Closing the loop: our approach to autonomous formulation discovery

Abstract

Personal care products are complex, multi-component soft materials. Their performance emerges from the interactions of dozens of ingredients and from the process used to combine them, and formulating them remains a largely empirical craft. This article reviews the research that makes a different approach possible: self-driving laboratories that pair robotics with sequential experimental design, machine learning models that predict molecular and material properties, language-model agents that plan and reason over scientific knowledge, and a small but growing body of work that applies these ideas to formulations. We then describe the approach Kodimo is taking, a closed loop that turns a product brief into candidate formulations, tests them in a robotic wet lab and learns from every result, and we close with the ways it could fail and what we are doing about each.

1Formulation is a hard search problem

A shampoo or a moisturizer looks simple on the shelf. Inside the bottle it is a carefully balanced system: ten to thirty ingredients drawn from a palette of thousands of surfactants, emollients, humectants, polymers, preservatives, actives and fragrances. What the product does (whether it stays stable for two years, how it flows out of the pump, how it foams, how it feels on skin, whether it is safe, what it costs) emerges non-linearly from both the composition and the process. The same ingredients combined in a different order, at a different temperature or under a different shear rate can give a different product.

That makes formulation a search problem with four unpleasant properties. The space is combinatorial: ingredient choice is categorical, concentrations are continuous and coupled by the constraint that everything sums to one hundred percent. The objectives conflict: mildness against cleansing power, richness against a light feel, performance against cost. Evaluation is slow and noisy: a single stability study can take weeks of accelerated ageing. And most of what the industry knows lives in lab notebooks and in the experience of senior formulators rather than in structured, reusable data.

The last point matters more than it seems. In a survey of more than 1,500 researchers, a majority reported having failed to reproduce another scientist's experiment, and many had failed to reproduce their own [1]. Small effect sizes and flexible analyses make it easy for published findings to be wrong [2]. Manual, sparsely logged experimentation makes both problems worse.

What has changed is that three lines of research have matured at the same time: robots that can run experiments reliably, machine learning models that can predict properties and choose experiments, and language models that can read and reason over the scientific literature [3, 4]. The rest of this section reviews each of them, and then the work that applies them to formulations.

2What the literature tells us

2.1From robot scientists to self-driving laboratories

The idea that a machine could run the scientific method end to end is older than today's AI boom. In 2009, the robot scientist Adam autonomously generated hypotheses about gene function in yeast, designed and ran the experiments to test them, and interpreted the results [5]. A decade later, the concept was formalized as the self-driving laboratory: a closed loop in which an algorithm plans an experiment, automation executes it, and the result updates the plan [6]. Coley, Eyke and Jensen framed the discovery of new molecules, materials and processes as a single search problem and mapped where automation could help at each step [7, 8].

The results since then have been striking. Ada, a modular self-driving lab, autonomously optimized thin-film materials for organic electronics [9]. A free-roaming mobile robot chemist ran 688 experiments over eight days across a ten-variable space and found photocatalyst mixtures six times more active than the starting point [10]. That result deserves attention from anyone who works on formulations: it was, in effect, a mixture optimization problem, solved by a robot guided by batched Bayesian optimization. More recently, cooperating mobile robots have run exploratory synthetic chemistry, using the same NMR and mass spectrometry instruments that human chemists use, and made decisions about what to try next [11]. Reviews in Nature Synthesis and Chemical Reviews catalogue dozens of such systems [4, 12].

Year	System	What it showed	Ref.
2009	Adam, robot scientist	Autonomous hypothesis generation and testing in yeast functional genomics	[5]
2020	Ada	Closed-loop optimization of thin-film materials	[9]
2020	Mobile robotic chemist	688 experiments in 8 days; mixtures 6× more active than baseline	[10]
2021	Robotic formulation platform	ML-driven design of experiments for liquid formulations	[13]
2023	A-Lab	Autonomous solid-state synthesis guided by computation and literature	[14]
2023	Coscientist	Language-model agent planning and running reaction optimization	[15]
2024	Mobile robots for synthesis	Exploratory chemistry with shared, unmodified instruments	[11]
2024	High-throughput shampoo dataset	812 formulations with stability and rheology labels	[16]

Across these systems, a pattern repeats. Closed loops work best when three conditions hold: the measurement is fast and informative, the search space is parameterized so that an algorithm can move through it, and the decision-making is sample-efficient enough that every experiment counts. Formulation can meet all three, but only if the lab is designed for it from the start.

