A large bio-material lab may hold 200-500 validated formulations, while inorganic materials databases contain millions. That gap explains why inverse design literature, built largely around crystal structure prediction, does not transfer directly to bio-materials. Inverse design takes property targets as input and composition as output, the reverse of forward modeling. At these data scales, Bayesian methods are usually more practical than deep generative models.
Inorganic materials provide a well-developed setting for inverse design, with established methods for crystal structure prediction and functional materials discovery. Bio-material formulation brings continuous variables, process dependence, smaller datasets, and greater uncertainty. This article examines the representations and methods that make inverse design usable at those scales.
Why bio-materials require different representations
In inorganic materials science, composition is often represented as a vector of elemental fractions with relatively few meaningful components. Crystal periodicity also makes the search space discrete in many cases: the task is to select elements and stoichiometries from a bounded set. Property models may draw on databases containing millions of computed and experimental data points.
Bio-material formulations differ on several fronts. Concentrations such as crosslinker ratio vary continuously rather than coming from a discrete list. Adding process conditions creates high-dimensional spaces: a collagen formulation with three composition variables and four process variables occupies a seven-dimensional space. A large bio-material lab may have only 200-500 validated formulations, not millions. The mapping from composition to property is also context-dependent. The same formulation can behave differently with another cell type or under different bioreactor conditions.
That is why variational autoencoders and generative adversarial networks, which can work well for inorganic materials design, do not carry over directly. Bio-material datasets generally lack the volume needed to train them. At this scale, practical choices include Gaussian processes, Bayesian neural networks, and Bayesian optimization algorithms.
Forward model first in inverse design
Practical bio-material inverse design begins with a forward model, meaning a predictor that maps composition to properties. The inverse step searches the composition space for inputs that the forward model predicts will satisfy the requested property targets.
There are several ways to conduct that search. Gradient-based optimization is suitable when the forward model is differentiable with respect to composition. Bayesian optimization is better when evaluation is costly, such as when each evaluation requires a batch, or when the space combines continuous and discrete variables. Monte Carlo sampling with rejection criteria can work with well-calibrated probabilistic models: sample compositions from a proposal distribution, then retain those whose predicted properties fall within the target range.
Celleste uses Bayesian optimization for its inverse design step. Its acquisition function balances exploitation, finding compositions predicted to have good performance, with exploration, finding uncertain compositions where new data could be informative. A user can adjust that balance for an exploration phase, focused on understanding the formulation space, or an optimization phase, focused on finding a strong formulation efficiently.
Handling multi-objective targets
Most bio-material applications impose several property targets at once. A cartilage engineering scaffold, for example, might need a Young's modulus above 0.5 MPa, cell adhesion above a defined threshold, and a degradation half-life between 4 and 12 weeks. Because these requirements can compete, the inverse design task becomes multi-objective optimization.
The formal answer is the Pareto front: compositions for which improving one property would worsen another. Computing the complete front is often expensive, and it may offer more detail than a laboratory can use. A smaller set of candidates is generally more actionable, provided the candidates represent different constraint strategies: one can favor mechanical properties, another biological activity, and a third can seek a balance.
Celleste structures its recommendations in this form, presenting three to five candidate formulations grouped by the strategy each represents and showing their trade-offs. The lab selects a direction according to its application requirements, rather than simply choosing the formulation with the highest aggregate score. That remains a scientific judgment for the researcher.
Inverse-design uncertainty management
Inverse design magnifies prediction uncertainty in a way that forward modeling does not. A forward prediction returns a property estimate with its confidence interval. When that estimate drives a composition recommendation, uncertainty also enters the search through the landscape, so uncertainty in the recommended composition can exceed uncertainty in any one property prediction.
For example, two compositions may have overlapping confidence intervals for a target property. In that situation, the inverse design algorithm cannot reliably tell them apart. It can recommend one, but the confidence that it is better than the other is limited by that overlap. More dependable recommendations require tighter intervals, obtained through additional lab-specific calibration data or a simpler formulation space with better model coverage.
We make this distinction visible in the recommendation output. If confidence that the leading composition is better than its alternatives falls below a threshold, we flag it and suggest testing both the top recommendation and its closest alternative. The single highest-ranked candidate is not treated as definitive.
The active inverse-design workflow in practice
For bio-material development, inverse design is most useful as an active iteration loop rather than a one-shot prediction. The lab sets initial property targets. The platform proposes a small candidate set. The lab runs one or two batches, measures the relevant properties, and records the results. The platform updates its model and produces a revised set. This repeats until the targets are reached or the model indicates that they are not achievable within the formulation space.
The experimental portion of an iteration typically takes one to two weeks, depending on the measurement protocol; the model update is instantaneous. In our experience with early-access labs, well-defined targets in formulation spaces with existing calibration reach a satisfactory formulation within four to eight iterations. Novel formulation systems and poorly specified targets generally require more.
The inverse design process is not a replacement for the researcher's judgment about which target to set, which measurements to run, or whether a recommendation is feasible under actual lab constraints. It compresses the interval between defining a formulation target and selecting a candidate to test, replacing a manual search through literature and intuition with a model-guided search of composition space. That benefit depends on model quality, target quality, and the experimental conditions used to generate the next data.
