A typical development program runs 20-100 batches. In a modest six-dimensional composition space, even 100 batches cover less than 0.3% of the available candidates. Generative design addresses that coverage gap by helping choose which formulations to test next.
This primer describes how the approach works in cultivated bio-material development, from defining the search space to proposing and ranking new compositions. The examples reflect the Celleste platform for bio-material formulation prediction and generative design.
Before generative design: trial-and-error formulation
Formulation work usually begins with a set of materials, concentrations, and process conditions that seem plausible from prior experiments or domain knowledge. Researchers prepare batches, measure properties such as strength, opacity, and processability, then adjust one or more variables for the next round.
That sequence provides useful evidence, but it samples only a small portion of the possible space. A team may learn a great deal about the neighborhood around its starting formulation while remaining unaware of better regions elsewhere. Each batch also consumes material, operator time, and measurement capacity.
The practical question is therefore not only what a tested formulation does. It is which untested formulation is most worth making next, given the evidence already collected and the properties the material must achieve.
Why the composition space is large
Consider a simple case with 3 compositional variables and 3 process variables. Together, they create a space of approximately 6 dimensions. If each variable is evaluated at five levels and three variants are considered, the grid contains 5^6 x 3 = ~47,000 candidate formulations.
A program that tests 20-100 batches visits only a small fraction of those options. Even 100 batches represent less than 0.3% of the 6D composition space. This is not merely a problem of having too many combinations. Interactions between variables can make a useful formulation occupy a narrow region that a sequential search never reaches.
Exhaustive testing is consequently unrealistic for many formulation programs. The search has to be selective, with each experiment chosen for its expected contribution to the target rather than simply added to a growing catalogue.
How a composition gets proposed
The process starts with training data assembled from published bio-materials literature and calibration data from validated batches. Measurements are linked to formulation and process variables, allowing a model to estimate how candidate compositions may perform across the defined property set.
Bayesian optimization is then used to search. It balances candidates that appear promising against candidates that could reduce uncertainty in less explored areas. The objective can include several target properties and practical constraints, rather than treating one measured value as the sole definition of success.
At each cycle, the system proposes a short list for laboratory evaluation. The intended output is 3-5 candidates, not 50. Results from those batches are returned to the model, which updates the search and proposes the next set.
What makes it generative rather than just predictive
Predictive modeling normally moves from a known composition to an estimated property. Generative design reverses that direction: the desired property target is supplied as the input, and the system returns compositions that may meet it.
The distinction matters because the output does not have to be a formulation already present in the training set. The search can combine variable values into novel compositions, subject to the allowed formulation and process constraints, and estimate how those combinations may behave.
Uncertainty is part of the proposal rather than hidden from the user. Candidates in poorly supported regions can be flagged explicitly or assigned lower priority, so a promising predicted result is considered alongside the strength of the evidence behind it.
Domain knowledge in generative design
Generative design does not replace material knowledge. Information about feasible concentrations, incompatible ingredients, manufacturing limits, and known failure modes helps define the search space and prevents the system from spending experiments on implausible options.
That knowledge can also be embedded in the training data and model objectives. For example, a formulation containing collagen at 20% by mass may be predicted as brittle, opaque, and difficult to process. Such a result reflects prior knowledge represented in the model, not only a new laboratory observation.
The limitation is equally important: the advantage is most apparent when the target application has good coverage in the training data. For genuinely novel applications, the first iterations remain exploratory regardless of the search method, because the model has little relevant evidence on which to orient its proposals.
Development with generative design in the loop
A program still begins with formulation goals, measurable property targets, and constraints on ingredients and processing. Existing literature and validated batches provide the initial calibration data, while early experiments establish how the selected material system behaves under the intended conditions.
Once enough calibration data exists to orient the search, the Celleste platform can rank candidate formulations against the target profile and identify a small batch of experiments. The team reviews those proposals, produces the selected batches, and feeds the resulting measurements back into the next optimization cycle.
This loop changes the role of experiments from broad, largely manual exploration to targeted data collection. It does not remove laboratory work. It makes each new batch part of an explicit search, with its value judged by both predicted performance and the information it adds.