2.2Learning the structure–property map

The second thread is the use of machine learning to predict what a molecule or material will do before anyone makes it. Learned graph representations of molecules now routinely match or outperform hand-crafted descriptors for property prediction [17], and treating reaction prediction as a translation problem pushed top-1 accuracy above ninety percent on a standard benchmark [18]. Generative models go further and propose new structures: an early variational autoencoder mapped molecules into a continuous space that could be searched with gradients [19], and diffusion models now generate stable inorganic crystals that satisfy target constraints [20].

Scale changes what is possible. Graph networks trained at scale identified about 2.2 million candidate crystal structures, including roughly 381,000 predicted to be newly stable [21]. A deep learning screen of chemical libraries found halicin, a structurally unusual antibiotic active against drug-resistant pathogens [22], and AlphaFold showed that learned models can reach near-experimental accuracy on one of biology's hardest prediction problems [23]. Even general-purpose language models, fine-tuned on small chemistry datasets, can match specialized models in the low-data regime [24], which is exactly the regime formulation lives in.

The lesson we take from this body of work is the one Rich Sutton called the bitter lesson: general methods that scale with data and compute tend to beat methods built on hand-engineered knowledge [25]. For formulations, the binding constraint is not the model. It is the data. There is no formulation equivalent of the Protein Data Bank, so the data has to be made.

2.3Language models as scientific collaborators

The third thread is the newest. Coscientist, an agent built on a large language model with access to search, code and laboratory automation, planned and executed chemical experiments, including optimizing palladium-catalysed cross-coupling reactions [15]. ChemCrow gave a language model eighteen expert chemistry tools and used it to plan and run syntheses of an insect repellent and three organocatalysts [26]. Surveys from the community have documented many other uses, from extracting data out of papers to acting as a natural-language interface to simulation tools [27].

Agents are also getting better at the part of science that happens before the experiment. PaperQA2 answered questions over the scientific literature with higher precision than PhD-level experts given the same task [28], and the Aviary framework showed that small open models can be trained to perform demanding scientific tasks well [29]. Multi-agent systems now generate and debate hypotheses: Google's AI co-scientist proposed drug-repurposing candidates and targets that were later supported by wet-lab experiments [30], and FutureHouse's Robin chained literature and data-analysis agents to propose a candidate therapeutic for dry age-related macular degeneration [31]. The AI Scientist went as far as writing and reviewing complete machine learning papers [32].

The most useful design pattern here is arguably the one behind FunSearch, which paired a language model that proposes programs with an automated evaluator that scores them, inside an evolutionary loop, and found new results in combinatorics [33]. Language models are powerful generators of ideas. They are unreliable judges of them. The systems that work pair them with a verifier, and in the physical sciences the most trustworthy verifier is an experiment.

2.4Deciding what to try next

Between a model and a robot sits a decision: which experiments to run next. Bayesian optimization is the standard answer when experiments are expensive. It fits a probabilistic surrogate model to the data so far and chooses the next experiment by trading off exploitation of promising regions against exploration of uncertain ones [34]. Practical recipes from machine learning [35] carried over to chemistry, where a head-to-head study found that Bayesian optimization matched or beat expert chemists at reaction optimization, in both efficiency and consistency [36].

