Process modeling, optimization and control using artificial neural networks

This thesis reports the results of research in modeling, optimization and run-to-run control of complex manufacturing processes through the use of artificial neural networks (ANNs). The test vehicle is the chemical vapor deposition of silicon in a barrel reactor. Other manufacturing processes are also included, where necessary, to elucidate some aspects of the research.In the introductory chapter, the motivation and objective of the present study are addressed, and a framework of the ANN control system is proposed.Chapter 2 describes the ANN modeling techniques for both dense data and sparse data. The principle and configuration of ANNs are briefly illustrated. The horizontal CVD reactor and boiling curve are chosen as examples for the dense data case. It is shown that ANNs can effectively map very nonlinear processes, including discrete relationships. When a physical model is available, enough data points generated by the physical model can be used to train an ANN model. This model supported by a physical model is referred to as the physico-neural model and is shown to have good generalization capability. The barrel CVD reactor is chosen as an example for the sparse data case. To map the input-output relationship, ANNs with different configurations are trained and tested by designed experimental data. A "simple to complex" approach is proposed to determine the best net configuration.Chapter 3 deals with process modeling and optimization technique. An artificial neural network response surface methodology (ANNRSM) is proposed for process modeling and optimization. An ANN model is built by "simple to complex" approach and is then used in conjunction with a gradient search scheme to ascertain input settings for optimal output. The results of using the ANNRSM in identifying optimum settings in the presence of noise show the applicability of the technique to noisy data. A laboratory experimental study based on a mock-up CVD reactor supports the optimum settings obtained by the ANNRSM. A comparison between the ANNRSM and regression RSM shows that the ANNRSM is able to build an accurate global model and find the optimum using fewer data especially when the data points are noisy.Chapter 4 deals with an ANN model based run-to-run controller. The ANN run-to-run controller is an integration of the ANN modeling, statistical process control (SPC), and automatic process control (APC) techniques. The procedure for design and optimization of an ANN run-to-run controller is discussed in detail. The controller model is extracted from the ANN process model by Taylor expansion and inversion. An EWMA technique is used to detect the process shift/drift and filter out the output noise. The control action is determined by feedback to compensate for the process shift and slow drift. It is found that a dual controller does a good job in controlling of a process with either a small or big shift. A total cost criterion is proposed for optimizing the run-to-run controller. Simulation calculations of the total cost are also performed. Experiments from the laboratory mock-up reactor demonstrate the effectiveness of the proposed ANN run-to-run controller.In the concluding chapter, the major contributions of this study are highlighted, and future work is proposed.