Mixture Design vs Factorial DOE for Formulators (2026)
A formulator sets up a textbook two-level factorial on a five-component coating, runs 32 batches over three weeks, fits the model, and finds that the main effects make no physical sense. Raising the binder "improves" gloss in one block and destroys it in the next. The culprit is not the lab work. It is the design of experiments itself: the components of a formulation must sum to 100%, so increasing one ingredient forces another to fall, and a factorial model that assumes independent factors is quietly fitting a geometry that does not exist. Choosing between mixture design and factorial DOE is the first decision in any formulation study, and it is the one most often made by habit rather than by the structure of the problem.
This article lays out when each design of experiments applies, how many runs each actually costs, what happens when a study needs both, and how AI-driven DOE changes the economics of the choice.
Why Design of Experiments Breaks When Components Must Sum to 100%
In a classical factorial, each factor is varied independently. Temperature can be 60 °C or 80 °C regardless of whether mixing speed is 200 rpm or 400 rpm. The experimental region is a cube, every corner is reachable, and the regression model carries an intercept plus main effects and interactions.
A formulation does not live in a cube. If a cleaning concentrate contains surfactant, solvent, chelant, and water, the four fractions are constrained by x₁ + x₂ + x₃ + x₄ = 1. The experimental region collapses from a four-dimensional cube to a three-dimensional simplex, a tetrahedron. You cannot "hold everything else constant" while raising the surfactant, because something must give. The response you measure depends on proportions, not on absolute amounts, and the response surface has to be described in terms of blends rather than factor levels.
Two practical consequences follow. First, the intercept of a standard polynomial becomes redundant, which is why mixture studies use Scheffé canonical polynomials with no constant term. Second, the pairwise interaction terms in a mixture model mean something different: a positive x₁x₂ coefficient indicates synergistic blending, not an interaction between two independent dials. A team that fits a factorial model to mixture data gets inflated variance, aliased terms, and the kind of sign-flipping coefficients described in the opening scenario.
Factorial DOE: Where It Still Earns Its Place in Chemical Formulation
None of this makes factorial DOE obsolete for formulators. It is the right tool whenever the variables really are independent, which in a chemical lab usually means process variables. Cure temperature, mixing time, shear rate, addition order, pH adjustment, and catalyst ppm (when the catalyst is a trace additive that does not materially shift the mass balance) all behave as independent factors.
The economics are well understood. A two-level full factorial with three process factors is 8 runs and resolves every interaction. Five factors at two levels is 32 runs; a half-fraction (2⁵⁻¹) brings that to 16 while still separating main effects from two-factor interactions. Push to three levels on five factors to capture curvature and the full factorial explodes to 243 runs, which is why response-surface variants such as central composite and Box-Behnken designs exist: they add axial and center points to a two-level core to estimate quadratic terms with 25 to 45 runs instead of hundreds.
For a process engineer validating a scale-up window, that is exactly the right machinery. The failure mode is applying it to the recipe itself.
Mixture Design: Simplex-Lattice, Simplex-Centroid, and Constrained Designs
Mixture designs place runs on the simplex rather than the cube. The three standard families differ in how they cover that space.
A simplex-lattice {q, m} design uses q components and m+1 equally spaced levels per component. For three components fit to a quadratic Scheffé model, that is 6 blends: the three pure components and the three 50/50 binaries. Move to five components and a quadratic lattice is 15 blends, which maps neatly onto the 15 coefficients of a five-component quadratic polynomial. A simplex-centroid design adds every equal-parts blend of every subset, giving 2^q − 1 points: 7 for three components, 15 for four, 31 for five.
Real formulations almost never permit pure components. Nobody will test a 100% surfactant "coating" or a 0% water concentrate. Once each ingredient carries upper and lower bounds, the feasible region becomes an irregular polytope inside the simplex, and the lattice points fall outside the allowed space. This is where extreme-vertices designs and D-optimal selection take over: the software enumerates the vertices, edge centroids, and overall centroid of the constrained region, then selects the subset that best supports the chosen model. A four-component adhesive with realistic bounds typically needs 12 to 16 runs for a quadratic model with replicates, compared with 15 for an unconstrained centroid or 81 for a naive 3⁴ factorial on the same ingredients.
Mixture Design vs Factorial DOE: Side-by-Side Comparison
| Dimension | Factorial / Response-Surface DOE | Mixture Design |
|---|---|---|
| Variable type | Independent factors (temperature, time, shear, pH) | Component proportions constrained to sum to 100% |
| Experimental region | Hypercube (every corner reachable) | Simplex or constrained polytope |
| Model form | Polynomial with intercept, main effects, interactions, quadratics | Scheffé canonical polynomial, no intercept |
| Meaning of x₁x₂ term | Interaction between two dials | Blending synergy or antagonism |
| Typical runs, 4 variables, quadratic | 25–30 (central composite) | 10–16 (lattice, centroid, or D-optimal) |
| Handles ingredient bounds | Not applicable | Extreme-vertices / D-optimal designs |
| Best for | Process windows, scale-up, cure schedules | Recipe optimization, ingredient substitution, cost-down |
| Classic failure mode | Applied to recipe proportions → aliased, nonsensical effects | Ignoring process factors → results don't transfer to plant |
The last row deserves emphasis. A mixture study run at a single mixing temperature may find a beautiful optimum that falls apart when the plant runs ten degrees hotter. That is the signal that a combined design is needed.
