(713a) Development of a Parallel Predictive Control Algorithm for Complex Processes with Multiscale Objectives

For many advanced materials and biological processes of interest, performance is measured both with respect to product yield and quality. The elusive latter term is usually a measure of certain aspects of product microstructure that needs to be within strict limits. Sensor information at the microstructure scale is impractical or infeasible, necessitating multiscale process modeling to provide soft sensors for controller synthesis.

Multiscale models are traditionally used to quantify process evolution across all relevant length scales and characterize product behavior within current computational limitations. Such models, however, pose significant challenges for controller design due to a) the unavailability of closed form equations of the molecular-level description components and b) the computational demands of the resulting system that prevent direct implementation inside the control structure. An industrially relevant example for this is the thin film deposition processes widely used by the microelectronics and solar energy industries (such as the production of photovoltaic systems). Motivated by the complex and nonlinear process dynamics, strict quality requirements, high product value, and high production costs, a significant amount of research has been applied to this problem. While earlier work focused on designing structures that allow the use of standard feedback controllers by directly using simulations in lieu of measurements, more recent results have been focused on combinations of model-based controllers with observers. These structures are based on models derived by using system identification methods. Unfortunately, the use of predictive controllers adds an additional computational burden due to mathematical complexity.

To address both the computationally demanding issues of process modeling and predictive control, we investigated the application of parallel computing. Based on our previous efforts, we developed a computationally tractable online system identification scheme based on a combination of subspace and fuzzy system identification methods. This scheme allows for the derivation of stochastic differential equations where both the expectation and drift terms are identified. This model was incorporated into a predictive control scheme that drives the process towards desired objectives. By employing optimization algorithms that account for the nonlinear and uncertain nature of the underlying dynamics, the predictive controller was capable of efficiently identifying near optimal manipulated variable profiles. The change in computing resources available in real time impacted both the design and the algorithms used for both the model and controller. Discussion of each algorithm and the total system are presented along with results for the process under control.