| Summary: | 碩士 === 中原大學 === 化學工程研究所 === 97 === In recent years, semiconductor manufacturing has swiftly evolved as a lucrative industry with high capital investment. Most actual manufacturing processes experience deterministic changes, such as set-point changes and process drifts. In addition, the output variable of the process is usually subject to random inputs, such as metrology delay of uncertain duration and non-Gaussian disturbances which can be treated as random variables that follow a specific probability density function (PDF). Conventional control is unable to minimize both sources of variation. Therefore, it is necessary to develop a novel control method to improve the performance of conventional control. In response to this need, this research proposes a hybrid run-to-run (RtR) controller called deterministic and stochastic model based control (DSMBC) strategy for the semiconductor production processes with the stochastic distribution characteristics. The DSMBC strategy merges the features of both models into an effective process control method.
The deterministic RtR controller (exponentially weighted moving average or EWMA controller) responds quickly to deterministic changes, such as set-point changes and drifts. But it could not overcome the hysteresis of control performance and closed loop instability caused by delay of information feedback of process outputs due to metrology delay of uncertain duration. The stochastic model based controller (minimum entropy control or MEC) is effective in minimizing effects due to the random disturbances and metrology lag of varying duration. But it is slow in responding to set-point changes. In designing an optimal controller that has the strengths of both controllers, the proposed DSMBC strategy formulates a novel performance index into a weighted sum of stochastic entropy of output errors, deterministic prediction of square of the output errors and the control energy constraint. This modification also allows the optimal controller to counteract the effects caused by the drift of aging of the process with batch processing and model mismatch. Model mismatch is eliminated as the DSMBC strategy uses the feedback information of the measured process output from every batch to accomplish the data-driven updating of the model parameters. In addition, the proposed data-driven RtR controller of the DSMBC strategy can also be applied to semiconductor industries which use multi-devices for parallel production of different products (also called MTMP processes). The observation vector is used to pool together the measured output information obtained at each batch to construct the deterministic and stochastic MTMP models.
Finally, simulations of the tungsten chemical-vapor deposition process and shallow trench isolation process demonstrate the power, effectiveness and feasibility of the proposed data-driven RtR controller employing the DSMBC strategy in STSP and MTMP processes.
|
|---|
