HYBRID EQUATION-BASED AND DATA-DRIVEN COMPUTATIONAL WORKFLOWS FOR ANALYSIS AND PREDICTION OF INDUSTRIAL DEPOSITION PROCESSES

Despite the leaps in quality and quantity of industrial data along with the increased interest in data-driven approaches brought about by Industry 4.0, there are still processes that are too complex to be accurately modeled via traditional first-principles methods, yet lack the necessary data for a purely data-driven approach.

Taking an industrial chemical vapor deposition (CVD) process as a key example, this work proposes a hybrid computational workflow involving equation-based (computational fluid dynamics - CFD) and machine learning (ML) methods for the modeling, investigation, and prediction of such complex processes. First, this work aims to provide a way of predicting process outcomes while also allowing the exploration of the process and obtaining insights regarding the several interplaying physical and chemical phenomena that govern it. The proposed CFD model can help with the exploration of the process and the prediction of the quality quantity of interest, which is the thickness of the deposited alumina. It can also shed light on the governing phenomena of the process. However, it comes with a high computational cost, which makes its use in everyday applications prohibitive. To overcome this, a purely data-driven predictive model which offers improved predictive and computational performance is proposed. A way of combining process data and the results of the CFD model is also proposed, via the GappyPOD method.

Subsequently, this work proposes a purely data-driven approach to identify potential critical process parameters based on a blend of supervised and unsupervised learning approaches. Following an initial clustering of the available process outcome data and the analysis of the resulting clusters, the differences between them can be matched to the differences in their respective process inputs, allowing the identification of potential key parameters. These parameters allow for deeper insight into the process and can then be used to develop data-driven models for the qualitative and quantitative prediction of the process. The versatility of this approach is then highlighted by its application to a vastly different process; the metabolism of astrocyte cells.