Using a neural network, a refractive index (RI) of silicon nitride film was predicted as a function of process parameters, including radio frequency (RF) power, pressure, substrate temperature, and SiH 4 , NH 3 , and N 2 flow rates. The film was deposited by a plasma-enhanced chemical vapor deposition (PECVD) system. The PECVD process was characterized by a 2 6 1 fractional factorial experiment. Particular emphasis was placed on examining temperature effects at low pressure. Model prediction accuracy was optimized as a function of training factors. Predicted parameter effects were experimentally validated. Plots generated from an optimized model were used to qualitatively estimate deposition mechanisms. It is noticeable that under various plasma conditions, the RI varied little with the temperature. The temperature effect was extremely sensitive to the pressure level. Enhanced ion bombardment at high temperatures yielded a Si-rich film. Effect of each gas was little affected by the temperature. The SiH 4 flow rate played the most significant role in determining the RI at low pressure.Index Terms-Modeling, neural network, plasma-enhanced chemical vapor deposition (PECVD), silicon nitride (SiN) film.
A GA-PNN was constructed and applied to a model plasma etch process. The etch process was characterised by a statistical experimental design. The performance of GA-PNN was compared with other models including the three types of statistical regression models, conventional backpropagation neural network, polynomial neural network and adaptive network based fuzzy inference system models. In all comparisons, the GA-PNN demonstrated signi cantly improved predictions. Moreover, the network complexity of conventional PNN could be signi cantly reduced. This indicates that the GA-PNN is an effective means to construct a predictive model for poorly de ned complex systems characterised by the limited data set. SE/S276 Drs Kim and Park are in theDepartment INTRODUCTIONPlasma etching is a key means of forming ne patterns for manufacturing integrated circuits.Owing to complex chemical reactions between plasma variables, ions and radicals, and material surfaces, it has been extremely dif cult to model plasma etching. There have been many reports on constructing predictive etch models using intelligent systems such as backpropagation neural network (BPNN) 1 -3 and fuzzy logic. 4 Recently, a polynomial neural network (PNN) 5 was applied and demonstrated improved prediction over conventional statistical regression models or BPNN. Despite the advantages, constructing a PNN model is complicated by the presence of several training factors, whose optimal values are initially unknown. These may include the number of partial descriptions (PDs) in each layer, the number of input variables connected to each PD, or the order of polynomials assignable to each PD. 6 In most cases, they are determined by the trial and error method, thereby causing a heavy computational burden and low ef ciency. Moreover, the heuristic nature of the PNN algorithm does not guarantee that the PNN model constructed has the best predictive ability. In this paper, a method of circumventing the drawbacks of current PNNs is presented. This is accomplished using a genetic algorithm (GA) 7 to optimise the training factors. For convenience, this type of PNN is referred to as a 'GA-PNN'. The GA is used to search for one optimal set of training factors, including the number of input variables to each PD, the selection of input variables and the type of polynomials of PDs. The performance of the GA-PNN is compared with statistical regression models, conventional BPNN and PNN, and adaptive networks based fuzzy inference system (ANFIS) models. Model
Deposition rate of silicon nitride films was modelled using a neural network in conjunction with Box—Wilson experimental design. The films were deposited by a plasma enhanced chemical vapour deposition system. A neural network model was constructed and tested with 33 and 12 experiments respectively. Model prediction performance was significantly improved by applying genetic algorithm to training factor optimisation. To qualitatively estimate deposition mechanisms with the parameters, several plots were generated from the model and emphasis was placed on the investigation of the radio frequency power effect on the deposition rate under various plasma conditions. An increase in the deposition rate with increasing power was ascribed to more Si deposition near the reaction surface. The SiH4 effect was the most complex depending on the powers. In contrast, the power effect was insensitive to the variations in N2 flow rate. The power effect at high pressure was attributed to increased concentration of Si radicals in the gas phase. A comparison with a refractive index model facilitated to infer deposition mechanisms.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.