Formulation stretches the standard toolkit. Variables are mixed (categorical ingredient choices and continuous concentrations), constrained (sums to one hundred percent, regulatory limits on specific ingredients), multi-objective and expensive at different fidelities, since a viscosity reading takes minutes and a stability study takes weeks. Optimizers developed for chemistry handle batched proposals and categorical variables [37], open frameworks make Monte Carlo acquisition functions practical on modern hardware [38], and recent self-driving-lab software supports multi-objective, constrained and multi-fidelity planning [39].

2.5Machine learning for formulations and personal care

Work applying these ideas directly to formulations is thinner but growing. Cao and colleagues coupled a machine learning classifier with multi-objective Bayesian optimization on a robotic platform to design liquid formulations, and found several recipes that met customer-defined targets within a few weeks of operation [13]. The same research community later released a dataset of 812 shampoo-type formulations built from a fixed set of ingredients, labelled for phase stability, turbidity and rheology and generated on a semi-automated platform [16]. It is one of very few public formulation datasets of its kind.

Other work targets the properties that govern formulation behaviour. Graph convolutional networks predict critical micelle concentrations across surfactant classes more accurately than earlier descriptor-based methods [40], and newer models extend this to binary surfactant mixtures [41], a step closer to real products. Models trained on hundreds of cleansing-foam formulations can predict cleansing performance from ingredient descriptors [42]. A 2024 review of AI in cosmetic and personal care formulation summarises the state of the field and its gaps [43].

Compared with molecules and crystals, formulation machine learning is data-starved. Public datasets are small, industrial datasets are proprietary and heterogeneous, and closed-loop demonstrations are few. We see this as the gap where a dedicated autonomous lab can contribute the most.

3Our approach

Kodimo is building an autonomous formulation lab around a single closed loop. A product brief goes in; candidate formulations are generated, made and measured; and every result makes the next round better. Figure 1 shows the loop.

3.1From brief to objective

Every campaign starts with a brief written the way a product team would write it, for example: a lightweight daily moisturizer, pH between 5.0 and 5.5, silicone-free, stable for twelve weeks at 40 °C, under three dollars per kilogram. A language model translates the brief into a structured specification: objectives to optimize, hard constraints such as ingredient limits for the target markets, softer preferences, and the laboratory measurements that act as proxies for each. A chemist reviews and signs off the specification before anything runs. Getting the objective right is half of the science.

3.2Generate

Candidate generation combines three components. Language-model agents grounded in the literature and in ingredient documentation propose chemically sensible starting points, in the spirit of literature-grounded agents such as PaperQA2 [28]. Mixture-aware property models, trained on our own data, predict how each candidate will behave. A constraint layer guarantees that every candidate is valid: concentrations sum correctly, regulatory limits hold, incompatible ingredients are excluded. As in FunSearch, generation is cheap and verification decides [33]. We predict before we pipette, so the robots spend their time on the candidates most worth testing.

3.3Test

The wet lab dispenses liquids and powders, mixes and emulsifies under controlled temperature and shear, and measures the properties that matter for a personal care product: pH, viscosity and rheology, droplet size, appearance and phase stability, and stability under accelerated ageing. We draw on the design lessons of earlier self-driving labs, especially the value of modular stations and standard instruments [10, 11].

Every experiment is recorded with complete provenance: ingredient lots, process parameters, instrument settings, raw signals and derived values, stored following the FAIR principles so that data is findable, accessible, interoperable and reusable [44]. The data is not a by-product of the lab. Over time it is the most valuable thing the lab produces.

3.4Learn

Results update the surrogate models, and a Bayesian optimization layer chooses the next batch [37, 38, 39]. It handles mixed variables and constraints, balances several objectives at once, and uses multiple fidelities: fast, early measurements screen candidates before the slow, expensive stability tests are spent on the most promising ones. Because every campaign writes to the same data layer, what the lab learns on one brief transfers to the next, and the models improve with every product we work on.

3.5Scientists in the loop

Autonomy does not mean the absence of people. Our chemists write and review briefs, audit surprising results and own every decision about safety. The machines handle volume, precision and bookkeeping; people provide judgment. We begin from ingredients with established safety profiles, and anything new goes through standard toxicological and regulatory review before it comes near a consumer.

