Explore the research

Chemistry

Can AI plan and run experiments that improve quality, reduce waste and scale up?

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.

Yield and selectivity35%
Transfer and scale30%
Substantial novelty15%
Route execution30%
Process burden10%
Experimental efficiency35%

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.

Experimental efficiency

35%estimated progress

What we track

Number of experiments, elapsed days, human hours, and total dollars to meet a frozen yield and purity specification.

Why this estimate?

A logged optimization campaign and instrument feedback support bounded automated experimentation. Full campaign economics and autonomous choice of research directions remain gaps.

Advancements

BMS Bayesian optimizer beats a 500-experiment campaign with 24 experiments2024. In work recognized by the ACS Green Chemistry Institute Pharmaceutical Roundtable's 2024 data science award, Bristol Myers Squibb's EDBO+ reached 80% yield and 91% ee in 24 experiments, where one-factor-at-a-time optimization had reached 70% yield and 91% ee in 500 experiments.
Chemma LLM helps optimize an unreported Suzuki coupling in 15 runs25 April 2025. Chemma, a language model fine-tuned on 1.28 million reaction question-answer pairs, was used in a human-AI loop that found a suitable ligand and solvent (1,4-dioxane) for an unreported Suzuki-Miyaura coupling of cyclic aminoboronates with aryl halides within 15 runs, reaching 67% isolated yield.
Minerva ML optimizes a pharma Buchwald-Hartwig step in 4 weeks, not 6 months12 July 2025. Roche and EPFL's Minerva framework, running 96-reaction automated batches, optimized a Pd-catalysed Buchwald-Hartwig step for an active pharmaceutical ingredient in about 4 weeks, compared with a previous 6-month development campaign. The study released 1,632 high-throughput reactions.
GPT-5.4 and Molecule.one's Maria lab run 10,080 reactions in two cyclesJune 17, 2026. OpenAI and Molecule.one report a near-autonomous workflow in which GPT-5.4, paired with the Maria chemistry AI and lab, ran 10,080 Chan-Lam coupling reactions across two experimental cycles, from a first prompt on March 4 to shared results on June 4. Humans selected proposals, corrected some details and validated results by hand.
RoboChem-Flex optimizes a labeling reaction in 37 experiments2026 publication. In photocatalytic isotope-labeling case study 3, optimization used seven initial experiments followed by 30 optimization iterations.

Read the dated findings and sources

Status: Affordable optimization platform; full campaign metric incomplete. 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

2026 publication: In photocatalytic isotope-labeling case study 3, optimization used seven initial experiments followed by 30 optimization iterations. RoboChem-Flex, Nature Synthesis.

Interpretation and tracking value

This gives a 37-experiment campaign baseline (project sum), not all-in dollars or time to an independently reproduced specification. Hardware affordability alone is not campaign efficiency.

Next comparable measurement

Run matched optimization campaigns; count failed experiments, consumables, analytics, operator time and elapsed days until the same specification is reached. 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

Process burden

10%estimated progress

What we track

Kilograms of waste per kilogram of product, solvent use, energy per batch, and total production cost.

Why this estimate?

Affordable automation is an enabling step, but no verified waste, energy or production-cost improvement per qualified kilogram is established in the baseline.

Advancements

Synthia-planned API route lifts overall yield from 8% to 26%October 2, 2024. For a helicase-primase inhibitor, a Synthia-proposed and experimentally validated route gave 26% overall yield versus 8% for the patented route, with building-block cost about 556-fold lower; a human-refined version reached 35% in six steps instead of seven and raised the GreenMotion score from 0 to 18. An unconstrained Synthia route failed, with one step giving only about 5%. No PMI values are reported.
Lonza AI route scouting cuts a seven-step synthesis to four stepsMay 22, 2025. A Lonza case study of AI-enabled route scouting reduced a route from seven steps and seven isolations to four of each, cut the required starting material by 70%, found starting materials priced 50% lower and saved four to six weeks of lab time.
RoboChem-Flex offers lower-cost lab automationApril 13, 2026. RoboChem-Flex provides a lower-cost automation setup for reaction optimization.

