S. Jamadagni
Procter & Gamble,
United States
Keywords: molecular modeling, machine learning
Summary:
Data for product development in the consumer goods industry is often expensive to generate and is relatively sparse – samples have to made and various measures have to be obtained (rheology, spectroscopy etc). Further, the diversity of the number of chemical species of interest is very large. Hence conventional ‘big-data’ approaches are usually of limited utility. Further, for models to be useful, they must be very interpretable. In such a situation, combining physics-based mechanistic models with Bayesian approaches to quantify uncertainty in the predictions and enables extrapolation via hierarchical models with only limited additional data is very useful. We are in the early days of adopting this approach in a number of different contexts – from predicting viscosity of formulations, product stability, microbial hostility etc. I will describe a few case studies that illustrate this approach.