Machine learning and design of experiments with an application to product innovation in the chemical industry

Design of Experiments and machine learning for product innovation: A systematic literature review

“…In the following sections, we will observe that, in addition to the advantages of the DOE + ML framework highlighted in the industrial application described by Arboretti et al, 6 several other advantages have been found in the literature.…”

“…In this section, as a motivating example, we present an industrial application that successfully employs the DOE + ML framework. The case study is taken from the work of Arboretti et al, 6 and consists in the application of DOE + ML as a way to collect and analyze data for the development of a new detergent. The main objective was to develop models that can reliably predict washing performance on three different types of stains, based on the components of the detergent's formula.…”

“…At the same time, the estimation of variable importance and uncertainty quantification, together with an extensive analysis and discussion of the results, prove that ML models can be applied on DOE data without necessarily losing much in terms of interpretability with respect to parametric approaches. The authors 6 state that the developed algorithms can be included in a semi-automatic system for prediction of the washing performance, thus reducing the need for future experiments and facilitating the process of new product development.…”

“…The analyst should strive to select or develop ML algorithms that can quantify uncertainty of predictions and can be, at least to a certain extent, interpretable. Some works in the literature already acknowledge the importance of these aspects, 6 and great potential is shown by the AL and BO approaches since both take the uncertainty of predictions into account when suggesting the next experimental configuration to test. 3,27 Furthermore, over the years, some model-agnostic tools have been proposed to make ML more interpretable, 40 but this is a growing field of research.…”

“…Some work has been carried out on this topic, but it only presented a general background of the literature and a case study application. 6 Therefore, the aim of this article is to present a systematic literature review (SLR) covering the publications on the joint application of DOE and ML in a PI setting. To the best of our knowledge, this is the first comprehensive literature review of these topics, therefore our aim is to identify the state of the art of this branch of literature for the first time, identify opportunities, and propose future research directions.…”