Read the dated findings and sources

Status: Required process-intensity baseline 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 13, 2026: RoboChem-Flex provides a lower-cost automation setup for reaction optimization. University of Amsterdam: RoboChem-Flex.

Interpretation and tracking value

Lower equipment purchase cost does not imply less waste, solvent, energy or production cost per kilogram. Those quantities are not established by the platform headline.

Next comparable measurement

Weigh all material inputs and waste; meter batch energy and calculate complete cost per kilogram of isolated qualified product. 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

Route execution

30%estimated progress

What we track

Target compounds obtained at required purity / all assigned synthesis targets, confirmed experimentally.

Why this estimate?

Computer-directed reaction execution is demonstrated on selected tasks. Reliable completion of diverse multi-step targets, purity validation and recovery still lack complete autonomous evidence.

Advancements

ChemCrow agent plans and executes syntheses of a repellent and 3 organocatalysts11 April 2023. ChemCrow, a language-model agent with 18 expert-designed chemistry tools, autonomously planned and executed the syntheses of an insect repellent and three organocatalysts, and guided the discovery of a novel chromophore.
Coscientist plans and runs palladium coupling experimentsDecember 20, 2023. Coscientist demonstrated planning and execution capabilities including palladium-catalysed coupling experiments.
MIT autonomous platform makes 294 previously unreported dye moleculesDecember 22, 2023. An integrated design-make-test-analyze platform realized 294 unreported molecules over three automatic iterations in one case study, and identified nine top-performing molecules in a second. Published in Science.
Liverpool mobile robots run and assess exploratory synthesis autonomouslyNovember 6, 2024. Two 1.75-metre mobile robots carried out reactions, product analysis and next-step decisions in structural diversification, supramolecular host-guest and photochemical synthesis; the AI logic's decisions were the same as or similar to a human chemist's and far faster. Published in Nature.
Yale's MOSAIC procedures yield over 35 new compounds at 71% successJanuary 19, 2026. MOSAIC, built from 2,498 specialized AI chemistry experts, generated experimental procedures used to synthesize more than 35 previously unreported compounds; the Nature abstract reports a 71% overall success rate.

Read the dated findings and sources

Status: Laboratory task demonstrations; target success denominator incomplete. 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

December 20, 2023: Coscientist demonstrated planning and execution capabilities including palladium-catalysed coupling experiments. Coscientist, Nature.

Interpretation and tracking value

Selected tasks do not establish a broad fraction of assigned multi-step synthesis targets obtained at required purity. Tool execution and final compound acceptance are different endpoints.

Next comparable measurement

Assign a fixed diverse target set; verify identity, isolated purity and yield, counting all failures, manual repairs and abandoned routes. 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

Substantial novelty

15%estimated progress

What we track

New transformations or materially different routes accepted after literature and patent review; distinguish routine parameter tuning.

Why this estimate?

Planning and execution provide foundations, but known transformations and parameter optimization do not establish independent discovery of substantively new chemistry.

Advancements

Coscientist demonstrates computer-directed chemistryDecember 20, 2023. Coscientist demonstrated computer-directed chemistry using known reaction capabilities.
Grzybowski group's algorithm designs new multicomponent reactions2024. A computer system given reaction-mechanism knowledge and physical-organic rules autonomously designed large numbers of mechanistically distinct multicomponent reactions, predicted yields and flagged organocatalytic candidates; predictions were validated by experiments across different reactivity modes and product scaffolds. Published in Nature Communications.

Read the dated findings and sources

Status: Autonomous execution evidence; qualifying novelty count 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

December 20, 2023: Coscientist demonstrated computer-directed chemistry using known reaction capabilities. Coscientist, Nature.

Interpretation and tracking value

Executing or optimizing an existing transformation does not meet the substantial-novelty rule. No qualifying count of independently adjudicated new transformations is extracted here.

