Gary S. May

Advantages of plasma etch modeling using neural networks over statistical techniques

Due to the inherent complexity of the plasma etch process, approaches to modeling this critical integrated circuit fabrication step have met with varying degrees of success. Recently, a new adaptive learning approach involving neural networks has been applied to the modeling of polysilicon film growth by low-pressure chemical vapor deposition (LPCVD). In this paper, neural network modeling is applied to the removal of polysilicon films by plasma etching. The plasma etch process under investigation was previously modeled using the empirical response surface approach. However, in comparing neural network methods with the statistical techniques, it is shown that the neural network models exhibit superior accuracy and require fewer training experiments. Furthermore, the results of this study indicate that the predictive capabilities of the neural models are superior to that of their statistical counterparts for the same experimental data. >

An optimal neural network process model for plasma etching

Neural network models of semiconductor processes have recently been shown to offer advantages in both accuracy and predictive ability over traditional statistical methods. However, model development is complicated by the fact that back-propagation neural networks contain several adjustable parameters whose optimal values are initially unknown. These include learning rate, initial weight range, momentum, and training tolerance, as well as the network architecture. The effect of these factors on network performance is investigated here by means of a D-optimal experiment. The goal is to determine how the factors impact network performance and to derive a set of parameters which optimize performance based on several criteria. The network responses optimized are learning capability, predictive capability, and training time. Learning and prediction accuracy are quantified by the experimental error of the model. The process modeled is polysilicon etching in a CCl/sub 4/-based plasma. Statistical analysis of the experimental results reveals that learning capability and convergence speed depend mostly on the learning parameters, whereas prediction is controlled primarily by the number of hidden layer neurons. An optimal network structure and parameter set has been determined which minimizes learning error, prediction error, and training time individually as well as collectively. >

Statistical experimental design in plasma etch modeling

The objective of this work is to obtain a comprehensive set of empirical models for plasma etch rates, uniformity, selectivity, and anisotropy. These models accurately represent the behavior of a specific piece of equipment under a wide range of etch recipes, thus making them ideal for manufacturing and diagnostic purposes. The response characteristics of a CCl/sub 4/-based plasma process used to etch doped polysilicon were examined via a 2/sup 6-1/ fractional factorial experiment followed by a Box-Wilson design. The effects of variation in RF power, pressure, electrode spacing, CCl/sub 4/ flow, He flow and O/sub 2/ flow on several output variables, including etch rate, selectivity, and process uniformity, were investigated. Etch anisotropy was also measured by scanning electron microscopy analysis on a 2/sup 6-2/ fraction of the original experiment. The screening factorial experiment was designed to isolate the most significant input parameters. Using this information as a platform from which to proceed, the subsequent phase of the experiment allowed the development of empirical models of etch behavior using response surface methodology (G. E. P. Box and N. D. Draper, 1987). The models were subsequently used to optimize the etch process. >

Neural network modeling of reactive ion etching using optical emission spectroscopy data

Neural networks are employed to model reactive ion etching (RIE) using optical emission spectroscopy (OES) data. While OES is an excellent tool for monitoring plasma emission intensity, a primary issue with its use is the large dimensionality of the spectroscopic data. To alleviate this concern, principal component analysis (PCA) and autoencoder neural networks (AENNs) are implemented as mechanisms for feature extraction to reduce the dimensionality of the OES data. OES data are generated from a 2/sup 4/ factorial experiment designed to characterize RIE process variation during the etching of benzocyclobutene (BCB) in a SF/sub 6//O/sub 2/ plasma, with controllable input factors consisting of the two gas flows, RF power, and chamber pressure. The OES data, consisting of 226 wavelengths sampled every 20 s, are compressed into five principal components using PCA and seven features using AENNs. Each method is subsequently used to establish multilayer perceptron neural networks trained using error back-propagation to model etch rate, uniformity, selectivity, and anisotropy. The neural network models of the etch responses using both methods show excellent agreement, with root-mean-squared errors as low as 0.215% between model predictions and measured data.

Manufacturing ICs the neural way

To offset the enormous investment of semiconductor companies, chip manufacturers must innovate to a greater degree in the fabrication processes themselves. The aim is to employ the latest computer hardware and software developments-namely, computer-integrated manufacturing of integrated circuits (or IC-CIM)-to optimize the cost-effectiveness of IC manufacture, much as computer-aided design (CAD) revolutionized the economics of circuit design. Here, the author describes how neural networks have recently emerged as a powerful aid in a computer-integrated manufacturing of ICs. Such neural networks help engineers infer subtle input-output relationships.<<ETX>>

Modeling the properties of PECVD silicon dioxide films using optimized back-propagation neural networks

