A data-driven methodology on deterministic linear models and probabilistic machine learning for material development

Bayesian optimization has become an appealing machine learning alternative to traditional design of experiments in optimization tasks. A nonparametric surrogate model and an iterative sampling strategy, however, make it difficult to evaluate and compare the effects of the controlled variables on the experimental objective. We report a data-driven methodology for combining the benefits of deterministic linear models and probabilistic machine learning to improve cellulosic airlaids by thermal pressing. Our approach starts with a fractional factorial design as the initial sampling strategy to quantify independent and interpretable variable effects and their interactions. We show how these resource-efficient designs can be easily complemented with few additional experiments to identify more complicated behavior using a formal statistical test and how these tests indicated that a linear interaction model was likely not sufficient at reliably estimating the changes in airlaid properties. We then identified three main weaknesses in traditional design of experiments for optimizing the pressing conditions for our airlaids and replaced these tools with Bayesian optimization. The Bayesian optimization algorithm improved the mechanical and physical properties of our airlaids and identified several promising conditions for airlaid pressing. Our work is an important contribution to improve airlaids by thermal pressing and to bridge the gap between traditional design of experiments and probabilistic machine learning for experimental material development.