The research results presented here have demonstrated the feasibility of an innovative temperature control technology for the Metal-Organic Chemical Vapor Deposition (MOCVD) process used in the fabrication of Multi-Quantum Well (MQW) LEDs. The control technology has the strong potential to improve both throughput and performance quality of the manufactured LED. The color of the light emitted by an LED is a strong function of the substrate temperature during the deposition process. Hence, accurate temperature control of the MOCVD process is essential for ensuring that the LED performance matches the design specification. The Gallium Nitride (GaN) epitaxy process involves depositing multiple layers at different temperatures. Much of the recipe time is spent ramping from one process temperature to another, adding significant overhead to the production time. To increase throughput, the process temperature must transition over a range of several hundred degrees Centigrade many times with as little overshoot and undershoot as possible, in the face of several sources of process disturbance such as changing emissivities. Any throughput increase achieved by faster ramping must also satisfy the constraint of strict temperature uniformity across the carrier so that yield is not affected.

SC’s Multiple Input Multiple Output (MIMO) temperature controllers use physics-based models to achieve the performance demanded by our customers. However, to meet DOE’s ambitious goals of cost reduction of LED products, a new generation of temperature controllers has to be developed to ramp up throughput while maintaining temperature accuracy and uniformity. SC believes that this new control technology will be made feasible by the confluence of mathematical formulations of engineering problems that may be well-approximated as convex optimizations, new efficient and scaleable algorithms for solving convex problems, and the increase in computational power available for real-time control.

We have formulated a control system (open and closed-loop) design which can steer the MOCVD heat transfer process that is described by a set of uncertain nonlinear differential equations from an uncertain initial state to as close as possible to a target state while maintaining control constraints and keeping the states (temperatures) within a desired range. Model errors and/or uncertain model parameters severely limit the performance. To overcome these limitations, we have therefore used a real-time, model-adaptation-with-learning strategy.