Faster Cosmetic Product Development with Explainable AI, when traditional machine learning isn’t enough

Think about how many attempts it takes to bring a new cosmetic product to the shelf. Behind every cream, shampoo or emulsion there is a long sequence of formulation trials: dozens of raw materials to combine, physicochemical properties to hit, regulatory constraints to respect, and a client brief that leaves little room for error. Traditionally, this is a trial-and-error process. A formulator draws on experience to design a prototype, the lab tests it, the results come back, the formula is adjusted — and the cycle starts again. Every iteration costs time, materials and laboratory capacity. And while the cycle runs, the client is waiting.

For cosmetics manufacturers, this is the reality of every project. The experience described in this case study concerns a Spanish company specialized in developing and manufacturing cosmetic and personal care products for retail chains and private labels. It operates internationally, runs an R&D team of more than 20 professionals and generates annual revenues above €180 million. Each client arrives with a different product concept, a different target market and different specifications — which means the company’s laboratories are constantly running formulation trials to meet requirements that are, by definition, unique.

The development process typically moves through four stages: a client briefing that defines target properties, regulatory requirements and cost constraints; the development of prototype formulations based on R&D expertise; laboratory testing of key attributes such as stability, viscosity, pH, spreadability and fragrance retention; and an iterative optimization loop that continues until every technical, regulatory and sensory requirement is met. When several iterations are needed, this last stage alone can absorb weeks.

Machine learning is an obvious candidate to compress this cycle: predict the properties of a new formulation before it enters the lab, and you can discard weak candidates on screen instead of on the bench. But two obstacles stand in the way.

The first is trust. Most predictive models behave as black boxes. Without visibility into how a model reaches its conclusions, it is difficult for formulators to trust its suggestions, for companies to meet regulatory expectations, and for R&D teams to extract insight they can actually act on.

The second is the shape of the data. Formulation knowledge lives in tables — spreadsheets of trials, ingredients and measured properties. Yet the value hides in the relationships between trials: which formulas are similar, which past experiments are relevant to a new brief, how knowledge transfers from one project to the next. Those relationships are never made explicit in tabular data, and conventional models cannot exploit them.

Intellico’s approach addresses both problems at once. Starting from ordinary tabular formulation data, our model learns the graph of relationships between formulations on its own. Each formula becomes a node; the connections between nodes — learned automatically by a Differentiable Graph Module (DGM) — capture how similar formulations are to each other. The model simultaneously learns this graph and uses it to predict the properties of new formulations, uncovering correlations that remain invisible in a spreadsheet.

On top of the predictive engine sit three components designed for the people who actually develop products. An Explainer applies graph-based explainability, perturbation and sensitivity analysis to show which characteristics drive a given property and which past formulas matter most. An Explorer lets the R&D team navigate the learned graph visually to investigate promising regions of the formulation space. And a Simulator predicts the outcome of a new trial before it is ever run in the laboratory.

The project started from the company’s historical trials in the emulsion category: 1,089 formulation trials described by 1,332 commercial ingredients, 915 INCI ingredients and 70 physicochemical properties. Rich, but messy — features were not consistently recorded across trials, and several physicochemical measurements were missing.

Working side by side with the company’s formulation experts, we reduced the feature space to 17 carefully selected parameters, prioritizing data completeness and relevance to the target properties: pH and viscosity. Incomplete formulations were filtered out, leaving 300 complete, reliable trials; remaining gaps were imputed and variables normalized. Deliberately minimal feature engineering kept the model interpretable and reduced the risk of overfitting.

The historical dataset also included formulation variations that were too broad and did not always reflect the formulators’ day-to-day activities. As a result, the model’s predictions were not sufficiently accurate when only minor adjustments were made to a formulation. To address this, we added more than 100 representative formulation trials, together with data on the molecular structures of the ingredients.

Predictive capability only generates value when it can be used simply and immediately. This principle underpins how Intellico’s MATILDE® solution is structured today: after simulating the viscosity of a candidate formula, they immediately see neighbouring formulations with similar expected behaviour, with colour intensity mapping the magnitude of the property. Alongside the graph, an ingredient impact analysis quantifies how a 5% increase in each ingredient shifts the predicted viscosity — showing, for example, that glycerin and sodium lactate push viscosity up while sodium lauryl sulfate and silica pull it down. Formulators don’t just get a prediction; they understand what is driving it.

Tested against historical formulation data, the models proved highly accurate. For viscosity, the model reached 95.6% accuracy with a mean absolute percentage error (MAPE) of 4.4%. For pH, accuracy was even higher: 96.2%, with a MAPE of 3.8%. In practical terms, the models reliably capture the relationship between ingredient combinations and the resulting product properties within the formulation space defined by the dataset.

For the company, this translates into a drastic reduction in laboratory effort and material waste: weak candidates are screened out digitally, and lab capacity is concentrated on the formulations most likely to succeed. Beyond the immediate gains, the explainability features help researchers focus on the ingredient combinations with the highest potential, while the structured, graph-based representation of formulation knowledge supports long-term organizational learning — preserving expertise, making historical data accessible, and easing the onboarding of new staff.

To turn the pilot into stable, long-term adoption, a two-phase roadmap is in place: first, enriching data quality across all key parameters through targeted, AI-driven data collection; second, extending the model to additional product categories — sunscreens, cleansers — and additional performance parameters such as process settings and sensory attributes.

The challenge this company faced is anything but unique. Wherever product development depends on long experimental histories — cosmetics, food, pharmaceuticals, materials science — the knowledge that determines competitiveness sits locked in tables of past trials and in the heads of experienced formulators. Machine learning can unlock it, but only if the model’s reasoning is visible, verifiable and usable by the experts who own the decision.