Next comparable measurement

Register proposals before experimentation; obtain literature/patent review and outside experimental reproduction before counting a new route or transformation. 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

Transfer and scale

30%estimated progress

What we track

Success on unseen substrate families and retention of yield, purity, and reproducibility at increased scale.

Why this estimate?

An eightfold laboratory scale-up provides concrete transfer evidence. Small-scale demonstrations do not establish unseen-family success or autonomous industrial process development.

Advancements

Minerva conditions hold at gram scale, cutting an impurity from 3.8% to <0.5%12 July 2025. Conditions found by Roche and EPFL's Minerva optimizer for a Pd-catalysed Buchwald-Hartwig API step were run at gram scale, reducing an impurity from 3.8% to below 0.5% with over 99% conversion.
AI-proposed TEMPO additive improves 11 of 14 substrate pairs at bench scaleJune 17, 2026. In the OpenAI and Molecule.one Chan-Lam study, optimized conditions with TEMPO improved yields for 88% of boronic acids and 83% of sulfonamides tested; at bench scale 11 of 14 substrate pairs gave higher yields, eight of them more than double.
A RoboChem-Flex reaction is scaled up eightfold2026 publication. Case study 3 was scaled eightfold to 1 mmol, yielding isolated product at 52% yield and 64% deuterium incorporation.

Read the dated findings and sources

Status: Multi-case demonstration; industrial-scale retention 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

2026 publication: Case study 3 was scaled eightfold to 1 mmol, yielding isolated product at 52% yield and 64% deuterium incorporation. RoboChem-Flex, Nature Synthesis.

Interpretation and tracking value

This is laboratory scale, not industrial manufacturing or a held-out substrate-family test. Scale-up is measured, broad transfer remains unquantified.

Next comparable measurement

Hold out substrate families; repeat accepted routes at increasing batch sizes, measuring yield, purity and between-batch variability. 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

Yield and selectivity

35%estimated progress

What we track

Isolated yield and measured selectivity under matched conditions; distinguish isolated product from inferred instrument signal.

Why this estimate?

Closed-loop optimization and isolated product measurements support useful bounded execution. The gap between instrument and isolated yield limits claims of dependable autonomous validation.

Advancements

RoboChem gives better yields than published work in about 80% of cases25 January 2024. The University of Amsterdam's RoboChem platform replicated work from four randomly selected papers and produced better yields in roughly 80% of cases and similar yields in the rest; it can optimize about 10 to 20 molecules in a week. Published in Science.
Minerva finds 76% yield where two chemist-designed plates found none12 July 2025. In a 96-well campaign on a nickel-catalysed Suzuki reaction, Minerva reached up to 76% area-percent yield and 92% selectivity while two chemist-designed plates found no successful conditions; for two API syntheses it found multiple conditions above 95 area-percent yield and selectivity. Values are area percent, not isolated yields.
GPT-5.4-proposed additive raises mean Chan-Lam yield from 16.6% to 25.2%June 17, 2026. In OpenAI and Molecule.one's near-autonomous Chan-Lam coupling study, adding TEMPO raised mean yield from 16.6% to 25.2% across the high-throughput set, and the share of reactions above 30% yield rose from 15.6% to 37.5%.
RoboChem-Flex reaches about 70% yield in a labeling reaction2026 publication. Case study 3 reached an instrument-estimated 70% ±10% yield and 71% deuterium incorporation; an isolated scale-up produced 52% yield with 64% incorporation.

Read the dated findings and sources

Status: Experimental optimization demonstrated; common numerical baseline unavailable. 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

2026 publication: Case study 3 reached an instrument-estimated 70% ±10% yield and 71% deuterium incorporation; an isolated scale-up produced 52% yield with 64% incorporation. RoboChem-Flex, Nature Synthesis.

Interpretation and tracking value

Instrument estimates and isolated yields are different endpoints. Deuterium incorporation is specific to this transformation, not a generic selectivity rate.

Next comparable measurement

For each reaction report starting conditions, isolated mass/purity, selectivity and uncertainty against an expert baseline under matched conditions. 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