Silicon dioxide films deposited by plasma-enhanced chemical vapor deposition (PECVD) are useful as interlayer dielectrics for metal-insulator structures such as MOS integrated circuits and multichip modules. The PECVD of SiO/sub 2/ in a SiH/sub 4//N/sub 2/O gas mixture yields films with excellent physical properties. However, due to the complex nature of particle dynamics within the plasma, it is difficult to determine the exact nature of the relationship between film properties and controllable deposition conditions. Other modeling techniques, such as first principles or statistical response surface methods, are limited in either efficiency or accuracy. In this study, PECVD modeling using neural networks has been introduced. The deposition of SiO/sub 2/ was characterized via a 2/sup 5-1/ fractional factorial experiment, and data from this experiment was used to train feed-forward neural networks using the error back-propagation algorithm. The optimal neural network structure and learning parameters were determined by means of a second fractional factorial experiment. The optimized networks minimized both learning and prediction error. From these neural process models, the effect of deposition conditions on film properties has been studied, and sensitivity analysis has been performed to determine the impact of individual parameter fluctuations. The deposition experiments were carried out in a Plasma Therm 700 series PECVD system. The models obtained will ultimately be used for several other manufacturing applications, including recipe synthesis and process control. >

Neural network-based real-time malfunction diagnosis of reactive ion etching using in situ metrology data

To mitigate capital equipment investments and enhance product quality, semiconductor manufactures are turning to advanced process control (APC) methods. With the objective of facilitating APC, this paper investigates a methodology for real-time malfunction diagnosis of reactive ion etching (RIE) employing two types of in situ metrology: optical emission spectroscopy (OES) and residual gas analysis (RGA). Based on metrology data, time series neural networks (TSNNs) are trained to generate evidential belief for potential malfunctions in real time, and Dempster-Shafer (D-S) theory is adopted for evidential reasoning. Successful malfunction diagnosis is achieved, with only a single missed alarm and a single false alarm occurring out of 21 test runs when both sensors are used in tandem. From the results, we conclude that the OES and RGA sensors, in conjunction with the TSNN models, can be effectively used for RIE monitoring and diagnosis. Furthermore, D-S theory is shown to be an appropriate inference methodology.

Time series modeling of reactive ion etching using neural networks

Neural networks have been used to model the behavior of real-time tool data in a reactive ion etch (RIE) process. An etch monitoring and data acquisition system for transferring data from the RIE chamber to a remote workstation was designed and implemented on a Plasma Therm Series 700 Dual Chamber etcher. This system monitors gas flow rates, RF power, temperature, pressure, and dc bias voltage. A neural network was trained on the monitored data using the feed-forward, error backpropagation algorithm. This network was used to perform three distinct modeling tasks. First, the network was trained on a subset of ten samples of the time series representing a single process run, and subsequently used to forecast the next data point. In the second task, the network was trained as in the first task, but used to predict the next ten values of the data sequence. In each of the first two tasks, the trained network yielded errors of less than 5%. In the final task, a neural net was used to generate a malfunction alarm when the sampled data did not conform to its previously established pattern. >

Using neural network process models to perform PECVD silicon dioxide recipe synthesis via genetic algorithms

Silicon oxide (SiO/sub 2/) films have extensive applications in integrated circuit fabrication technology, including passivation layers for integrated circuits, diffusion or photolithographic masks, and interlayer dielectrics for metal-insulator structures such as MOS transistors or multichip modules. The properties of SiO/sub 2/ films deposited by plasma enhanced chemical vapor deposition (PECVD) are determined by the nature and composition of the plasma, which is in turn controlled by the deposition variables involved in the PECVD process. The complex nature of particle dynamics within a plasma makes it very difficult to quantify the exact relationship between deposition conditions and critical output parameters reflecting film quality. In this study, the synthesis and optimization of process recipes using genetic algorithms is introduced. In order to characterize the PECVD of SiO/sub 2/ films deposited under varying conditions, a central composite designed experiment has been performed. Data from this experiment was then used to develop neural network based process models. A recipe synthesis procedure was then performed using the optimized neural network models to generate the necessary deposition conditions to obtain several novel film qualities, including zero stress, 100% uniformity, low permittivity, and minimal impurity concentration. This synthesis procedure utilized genetic algorithms, Powells algorithm, the simplex method, and hybrid combinations thereof. Recipes predicted by these techniques were verified by experiment, and the performance of each synthesis method are compared. It was found that the genetic algorithm-based recipes generally produced films of superior quality. Deposition was carried out in a Plasma Therm 700 series PECVD system.

Reactive ion etch modeling using neural networks and simulated annealing

Silicon dioxide films are useful as interlayer dielectrics for integrated circuits and multichip modules (MCMs), and reactive ion etching (RIE) in RF glow discharges is a popular method for forming via holes in SiO/sub 2/ between metal layers of an MCM. However, precise modeling of RIE is difficult due to the extremely complex nature of particle dynamics within a plasma. Recently, empirical RIE models derived from neural networks have been shown to offer advantages in both accuracy and robustness over more traditional statistical approaches. In this paper, a new learning rule for training back-propagation neural networks is introduced and compared to the standard generalized delta rule. This new rule quantifies network memory during training and reduces network disorder gradually over time using an approach similar to simulated annealing. The modified neural networks are used to build models of etch rate, anisotropy, uniformity, and selectivity for SiO/sub 2/ films etched in a chloroform and oxygen plasma. Network training data was obtained from a 2/sup 4/ factorial experiment designed to characterize etch variation with RF power, pressure, and gas composition. Etching took place in a Plasma Therm 700 series RIE system. Excellent agreement between model predictions and measured data was obtained.