on 900 chemistries in 3 monthsSeptember 30, 2025. The CRESt system, which combines multimodal models with robotic synthesis and testing, explored more than 900 chemistries and ran 3,500 electrochemical tests over three months to find a fuel-cell catalyst. Published in Nature.
A-Lab correction revises which syntheses count as confirmedJanuary 19, 2026 correction to November 29, 2023 paper. The January 2026 amendment changes which A-Lab synthesis outcomes can be counted as confirmed.

Read the dated findings and sources

Status: Historical campaign requires corrected outcome accounting. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 19, 2026 correction to November 29, 2023 paper: The January 2026 amendment changes which A-Lab synthesis outcomes can be counted as confirmed. A-Lab author correction, Nature.

Interpretation and tracking value

A historical campaign's duration cannot be divided by its unrevised success total. No all-in time/cost baseline per independently reproduced qualified material is established here.

Next comparable measurement

Use the corrected campaign log; count attempts, staff time and full costs through outside synthesis and property verification. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Functional improvement

40%estimated progress

What we track

Application-specific measurements such as battery Wh/kg and cycle life, conductivity at specified temperature, or strength at specified density. Fix test conditions before selection.

Why this estimate?

Measured and certified device efficiency supports strong bounded optimization. The result is application-specific and does not establish autonomous development across material classes.

Advancements

Self-driving labs in five sites find 21 top-performing organic laser emittersMay 16, 2024. A cloud-coordinated closed loop across five laboratories made and tested over 1,000 candidate molecules and found at least 21 top-performing organic solid-state laser gain materials; gram-scale synthesis verified best-in-class stimulated emission in a thin-film device. Published in Science.
Closed-loop AI experiments yield light-harvesting molecules up to 4x more stableAugust 28, 2024. Five rounds of closed-loop experimentation by University of Illinois Urbana-Champaign and University of Toronto researchers produced 30 new chemical candidates, with molecules up to four times more light-stable than the starting point. Published in Nature.
MatterGen-designed TaCr2O6 is synthesized and measures 169 GPa against 200 targetJanuary 16, 2025. A material generated by Microsoft's MatterGen for a target bulk modulus of 200 GPa, TaCr2O6, was synthesized by a team at the Shenzhen Institutes of Advanced Technology; the measured bulk modulus was 169 GPa, a relative error below 20%, with compositional disorder between Ta and Cr. Published in Nature.
Toronto team's Bayesian-optimized carbon nanolattices reach 2.03 MPa m3/kgJanuary 24, 2025. Nanolattice designs optimized with multi-objective Bayesian optimization from about 400 data points withstood a stress of 2.03 megapascals for every cubic metre per kilogram of density, about five times higher than titanium. Published in Advanced Materials.
Argonne's Polybot produces polymer films averaging over 4,500 S/cmFebruary 17, 2025. The autonomous Polybot platform optimized processing in a seven-dimensional space and produced transparent conductive polymer thin films with averaged conductivity exceeding 4,500 S/cm, along with recipes intended for scale-up. Published in Nature Communications.
AI-guided perovskite cell reaches 27.22% efficiencyApril 14, 2026. AI-guided perovskite passivation achieved 27.22% efficiency on a 0.05 cm² cell, with certified maximum-power-point tracking efficiency of 27.18%.

Read the dated findings and sources

Status: Measured photovoltaic performance. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

April 14, 2026: AI-guided perovskite passivation achieved 27.22% efficiency on a 0.05 cm² cell, with certified maximum-power-point tracking efficiency of 27.18%. Autonomous perovskite optimization, Nature.

Interpretation and tracking value

This is an application-specific device result, not a generic material-improvement percentage. Certification of one result does not establish mass-production yield.

Next comparable measurement

Retest equivalent device area, illumination and stabilization protocol; report certified efficiency, device counts and distribution rather than only the best device. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Identity and novelty

15%estimated progress

What we track

Count materials with measured composition and structure, prior-art search, and independent review distinguishing new phases from known materials or routine substitutions.

Why this estimate?

The corrected evidence demonstrates the need to separate platform novelty from scientific novelty; autonomous identification and prior-art adjudication remain weakly established.

Advancements

GNoME predicts 2.2 million crystals; 736 already made independentlyNovember 29, 2023. Google DeepMind's GNoME predicted 2.2 million new crystal structures, of which 380,000 are the most stable; external researchers had independently created 736 of the predicted structures in the lab.
PRX Energy analysis concludes A-Lab discovered no new materialsMarch 2024. Leeman and colleagues examined all 43 synthetic products reported by the A-Lab and concluded that no new materials had been discovered; about two-thirds of the claimed successes are likely known compositionally disordered versions of the predicted ordered compounds.
Cheetham and Seshadri find little novelty in sampled GNoME compoundsApril 8, 2024. A Chemistry of Materials perspective reviewed the first 250 of 2,047 GNoME Explorer entries and 10 random entries from the 384,870-entry stable-structure database; all 10 matched structures already in the ICSD, and the authors found little evidence of compounds meeting novelty, credibility and utility together.
MIT SCIGEN-guided generation leads to two newly synthesized compoundsSeptember 22, 2025. Using SCIGEN to constrain a generative model to Archimedean lattices, researchers generated over 10 million candidates, screened about 1 million for stability, simulated 26,000 and synthesized two previously undiscovered compounds, TiPdBi and TiPbSb, whose measured properties largely aligned with predictions. Published in Nature Materials.
A-Lab clarifies its materials were new to its platformJanuary 19, 2026 correction to November 29, 2023 paper. The correction clarifies that A-Lab's materials were new to its prediction platform, not necessarily new to science.
GPT-4o-based agent crystallizes COF-2000, a previously unmade frameworkMarch 9, 2026. A GPT-4o-based agent from Omar Yaghi's group, refining conditions after each round of 96-well screening, raised crystallinity of a benchmark covalent organic framework by 350% and produced COF-2000, a material not made before. Published in the Journal of the American Chemical Society.

