LAUSR

Abstract Refractive index (RI) is a highly relevant property for the design of new polymeric materials for very specific applications in the telecommunications industry, medicine, and analytical chemistry, among many others. A particular case is that of plastic optical fibers, in which the information is transmitted by photons and then RI takes center stage. Therefore, the modeling and prediction of this property play a key role when characterizing and designing materials for these important industries. Over the last decades, the use of Machine Learning (ML) algorithms in the modeling of properties for the design of new materials has been consolidated thanks to the gradual increase in the available databases. In particular, the development of Quantitative Structure-Property Relationship (QSPR) models has benefited from these emerging technologies, providing the possibility of generating in silico testing strategies applicable to the early stages of the design of new materials. However, in many cases, it has been observed that using ML algorithms in a fully automatic way, without human intervention in the QSPR model design process, tend to generate black-box models that have a difficult interpretation and can lose sight about relevant aspects that require both criteria and an expert’s knowledge in the chemical domain. For this reason, interactive ML methodologies that combine computational outputs with experts’ knowledge, usually known as expert-in-the-loop strategies, are becoming more frequent. In this article, we present the design of QSPR models for RI modeling following two different approaches, a black-box ML methodology and an Interactive Machine Learning (IML) methodology with expert-in-the-loop, from a database whose curation is also described in the present work. In this regard, visual analytics strategies were used to capture the expert’s knowledge, facilitating an effective and rapid interaction between the outputs provided by ML and the chemical analyst. In addition, we contrast the best models obtained by both approaches against two other predictive models for RI estimation reported in the literature, achieving promising performances in terms of cardinality and accuracy when the expert interacts during modeling. In summary, the obtained results allow us to claim that the expert-in-the-loop approach provides QSPR models with better generalizability properties and more interpretable from a physicochemical point of view, without losing accuracy. Finally, in addition to providing high quality QSPR models to predict the RI of polymeric materials, the present work lays the foundation for defining an effective methodology to incorporate experts’ knowledge in the design of other material properties.