Table of Contents
Design of experiments (DoE) fixes a set of runs in advance and fits a model that explains how each factor affects the result. Bayesian optimisation picks a few runs at a time, guided by a model of every result so far, to reach a target in fewer tests. DoE explains a process; Bayesian optimisation searches it. Many teams need both.
Key takeaways
- DoE is a planned set of runs, and good practice is a sequence of small experiments, each informed by the last (NIST/SEMATECH).
- Bayesian optimisation suits tests that take minutes or hours, tolerates noisy results, and works best with fewer than 20 continuous variables (Frazier, 2018).
- Head-to-head evidence is mixed. In a pilot-scale study, Bayesian optimisation found comparable conditions but did not reduce the number of experiments; in cell culture media, authors estimate it used 3 to 30 times fewer experiments than standard DoE.
- A DoE model is valid across the whole region studied. A Bayesian optimisation model is usually most accurate near the optimum.
- A practical pattern: screen and seed with a small classical design, then let the optimiser choose the next batches.
What is design of experiments?
The NIST/SEMATECH e-Handbook defines DoE as an efficient procedure for planning experiments so that the data can support valid and objective conclusions. You choose the factors and their ranges, select a design, run every point, then fit a statistical model, usually a linear or quadratic regression.
The design depends on the question. NIST's design selection guideline maps objectives to designs: comparative designs to test whether one factor matters, full or fractional factorials and Plackett-Burman designs to screen many factors, and central composite or Box-Behnken designs to model curvature once only 2 to 4 factors remain. With 5 or more factors, its advice for a response surface is to screen first.
Run counts grow quickly. A two-level full factorial with six factors needs 64 runs before centre points, which is why fractional designs exist. Response surface designs cost more per factor:
| Factors | Central composite design | Box-Behnken design |
|---|---|---|
| 3 | 20 runs (6 centre points) | 15 runs |
| 4 | 30 runs (6 centre points) | 27 runs |
| 5 | 33 (fractional) or 52 (full factorial) | 46 runs |
| 6 | 54 (fractional) or 91 (full factorial) | 54 runs |
Source: NIST/SEMATECH e-Handbook, Table 3.28.
Two points are often forgotten. First, DoE is not meant to be one big experiment. NIST states that an iterative approach is best and, in the end, most economical, with each stage showing how the next should be run. Second, formulations need special designs: in a mixture experiment the factors are proportions that must sum to 100%, which is handled with simplex-lattice and simplex-centroid designs.
What is Bayesian optimisation?
Bayesian optimisation is a loop with two parts. A surrogate model, usually a Gaussian process, predicts the result for any untested recipe together with an uncertainty that is tight in well-explored regions and wider away from them. An acquisition function, such as expected improvement, then scores every candidate by balancing two goals: testing where the model predicts a good result, and testing where it knows little. The best-scoring candidates go to the lab, their results update the model, and the loop repeats.
Frazier's tutorial sets out where the method fits: objectives that take minutes or hours to evaluate, noisy measurements, and continuous domains of less than 20 dimensions. Lab formulation work often matches that profile, although some projects have more variables than that and need narrowing first.
In chemistry, the best-known test is by Shields and colleagues, who benchmarked the method against expert chemists and engineers on a palladium-catalysed reaction. It outperformed human decision-making in both average efficiency and consistency. Their software, EDBO, is open source under the MIT licence.
Industrial constraints are supported by standard tools. BoTorch can restrict candidates with linear equality and inequality constraints, which covers mixture totals and dosage limits, and its multi-objective methods learn the Pareto front, the set of trade-offs where improving one target worsens another.
How do the two approaches compare?
| Classical DoE | Bayesian optimisation | |
|---|---|---|
| Main question | Which factors matter, and how? | Which experiment should we run next to reach the target? |
| How runs are chosen | All runs of a stage fixed in advance | A few at a time, after each round of results |
| Model | Linear or quadratic regression with effect estimates | Gaussian process or another surrogate with uncertainty |
| Where the model is reliable | Across the whole region studied | Mostly near the best results found |
| Objective | Can study many responses without choosing a winner | Needs a clear objective, or explicit trade-offs between several |
| Mixtures | Dedicated mixture designs | Linear equality constraints |
| Uses past, unplanned data | Not in the standard analysis | Yes, as starting points for the model |
| Easiest to explain to an auditor | Yes: fixed plan, standard statistics | Needs a log of each round and why each run was proposed |
| Typical tools | Any statistical package | BoTorch, Ax, EDBO and similar Python libraries |
What do head-to-head studies show?
Published comparisons do not support the claim that Bayesian optimisation always needs fewer experiments.
Wood delignification, pilot scale. Researchers at VTT compared a 15-run, three-factor Box-Behnken design with Bayesian optimisation that started from 5 of those runs and added 10 of its own. Both methods reached comparable pilot-scale results, and Bayesian optimisation did not reduce the number of experiments needed. Its model was more accurate near the optimum, while the DoE model had lower errors across the full design space. Simulations showed that the choice of initial experiments strongly affected convergence and that noise slowed it. The authors concluded that the two methods are not mutually exclusive and advised using domain knowledge to choose the initial design.
Cell culture media. A 2025 study in Nature Communications optimised media blends, cytokine mixtures and carbon sources with Bayesian optimisation, including a categorical factor with 19 options. The authors report improved conditions using 3 to 30 times fewer experiments than estimated for standard DoE. That comparison is an estimate, not a parallel DoE campaign, and the largest saving relates to the case with 9 factors and categorical variables. They also note a trade-off: small batches need more rounds to converge, while large batches raise the total number of experiments.
