ABSTRACT Despite their ubiquitous presence in scientific publications and media, the number of artificial intelligence and machine learning (AI/ML) systems installed and operating in industry is still incipient. Issues such as robustness, memory requirements, maintenance, and interpretability hinder their diffusion and wider adoption. In this article, we report a system developed to operate in the circular economy of lubricant oil that cleared all of these requirements and is now under supervised adoption in industry. We also share important refinement stages that were crucial for the ML algorithm to work properly and deliver the aimed aided‐value for decision‐making. More specifically, to expedite decision‐making and mitigate safety risks, a robust classification method is proposed to predict the coagulation potential of waste lubricant oil (WLO), a hazardous waste according to the European legislation. The occurrence of the coagulation phenomena renders WLO regeneration into base oil production unfeasible. Leveraging on Fourier‐transform infrared (FTIR) spectroscopy and ML, the proposed methodology comprises three key modules: feature extraction and validation of the model, robust classification using risk‐adjusted rules, and soft validation of negative predictions. The original PAT‐based ML classification model achieved only 70% accuracy, mostly due to misclassification of WLO that coagulates, which represents the highest risk for the process operation. In turn, the proposed methodology was able to reliably classify the WLO samples, showcasing its value for decision‐making together with a significant reduction in laboratory workload.
Gariso et al. (2026) studied this question.