4What could go wrong

We think a research program should be judged partly by how clearly it states its own failure modes. Here are the ones we worry about most.

4.1Throughput is not discovery

The A-Lab reported the autonomous synthesis of dozens of new inorganic compounds over seventeen days of continuous operation [14]. A subsequent analysis by independent materials chemists re-examined the claimed products and argued that problems with automated phase identification meant that none of them was convincingly shown to be a new material [45]; the dispute was widely reported [46]. The lesson is not that autonomous labs fail. It is that speed amplifies whatever errors are in the evaluation pipeline. Our defence is to verify top candidates independently, with expert review and repeat measurements, to fix success criteria for each brief before a campaign starts, and to always compare against formulations designed by experienced human chemists.

4.2Language models make things up

Language-model agents can hallucinate facts, behave inconsistently from run to run and judge their own output poorly. Even flagship demonstrations operate with guardrails and human checks [15, 26], and in the most impressive hypothesis-generation results, people carried out and validated the physical experiments [30, 31]. In our loop a language model never has the final word. Every proposal must pass the constraint layer and then an experiment.

4.3The data desert

Formulation data is scarce, proprietary and measured inconsistently from lab to lab [43]. Models trained on such data can look good and generalize poorly. We are addressing this by generating our own data at scale under uniform protocols, measuring uncertainty alongside every value, and contributing benchmarks back to the community where we can, following the example of open datasets such as [16].

4.4Proxies are not people

Viscosity, droplet size and stability are proxies for what a consumer actually experiences: how a cream spreads, how quickly it absorbs, how it smells. We will calibrate our proxies against trained sensory panels and consumer testing, and human evaluation stays as the final gate for any product.

4.5Bench is not factory

A formulation that behaves well at ten millilitres can behave differently at a thousand kilograms. We include process parameters in our models from the first experiment and plan pilot-scale validation with manufacturing partners.

5Outlook

Hiroaki Kitano has proposed building AI systems capable of making discoveries worthy of a Nobel Prize [47]. Our near-term goal is more concrete: to make formulation science systematic, fast and cumulative, so that every experiment adds to what we know instead of disappearing into a notebook. Everyday chemistry is a good place to start. The products are used by billions of people, the science is genuinely hard, and better answers make a direct difference to safety, sustainability and the quality of daily life.

We will share results here as the lab comes online. If this is the kind of problem you want to work on, we are hiring.