Read the dated findings and sources

Status: Corrected claim; substantial novelty not established. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 19, 2026 correction to November 29, 2023 paper: The correction clarifies that A-Lab's materials were new to its prediction platform, not necessarily new to science. A-Lab author correction, Nature.

Interpretation and tracking value

Do not retain the original headline as a count of novel materials. Identity, platform novelty and scientific novelty need separate adjudication.

Next comparable measurement

Require composition/structure evidence, a dated prior-art search and independent classification before adding each new material to the count. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Scale and durability

30%estimated progress

What we track

Yield, batch variability, degradation, and property retention when production scale increases.

Why this estimate?

Mini-module and stability testing provide real engineering progress. They do not yet establish autonomous scaling with representative manufacturing yield and field durability.

Advancements

AI-chemist Mars-meteorite catalyst runs over 550,000 seconds without degradation13 November 2023. The oxygen-evolution catalyst selected by USTC's robotic AI chemist operated steadily for over 550,000 seconds at 10 mA/cm2 with an overpotential of 445.1 mV, and produced oxygen at -37 C without apparent degradation.
AP-Lab autonomous pilot platform scales nanoparticle reagents to 500 g batchesFebruary 12, 2026. An AI-driven pilot-scale platform optimized magnetic-nanoparticle nucleic-acid extraction reagents in about three weeks, then scaled to 500 g of nanoparticle modification and 400 L of reagents per batch; four scale-up batches showed uniform PCR performance, with consistency scores of at least 0.9 versus 0.5 to 0.7 for human-in-the-loop runs. Published in Advanced Science.
Perovskite mini-module keeps 98.7% of its output after 1,200 hoursApril 14, 2026. The study reports 23.49% efficiency on a 21.4 cm² mini-module and 98.7% retention after 1,200 hours under ISOS-L-1I testing.

Read the dated findings and sources

Status: Measured scale and stability demonstrations. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

April 14, 2026: The study reports 23.49% efficiency on a 21.4 cm² mini-module and 98.7% retention after 1,200 hours under ISOS-L-1I testing. Autonomous perovskite optimization, Nature.

Interpretation and tracking value

Do not assume the durability test and mini-module refer to identical specimens. Laboratory accelerated testing is not a decades-long field lifetime.

Next comparable measurement

Track device-specific test conditions and sample counts; scale production while measuring yield, variability, outdoor degradation and retained performance. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Synthesis success

30%estimated progress

What we track

Independently confirmed target materials / all selected candidates attempted, with unsuccessful recipes retained.

Why this estimate?

Corrected synthesis confirmations support bounded laboratory execution. Inconclusive identities and missing independent replication of all attempts limit validation credit.

Advancements

A-Lab reports making 41 of 58 target compounds in 17 daysNovember 29, 2023. Berkeley Lab's autonomous A-Lab synthesized 41 of 58 attempted target compounds over 17 days of independent operation. Later analyses disputed how many were correctly identified or new.
LLM-interfaced robot at Xiamen synthesizes 13 inorganic compoundsJune 17, 2025. A natural-language-interfaced robotic platform synthesized 13 compounds across four classes (coordination complexes, metal-organic frameworks, nanoparticles and polyoxometalates) and its AI-copilot-guided exploration found a previously unreported family of Mn-W polyoxometalate clusters. Published in the Journal of the American Chemical Society.
Reanalysis supports 36 of 40 A-Lab synthesesJanuary 19, 2026 correction to November 29, 2023 paper. Manual reanalysis supported 36 of 40 reported platform successes; four were inconclusive by diffraction. A training-data compound was also removed from discussion.

Read the dated findings and sources

Status: Corrected measured identification result. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

January 19, 2026 correction to November 29, 2023 paper: Manual reanalysis supported 36 of 40 reported platform successes; four were inconclusive by diffraction. A training-data compound was also removed from discussion. A-Lab author correction, Nature.

Interpretation and tracking value

36/40 (90%, project calculation) is confirmation among reported successes, not success among all attempted candidates, and is not independent laboratory replication.

Next comparable measurement

Reconstruct all attempted targets and corrected outcomes; report synthesis success/all attempts separately from identity confirmation and outside replication. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note