Choice of surrogate. Across five experimental materials systems, Liang and colleagues found that random forests and Gaussian processes with anisotropic kernels performed comparably, and both did better than the commonly used Gaussian process without automatic relevance determination. The setup of the model matters more than the label "Bayesian".
The pattern is consistent: the gain from Bayesian optimisation grows with the number of factors and the cost per test, and shrinks when the problem is small enough for a compact classical design.
When should you use DoE, and when Bayesian optimisation?
Choose classical DoE when:
- you need to know which factors matter and by how much, for example to set an operating window or a specification range;
- there are 2 to 4 factors and a response surface design fits your budget;
- the result will be reviewed by quality or regulatory staff who expect standard statistics;
- you have several responses and no agreed way to rank them yet.
Choose Bayesian optimisation when:
- each test is slow or costly and you want to reach a target, not map the whole space;
- there are more factors than a response surface design can afford, but fewer than about 20;
- you have useful history from past projects that a fixed design would ignore;
- the objective, or the trade-off between several targets, can be stated clearly. The VTT authors stress that the method requires a clearly defined performance indicator.
How do you combine them in practice?
- Write the objective and constraints. State the target, the acceptable range of every other property, cost limits and banned or capped components.
- Screen if there are many factors. A fractional factorial or Plackett-Burman design removes factors with little effect, as NIST recommends before a response surface.
- Seed the model. Use a small classical design, past records, or both, chosen with process knowledge so the first points cover the region you care about.
- Set the batch size to your lab. Match it to how many samples your technicians and equipment can run per round, keeping room for replicates.
- Run rounds and log everything. For each proposed recipe, record the predicted value, the measured value and why it was proposed.
- Stop on a rule agreed in advance, such as a recipe meeting the full specification or no improvement over two rounds.
- Confirm and explain. Replicate the best recipe. If you need a robust operating window, run a small classical design around it.
What we have seen at KABAS AI
Our Virtual Lab is used for formulation R&D, where this choice comes up in every project. The two deployments we can describe publicly follow the combined pattern above.
At ADFC (Sonatrach Group), the published workflow starts by defining performance targets for each drilling section and simulating from historical reference recipes. It then uses multi-criteria optimisation and "what-if" scenarios to produce a focused laboratory test playbook, compares simulated and actual results, and stores validated formulations for reuse. At ENAD Shymeca, historical formulation data was migrated before any optimisation, and objectives such as cleaning efficacy against cost are balanced together. The case study reports a 40% reduction in physical testing cycles.
The practical lesson is the one the VTT study points to: the result depends less on the algorithm than on the starting data and a clear objective. Past recipes that were never part of a designed experiment are often the best seed, provided they are recorded consistently. A short AI readiness assessment shows whether yours are.
For the wider picture, read our pillar guide on the virtual lab for industrial R&D and the step-by-step AI formulation optimisation guide. For production rather than R&D, see AI process control in chemical manufacturing. Our chemistry solutions page shows how this fits a chemical company's R&D department.
Sources
- NIST/SEMATECH e-Handbook: What is design of experiments?
- NIST/SEMATECH e-Handbook: What are the steps of DOE?
- NIST/SEMATECH e-Handbook: Select and use an appropriate experimental design
- NIST/SEMATECH e-Handbook: Fractional factorial designs
- NIST/SEMATECH e-Handbook: Comparisons of response surface designs
- NIST/SEMATECH e-Handbook: Mixture designs
- Frazier, A Tutorial on Bayesian Optimization, arXiv 1807.02811, 2018
- Ax documentation: Introduction to Bayesian optimization
- Shields et al., Bayesian reaction optimization as a tool for chemical synthesis, Nature 590, 2021
- EDBO: Experimental Design via Bayesian Optimization (GitHub)
- BoTorch documentation: Constraints
- BoTorch documentation: Multi-objective Bayesian optimization
- Rummukainen et al., Traditional or adaptive design of experiments? A pilot-scale comparison on wood delignification, Heliyon 10, 2024
- Narayanan et al., Accelerating cell culture media development using Bayesian optimization-based iterative experimental design, Nature Communications, 2025
- Liang et al., Benchmarking the performance of Bayesian optimization across multiple experimental materials science domains, arXiv 2106.01309, 2021
Frequently asked questions
Is Bayesian optimisation a kind of design of experiments?
In the broad sense, yes: it is a method for deciding which experiments to run. It is often called adaptive or sequential experimental design. The difference from classical DoE is that it chooses a few runs at a time using a model of all results so far, instead of fixing the full set of runs before the first test.
How many experiments does Bayesian optimisation need?
There is no fixed number. It depends on how many factors you vary, how noisy your test methods are and how the first experiments were chosen. In one published pilot-scale study, the optimiser started from 5 designed runs and added 10 of its own; in a cell culture study, a four-component media blend was optimised in 4 rounds of 6 experiments.
Does Bayesian optimisation tell me which factors matter?
Not as directly as DoE. Its model is built to find a good result, so it tends to be most accurate near the optimum and less accurate elsewhere. If you need effect estimates and a model that is valid over the whole region, for example to set operating windows, a classical design is the better tool.
Can Bayesian optimisation handle recipes whose components must add up to 100%?
Yes. Mixture constraints are linear equality constraints, and open-source libraries such as BoTorch accept linear equality and inequality constraints when choosing the next candidates. Classical DoE handles the same situation with dedicated mixture designs such as simplex-lattice and simplex-centroid designs.
Which software do we need?
Classical designs are available in all common statistical packages. For Bayesian optimisation, open-source Python libraries such as BoTorch, Ax and EDBO are widely used. The harder part is usually not the software but a clean record of recipes, conditions and results that the model can learn from.