Combined Mixture-Process Designs: When Design of Experiments Needs Both
Most industrial formulation problems carry both kinds of variables. An adhesive has four resin and filler components that sum to 100% and two process factors, cure temperature and press time. The classical approach crosses a mixture design with a factorial: a 10-point quadratic mixture design for four components crossed with a 2² factorial on process factors gives 40 runs. That is tractable but expensive if each run is a half-day of lab time, and it balloons quickly; add a third process factor and you are at 80.
The modern alternative is a D-optimal or I-optimal design computed over the combined model. Instead of crossing full designs, the algorithm selects the 18 to 22 runs that best estimate the mixture terms, the process terms, and the mixture-by-process cross terms that matter. The workflow looks like this:
Mixture components vs independent process factors
Upper/lower bounds, ratio limits, cost ceiling
Scheffé quadratic × linear process + cross terms
18–22 runs instead of 40–80 crossed
Add runs only where the model is uncertain
Step 5 is where the practice has shifted most in the past few years. Rather than committing to all runs up front, teams run a first block, fit the model, and let the design be augmented sequentially where prediction variance is highest. For a broader view of how AI-generated designs compare across industries, see the earlier guide to AI-powered DOE across global industries.
How AI Changes Design of Experiments for Formulators
Classical optimal design answers "which runs best estimate a model I have already chosen?" AI-driven DOE answers a harder question: "which runs most quickly get me to a formulation that meets all my specs?" The distinction matters when the response surface is not quadratic, when there are six or more competing specifications, or when a team already has 180 historical batch records that a fresh design would otherwise ignore.
Consider a coatings team with 180 legacy batches of a VOC-compliant topcoat spread across spreadsheets and an ELN. A gradient boosting or Gaussian process model trained on those records already knows which regions of the simplex produce poor adhesion. Rather than spending the first 15 runs of a simplex-centroid rediscovering that, the model proposes the next batch where expected improvement is highest, subject to the mixture constraint, a cost ceiling, and a regulatory filter that excludes any ingredient flagged under REACH or TSCA. In practice this is the Bayesian loop described in the primer on Bayesian optimization: a surrogate model, an acquisition function, and a human reviewing each suggestion before it goes to the bench.
The run-count difference is substantial. A 48-run crossed mixture-process design that would have consumed six weeks of bench time can routinely be replaced by a 6-run seed design plus 6 to 10 sequential suggestions, with the model's uncertainty estimates telling the formulator when to stop. ChemCopilot implements this as AI-driven DOE on top of its predictive formulation models, taking the mixture constraints, process factors, and compliance rules as inputs and returning the next suggested batch along with the predicted property window and the SHAP-style explanation of why that blend was chosen, so a formulator can challenge the suggestion rather than accept it blindly.
What AI does not do is remove the need to classify variables correctly. A surrogate model fed proportions as if they were independent factors will learn the same aliased relationships a factorial would. The mixture constraint has to be encoded in the design space before any optimizer runs.
A Decision Framework for Choosing Your Design of Experiments
| Situation | Recommended approach | Indicative runs |
|---|---|---|
| Fixed recipe, tuning cure or mixing conditions | Fractional factorial → central composite | 8–30 |
| Reformulating 3–5 components, no process change | Constrained mixture (extreme vertices / D-optimal) | 10–16 |
| New product, components and process both open | Optimal combined mixture-process design | 18–22 |
| Rich historical data, many specs, cost + compliance limits | AI-driven sequential DOE on constrained space | 6 seed + 6–10 sequential |
| Ingredient substitution (one component swapped) | Mixture design on affected sub-simplex, process held | 6–10 |
The honest summary is that the question in the title is slightly wrong. Mixture design and factorial DOE are not competitors; they describe two different kinds of variable, and the real decision is how to combine them efficiently once both are present. Teams that get the classification right at the start spend their bench time learning about the product. Teams that do not spend it learning about regression artifacts.
FAQ
Can I just use a factorial DOE on formulation components if I let one component "float" to absorb the difference?
This slack-variable approach is common and sometimes workable when the floating component is inert and abundant, such as water in a dilute system. It fails when the slack component itself affects the response, because its effect becomes confounded with everything else. A proper mixture model avoids the problem rather than hiding it.
How many runs does a mixture design need for five components?
A quadratic simplex-lattice needs 15 blends, a simplex-centroid 31. With realistic ingredient bounds and D-optimal selection, 15 to 20 runs including replicates and a center point is typical. Adding a special cubic term for strong three-way synergies raises this by roughly 10 runs.
Does AI-driven DOE replace mixture design, or build on it?
It builds on it. The mixture constraint and ingredient bounds define the feasible space; the AI model decides which points inside that space to test next. Skipping the constraint definition produces suggestions that cannot be made in the lab.
Do I need coding skills to run a combined mixture-process or AI-driven design?
Classical designs are available in most statistical packages through menu-driven interfaces. No-code AI formulation platforms now expose the same workflow, including constraint entry, sequential suggestion, and model explanation, without scripting, which is what allows formulators rather than data scientists to own the loop.
What is the biggest mistake teams make when moving from mixture DOE to AI optimization?
Treating historical batch records as if they were designed experiments. Legacy data is clustered around past recipes and leaves large regions of the simplex unobserved; a model trained on it is confident where it has data and blind elsewhere. The remedy is to use the model to direct a small number of exploratory runs into those gaps before trusting its optimum.
If your team is deciding how to structure its next formulation study, or wants to see how an existing batch history can seed an AI-driven design, talk to the ChemCopilot team.
