Generative models for formulation: learning from twenty years of trials

Ask a senior formulator how they approach a new brief and you will rarely hear anything that sounds like search. They will tell you about a base they trust, a raw material that behaves badly above a certain level, a failure mode they have seen before, or a trick a colleague taught them fifteen years ago. It is deep expertise, and it lives mostly in people.

The company has already paid for a much larger version of that knowledge. Every trial that was weighed, measured, smelled, rejected or reformulated over the last two decades is a data point about how formulation space behaves. Most of that history is recorded as recipes, spreadsheets, reports and notes written for humans, not models.

That is why formulation AI is still difficult in practice.

Suppose the history exists in structured form. What can you actually do with it? Quite a lot, and the useful answer starts with the workflow formulators already know: design of experiments.

From DoE to cumulative learning

DoE is already the right mental model. Define the variables, choose experiments that cover the space intelligently, run them, measure the outcomes and learn what matters.

The problem is that each campaign often behaves too much like a fresh start.

A team may spend weeks exploring ingredient levels and process conditions for one brief, then begin the next project with only the formulator’s memory carrying that knowledge forward. Once historical trials are structured, that changes.

The first campaign may still need broad experimental coverage. But after the system has seen hundreds or thousands of trials, the next design does not have to pretend that nothing is known. Predictive models can estimate which regions are promising, which combinations repeatedly fail, and where uncertainty remains high. Sequential design methods, including Bayesian optimization, can then help choose the next experiments based on what has already been learned.

This is where Bayesian methods are useful: bench experiments are expensive, and the next experiment should be informed by the previous ones. Reaction-optimization work has shown that Bayesian optimization can reduce experimental burden, and formulation-specific work has applied related ideas to ingredient selection and concentration optimization.

The goal is simple: every DoE should make the next DoE smaller, smarter, or both.

The hard data is in the notes

The numerical part is often the easy part.

Ingredient levels, process temperatures, viscosity, pH, stability measurements and cost fit naturally into ordinary supervised-learning pipelines. In many cases, tree-based models and boosting methods are strong baselines before anything more exotic is justified.

The hard part is everything a formulator writes next to the numbers.

Those comments often contain the reason an experiment failed, and therefore the information needed to avoid repeating it.

A binary success/failure column collapses all of those outcomes into the same label. A useful formulation data layer has to preserve the original note and connect it to structured concepts such as sensory defects, processing failures, instability, texture problems or regulatory constraints. The vocabulary cannot be imposed too aggressively: formulation language is specific to the product, process and company.

This is a place where language models can help without becoming the scientific model themselves. They can normalize inconsistent notes, propose structured tags, link synonyms and surface recurring failure modes while keeping the raw observation attached.

The model should learn from the formulator’s knowledge, not erase it during preprocessing.

Ranking before the bench

Once that history exists, the first useful system does not need to generate magical new recipes. It needs to rank good hypotheses.

Given a new brief, retrieve relevant historical formulas. Predict likely properties and failure modes. Identify ingredient or process regions associated with previous problems. Then propose a smaller set of experiments balancing expected performance with uncertainty and diversity.

Bench time remains the scarce resource. The model’s job is to make a week of experiments start from what the company already knows instead of from a blank page.

Every edit, rejection and explanation adds information. A formula rejected because it is too expensive is different from one rejected because the emulsion breaks. Capturing those decisions means each development cycle improves the dataset used by the next one.

Where generative models enter

Generative models become interesting once the historical layer is rich enough to represent what plausible formulas look like and how formulations relate to measured properties.

There is already evidence that this is possible. Variational autoencoders trained on large curated formulation datasets have learned useful representations and generated ingredient-amount suggestions for chemical product deformulation.

The more interesting long-term problem is reformulation: start from something that already works, change a constraint, and preserve the properties that matter. Replace a restricted ingredient. Reduce cost. Move toward a cleaner label. Adapt a formula to a new market while disturbing the rest as little as possible.

That is still a research problem, especially when public data is small and performance depends on process history and poorly measured sensory properties. But it becomes tractable only after the underlying formulation history has been structured.

Our bet is that generative formulation will not arrive because one model architecture suddenly becomes good enough. It will arrive because companies finally start accumulating reusable experimental knowledge.

A good formulation lab already learns from every experiment.

The opportunity is to make the software learn too.

References