References

1. 1Baker, M. 1,500 scientists lift the lid on reproducibility. Nature 533, 452–454 (2016). doi.org/10.1038/533452a
2. 2Ioannidis, J. P. A. Why most published research findings are false. PLOS Medicine 2, e124 (2005). doi.org/10.1371/journal.pmed.0020124
3. 3Wang, H. et al. Scientific discovery in the age of artificial intelligence. Nature 620, 47–60 (2023). doi.org/10.1038/s41586-023-06221-2
4. 4Tom, G. et al. Self-driving laboratories for chemistry and materials science. Chemical Reviews 124, 9633–9732 (2024). doi.org/10.1021/acs.chemrev.4c00055
5. 5King, R. D. et al. The automation of science. Science 324, 85–89 (2009). doi.org/10.1126/science.1165620
6. 6Häse, F., Roch, L. M. & Aspuru-Guzik, A. Next-generation experimentation with self-driving laboratories. Trends in Chemistry 1, 282–291 (2019). doi.org/10.1016/j.trechm.2019.02.007
7. 7Coley, C. W., Eyke, N. S. & Jensen, K. F. Autonomous discovery in the chemical sciences part I: Progress. Angewandte Chemie International Edition 59, 22858–22893 (2020). doi.org/10.1002/anie.201909987
8. 8Coley, C. W., Eyke, N. S. & Jensen, K. F. Autonomous discovery in the chemical sciences part II: Outlook. Angewandte Chemie International Edition 59, 23414–23436 (2020). doi.org/10.1002/anie.201909989
9. 9MacLeod, B. P. et al. Self-driving laboratory for accelerated discovery of thin-film materials. Science Advances 6, eaaz8867 (2020). doi.org/10.1126/sciadv.aaz8867
10. 10Burger, B. et al. A mobile robotic chemist. Nature 583, 237–241 (2020). doi.org/10.1038/s41586-020-2442-2
11. 11Dai, T. et al. Autonomous mobile robots for exploratory synthetic chemistry. Nature 635, 890–897 (2024). doi.org/10.1038/s41586-024-08173-7
12. 12Abolhasani, M. & Kumacheva, E. The rise of self-driving labs in chemical and materials sciences. Nature Synthesis 2, 483–492 (2023). doi.org/10.1038/s44160-022-00231-0
13. 13Cao, L. et al. Optimization of formulations using robotic experiments driven by machine learning DoE. Cell Reports Physical Science 2, 100295 (2021). doi.org/10.1016/j.xcrp.2020.100295
14. 14Szymanski, N. J. et al. An autonomous laboratory for the accelerated synthesis of inorganic materials. Nature 624, 86–91 (2023). doi.org/10.1038/s41586-023-06734-w
15. 15Boiko, D. A., MacKnight, R., Kline, B. & Gomes, G. Autonomous chemical research with large language models. Nature 624, 570–578 (2023). doi.org/10.1038/s41586-023-06792-0
16. 16Chitre, A. et al. Accelerating formulation design via machine learning: generating a high-throughput shampoo formulations dataset. Scientific Data 11, 728 (2024). doi.org/10.1038/s41597-024-03573-w
17. 17Yang, K. et al. Analyzing learned molecular representations for property prediction. Journal of Chemical Information and Modeling 59, 3370–3388 (2019). doi.org/10.1021/acs.jcim.9b00237
18. 18Schwaller, P. et al. Molecular Transformer: a model for uncertainty-calibrated chemical reaction prediction. ACS Central Science 5, 1572–1583 (2019). doi.org/10.1021/acscentsci.9b00576
19. 19Gómez-Bombarelli, R. et al. Automatic chemical design using a data-driven continuous representation of molecules. ACS Central Science 4, 268–276 (2018). doi.org/10.1021/acscentsci.7b00572
20. 20Zeni, C. et al. A generative model for inorganic materials design. Nature 639, 624–632 (2025). doi.org/10.1038/s41586-025-08628-5
21. 21Merchant, A. et al. Scaling deep learning for materials discovery. Nature 624, 80–85 (2023). doi.org/10.1038/s41586-023-06735-9
22. 22Stokes, J. M. et al. A deep learning approach to antibiotic discovery. Cell 180, 688–702.e13 (2020). doi.org/10.1016/j.cell.2020.01.021
23. 23Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021). doi.org/10.1038/s41586-021-03819-2
24. 24Jablonka, K. M., Schwaller, P., Ortega-Guerrero, A. & Smit, B. Leveraging large language models for predictive chemistry. Nature Machine Intelligence 6, 161–169 (2024). doi.org/10.1038/s42256-023-00788-1
25. 25Sutton, R. The bitter lesson. Essay (not peer reviewed) (2019). www.incompleteideas.net/IncIdeas/BitterLesson.html
