Development of explainable AI methods for semiconductor metrology systems

This thesis was written during my internship period at ASML in Veldhoven, Netherlands. During this internship I studied and tested different AI explainability techniques to analyze the internal decision-making process of Machine Learning models trained on optical metrology data measured during semiconductor production. I first give an introduction about the tested methods, from the most simple to the most complicated ones, giving a theoretical comparison of their behaviour and analyzing how each of them fits the constraints posed by the problem. Different metrics are then introduced: they aim to give a quantitative comparison between the results of all the approaches in order to understand how accurately they capture the true reasoning of the model. After a first successful test on a simulation of the real environment, the methods are applied to production data. The results on real data show that gradient-based methods, which evaluate an explanation using the gradients from the non-linear model successfully pass all the metrics, maintain an extremely efficient computation cost and give similar results on different datasets. This analysis proves that the studied methods can be safely used for the explainability, facilitating the deployment of the models on real-world applications.