Machine Learning Approaches Optimizing Semiconductor Manufacturing Processes

Machine learning (ML) technique applied in this study is based on a regression algorithm for optimization, particularly of the deposition process condition and the plasma distribution control knob, the latter being related to the uniformity of film thickness and etch rate within a wafer, which are important properties in semiconductor manufacturing. The ML approach adopted herein is governed by an algorithm of building regression models with tuning parameters as explanatory variables, and the uniformity of film thickness and etch rate as objective variables.<br/> <br/>By definition, film-thickness nonuniformity refers to the 1σ of thickness at 49 points within a wafer. Nonuniformity aims to achieve a target value of close to zero. During the experiment, capacitively coupled plasmas (CCPs) were ignited in an Ar/O₂ gas mixture at 450-kHz frequency, and the input RF power, duration time, and four gas flow rates were tuned. Process optimization was performed separately by an engineer and by application of the ML algorithm. After the fifth iteration, the nonuniformity and its degree of convergence were evaluated. The engineer did not refer to the results by the ML approach during the whole experiment and the database for ML.<br/> <br/>During the film stress optimization of PEALD SiO₂, the RF power, duration time, and four gas flow rates were tuned, whereas the other parameters were fixed. In this study, the target film stress was from −100 MPa to zero.<br/>Accordingly, the process knob parameters were optimized for control of the process profiles. Here profile tuning was performed to control the take-off position of the thickness profile, which had not been achieved by conventional optimization methods.<br/> <br/>The process profiles were controlled with the prepared separated plasma electrode. The separated plasma electrode was installed in CCPs in the Ar/O₂ gas mixture, at a frequency of 13.56 MHz. Moreover, the plasma distribution was controlled by tuning the respective capacitances of the center and the outer (ring) electrodes.<br/>Two experiments were carried out to evaluate the applicability of ML in process tuning: plasma etching for carbon films and PEALD for SiO₂ films. Profiles of the film thickness were measured at 49 points within the wafer, where each input RF power, pressure, gas flow rate and capacitance at each electrode were tuned.<br/> <br/>As mentioned earlier, optimization of the PEALD film-thickness nonuniformity within a wafer was carried out by an engineer and by the ML approach, both separately. Such a wide variation of nonuniformity was attributed to the DOE initial conditions. The engineer could not settle the variation of nonuniformity after the five trials, which was effectively settled by the ML approach. Moreover, the range of nonuniformity by the ML approach at the second trial was 0.7%. These results suggest that the ML approach can find the optimum condition quickly and settle the variation.<br/> <br/>Similarly, stress optimization was performed by both the engineer and the ML approach. The condition obtained by the ML approach in the nonuniformity optimization of TiO₂ was transferred to this SiO₂ process. After 52 attempts of optimizing the process condition based on the engineer’s knowledge, the engineer was unable to achieve the target stress of −100 MPa. By contrast, the conditions created by the ML approach successfully met the target stress.<br/> <br/>The ML approach can be used as a supplement. For example, the engineer can proceed with the experiment by referring to the results of the ML approach. Once the key parameter is suggested via ML, the desired target in the extrapolation area is achievable by human judgment. Fundamentally, a learning result can be transferred when the process conditions are similar. Thus, it is efficient to build a database by using data that are relatively easy to obtain.