Anastasiia DOINYCHKO

[ABSTRACT]:

One of the leading open challenges in the Advanced Process Control (APC) field in Semiconductor Manufacturing (SM) industry is the ability to leverage the wealth of information in order to fully characterize the process in order to propose solutions for automotive process diagnostic and product management. In this thesis, we investigate Machine Learning (ML) based approaches to developing a unified framework for process analysis and modeling from highly diversified multi-view (that comes from different sources) Semiconductor Manufacturing (SM) data. We start with the analysis of existing data treatment techniques and their limitations. Then, we focus on proposing a strategy that consists of data cleaning, features extraction, and features selection steps to deal with data imperfections usually expected in the field, like noise, irregular sampling steps in sensory time series data, and incomplete records, all due to natural corruption-error rate of recording tools.

Next, this thesis intends to expand the scope of traditional process modeling in SM by cross-process analysis. Product manufacturing is a sequential procedure of applying ordered processes to deposit new layers of features; then, one can use the precedent history to learn its impact on the current modeling target of interest. Accordingly, we propose a methodology that benefits not only from different types of measurements but from dependencies between different process steps to make processes more predictable and productive.

Moreover, we study the problem of missing data, mainly when one of the views is entirely missing, which is another open challenge in the field. Some studies tackle this problem by assuming the existence of view generation functions to approximately complete the absent views. However, these functions generally require an external resource to be set, and their quality directly impacts the performance of the final predictive model learned over the completed training set. Instead, in this work, we propose to address this problem by jointly learning the missing views and the multi-view target estimator using an adversarial learning approach inspired by the ability of Generative Adversarial Networks (GANs) to seize the underlying distribution of the data and create new samples.

Finally, for all the hypotheses introduced above in this work, we consider the APC tasks like Virtual Metrology (VM) and Predictive Maintenance (PdM) to conduct experiments using the real data collections provided by leading in Europe SM fabrication facilities that we collaborated with, within the scope of the Metrology Advances for Digitized Electronic Components and Systems (ECS) Industry 4.0 (MADEin4) Project. We also consider at how we can use the models we have developed to classify multilingual documents and Electronic Health Record (EHR) data.