26. 26Bran, A. M. et al. Augmenting large language models with chemistry tools. Nature Machine Intelligence 6, 525–535 (2024). doi.org/10.1038/s42256-024-00832-8
27. 27Jablonka, K. M. et al. 14 examples of how LLMs can transform materials science and chemistry: a reflection on a large language model hackathon. Digital Discovery 2, 1233–1250 (2023). doi.org/10.1039/d3dd00113j
28. 28Skarlinski, M. D. et al. Language agents achieve superhuman synthesis of scientific knowledge. arXiv:2409.13740 (2024). arxiv.org/abs/2409.13740
29. 29Narayanan, S. et al. Aviary: training language agents on challenging scientific tasks. arXiv:2412.21154 (2024). arxiv.org/abs/2412.21154
30. 30Gottweis, J. et al. Towards an AI co-scientist. arXiv:2502.18864 (2025). arxiv.org/abs/2502.18864
31. 31Ghareeb, A. E. et al. Robin: a multi-agent system for automating scientific discovery. arXiv:2505.13400 (2025). arxiv.org/abs/2505.13400
32. 32Lu, C. et al. The AI Scientist: towards fully automated open-ended scientific discovery. arXiv:2408.06292 (2024). arxiv.org/abs/2408.06292
33. 33Romera-Paredes, B. et al. Mathematical discoveries from program search with large language models. Nature 625, 468–475 (2024). doi.org/10.1038/s41586-023-06924-6
34. 34Frazier, P. I. A tutorial on Bayesian optimization. arXiv:1807.02811 (2018). arxiv.org/abs/1807.02811
35. 35Snoek, J., Larochelle, H. & Adams, R. P. Practical Bayesian optimization of machine learning algorithms. Advances in Neural Information Processing Systems (NeurIPS) 25 (2012). proceedings.neurips.cc/paper/2012/hash/05311655a15b75fab86956663e1819cd-Abstract.html
36. 36Shields, B. J. et al. Bayesian reaction optimization as a tool for chemical synthesis. Nature 590, 89–96 (2021). doi.org/10.1038/s41586-021-03213-y
37. 37Häse, F., Roch, L. M., Kreisbeck, C. & Aspuru-Guzik, A. Phoenics: a Bayesian optimizer for chemistry. ACS Central Science 4, 1134–1145 (2018). doi.org/10.1021/acscentsci.8b00307
38. 38Balandat, M. et al. BoTorch: a framework for efficient Monte-Carlo Bayesian optimization. Advances in Neural Information Processing Systems (NeurIPS) 33 (2020). proceedings.neurips.cc/paper/2020/hash/f5b1b89d98b7286673128a5fb112cb9a-Abstract.html
39. 39Hickman, R. J. et al. Atlas: a brain for self-driving laboratories. Digital Discovery 4, 1006–1029 (2025). doi.org/10.1039/D4DD00115J
40. 40Qin, S., Jin, T., Van Lehn, R. C. & Zavala, V. M. Predicting critical micelle concentrations for surfactants using graph convolutional neural networks. The Journal of Physical Chemistry B 125, 10610–10620 (2021). doi.org/10.1021/acs.jpcb.1c05264
41. 41Choudhary, A. et al. Predicting the critical micelle concentration of binary surfactant mixtures using machine learning. Journal of Cheminformatics 17, 170 (2025). doi.org/10.1186/s13321-025-01112-2
42. 42Hamaguchi, M., Miwake, H., Nakatake, R. & Arai, N. Predicting the performance of functional materials composed of polymeric multicomponent systems using artificial intelligence: formulations of cleansing foams as an example. Polymers 15, 4216 (2023). doi.org/10.3390/polym15214216
43. 43Xin, H. et al. Applications of artificial intelligence and machine learning on critical materials used in cosmetics and personal care formulation design. Current Opinion in Colloid & Interface Science 73, 101847 (2024). doi.org/10.1016/j.cocis.2024.101847
44. 44Wilkinson, M. D. et al. The FAIR Guiding Principles for scientific data management and stewardship. Scientific Data 3, 160018 (2016). doi.org/10.1038/sdata.2016.18
45. 45Leeman, J. et al. Challenges in high-throughput inorganic materials prediction and autonomous synthesis. PRX Energy 3, 011002 (2024). doi.org/10.1103/PRXEnergy.3.011002
46. 46Peplow, M. (news) Robot chemist sparks row with claim it created new materials. Nature news (2023). www.nature.com/articles/d41586-023-03956-w
47. 47Kitano, H. Nobel Turing Challenge: creating the engine for scientific discovery. npj Systems Biology and Applications 7, 29 (2021). doi.org/10.1038/s41540-021-00189-3