Graph Machine Learning for Fast Product Development from Formulation Trials

Product development is the process of creating and bringing a new or improved product to market. Formulation trials constitute a crucial stage in product development, often involving the exploration of numerous variables and product properties. Traditional methods of formulation trials involve time-consuming experimentation, trial and error, and iterative processes. In recent years, machine learning (ML) has emerged as a promising avenue to streamline this complex journey by enhancing efficiency, innovation, and customization. One of the paramount challenges in ML for product development is the models’ lack of interpretability and explainability. This challenge poses significant limitations in gaining user trust, meeting regulatory requirements, and understanding the rationale behind ML-driven decisions. Moreover, formulation trials involve the exploration of relationships and similarities among previous preparations; however, data related to formulation are typically stored in tables and not in a network-like manner. To cope with the above challenges, we propose a general methodology for fast product development leveraging graph ML models, explainability techniques, and powerful data visualization tools. Starting from tabular formulation trials, our model simultaneously learns a latent graph between items and a downstream task, i.e. predicting consumer-appealing properties of a formulation. Subsequently, explainability techniques based on graphs, perturbation, and sensitivity analysis effectively support the R&D department in identifying new recipes for reaching a desired property. We evaluate our model on two datasets derived from a case study based on food design plus a standard benchmark from the healthcare domain. Results show the effectiveness of our model in predicting the outcome of new formulations. Thanks to our solution, the company has drastically reduced the labor-intensive experiments in real laboratories and the waste of materials.