Sidi Mohammed ALAOUI will defend his doctoral thesis on Thursday, November 6, 2025: “Optimization of a Sensor Network using Machine Learning Techniques for Pollutant Source Identification”

This thesis focuses on the optimization of sensor networks using machine learning techniques for the fast and reliable localization of atmospheric pollutant sources. It combines dispersion models and source identification methods within a unified framework. The objective is to design compact, efficient, and operationally deployable networks capable of accurately estimating both the position and emission rate of an accidental release. A comprehensive state of the art was first established, distinguishing two major families of approaches: the clustering problem, which aims to reduce redundancy and select representative sensors, and the combinatorial optimization problem, which requires defining and minimizing an appropriate cost function. Three benchmark datasets were developed as testbeds: a fully synthetic dataset for methodological comparison, a semi-synthetic dataset derived from the Indianapolis field campaign, and an extended dataset built from eight years of hourly meteorological measurements (2016–2023). In this work, three main contributions are presented. First, a comparative study is conducted between several classification-based approaches applied to sensor network optimization. Second, a new similarity measure is introduced; it is learned through a neural network and improves both spatial coverage and estimation accuracy. Third, a multi-learning method is proposed to combine several specialized models, increasing the robustness of the optimization. Finally, a predictive approach is developed to directly determine the optimal configuration of the sensor network. These results demonstrate that machine learning, whether integrated into classification frameworks or used directly to infer optimal configurations, enables the design of more compact, robust, and operational sensor networks. Future perspectives include the integration of physical constraints, adaptation to real data, uncertainty quantification, and extension to dynamic urban environments.