Search

A new technique for design centering and yield enhancement of devices and circuits is presented. The proposed method uses neural networks for device and/or circuit modeling and genetic algorithms for parametric yield optimization. It uses a Monte Carlo-based method for yield estimation via the neural models (thus consuming less time) and genetic algorithms for efficient design centering. The neurogenetic methodology has been used for design centering of SiGe heterojunction transistors and millimeter-wave voltage controlled oscillators. It results in significant yield enhancement of the SiGe heterojunction bipolar transistors (from 25% to 75%) and voltage controlled oscillators (from 8% to 85%). To the best of our knowledge, this method has not been reported previously. Index Terms--Genetic algorithms (GA), Monte Carlo method, neural networks, parametric yield estimation.

A hybrid neural network model is constructed by characterizing the growth of Ga[As.sub.1-y][P.sub.y]-GaAs superlattices (SLs) grown on (001) GaAs substrates by molecular beam epitaxy. These heterostruetures are formed by the [P.sub.2] exposure of an As-stabilized GaAs surface, and ex situ high-resolution X-ray diffraction (HRXRD) is performed to determine the phosphorus composition at the interfaces. A first-order kinetic model is then developed to describe the mechanisms of anion exchange, surface desorption, and diffusion. A semi-empirical hybrid neural network is used to estimate the parameters of the kinetic model and analyze the microscopic processes occurring at the interfaces of the mixed anion III-V heterostructures. The phosphorus diffusion process in GaAs is estimated to have a diffusion coefficient of D = 1.4 x [10.sup.-14] exp(-0.11 eV/[k.sub.B][T.sub.s]) [cm.sup.2] * [s.sup.-1] for samples with [P.sub.As4] = 4 x [10.sup.-6] torr and exhibits enhanced phosphorus intermixing for samples with lower As-stabilizing fluxes. Index Terms--Anion exchange, hybrid neural networks, kinetic modeling, molecular beam epitaxy.

To achieve timely and accurate fault detection in reactive ion etching, neural networks (NNs) have been applied for the fusion of data generated by two in-situ sensors: optical emission spectroscopy (OES) and residual gas analysis (RGA). While etching is performed, OES and RGA data are simultaneously collected in real time. Several pre-determined, statistically significant wavelengths (for OES data) and atomic masses (for RGA signals) are monitored. These data are subsequently used for training NN-based time series models of process behavior. Such models, referred to herein as time series NNs (TSNNs), are realized using multilayered perceptron NNs. Results indicate that the TSNNs not only predict process parameters of interest, but also efficiently perform as sensor fusion of the in-situ sensor data. Index Terms--Neural networks (NNs), optical emission spectroscopy (OES), reactive ion etching (RIE), real-time fault detection, residual gas analysis (RGA), sensor fusion, time series modeling.

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. Index Terms--Dempster-Shafer (D-S) theory, malfunction diagnosis, neural networks, optical emission spectroscopy (OES), reactive ion etching (RIE), residual gas analysis (RGA).

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 S[F.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. Index Terms--Autoencoder neural networks, neural networks, optical emission spectroscopy, principal component analysis, reactive ion etching.

This paper explores the use of neural networks for real-time, model-based feedback control of reactive ion etching (RIE). This objective is accomplished in part by constructing a predictive model for the system which can be approximately inverted to achieve the desired control. An indirect adaptive control (IAC) strategy is pursued. The IAC structure includes a controller and plant emulator, which are implemented as two separate back-propagation neural networks. These components facilitate non-linear system identification and control, respectively. The neural network controller is applied to controlling the etch rate and dc bias while processing a GaAs/AlGaAs metal--semiconductor--metal (MSM) structure in a B[Cl.sub.3]/[Cl.sub.2] plasma using a Plasma Therm 700 SLR series RIE system. Results indicate that in the presence of disturbances and shifts in RIE performance, the IAC neural controller is able to adjust the recipe to match the etch rate and dc bias to that of the target values in less than 5 s. These results are shown to be superior to those of a more conventional LQG/LTR control scheme based on a linearized transfer function model of the RIE system. Index Terms--Neural networks, reactive ion etching, real-time control.

Multichip module (MCM) technology is considered a strategic solution in electronics packaging because this approach offers significant advantages in electrical and thermal performance and reliability. However, manufacturing cost is a critical issue for mass production of high-performance MCM packages. To realize low-cost manufacturing technology, process modeling, optimization, and control techniques are required. In this paper, a modeling approach for via formation in MCM dielectric layers composed of photosensitive benzocyclobutene (BCB) is presented. A series of designed experiments are used to characterize the via formation workcell (which consists of the spin coat, soft bake, expose, develop, cure, and plasma descum unit process steps). The output characteristics considered are film thickness, refractive index, uniformity, film retention, and via yield. Sequential neural network process models are constructed to characterize the entire process. In the sequential scheme, each workcell subprocess is modeled individually, and each subprocess model is linked to previous subprocess outputs and subsequent subprocess inputs. This modeling scheme is compared with both the global and unit process modeling approaches to evaluate model prediction capability. The sequential method shows superior capability, with an average rms prediction error of 6.40% over all responses, compared to a 11.61% rinse for the global model and a 12.05% error for the unit process models. Index Terms - MCM's, modeling, neural networks, via formation.

This paper presents a novel technique for monitoring film thickness in reactive ion etching by incorporating a micromachined sensor. The prototype sensor correlates film thickness with the change in resonant frequency that occurs in a micromachined platform during etching. The platform is suspended over a drive electrode on the surface of the substrate and electrically excited into resonance. As material is etched from the platform, its resonant vibrational frequency shifts by an amount proportional to the amount of material etched, allowing etch rate to be inferred. As a proof-of-concept experiment, a platform made of DuPont 2611 polyimide has been fabricated. The sensor is driven into resonance electrostatically, and the shift in resonance is detected by monitoring the change in impedance between the drive electrode and platform as the drive frequency is swept. To enhance filtering of the sensor signal in the noisy plasma environment, the platform is designed so that the ratio of the plasma frequency to the fundamental mode of vibration is approximately 400:1. The prototype was etched in a Plasma Therm 700 series reactive ion etching (RIE) system in a CH[F.sub.3]/[O.sub.2] plasma. Electrical contact was made with the sensor using a feedthrough attached to the vacuum line beneath the process chamber to facilitate in situ excitation and measurement. The sensor is shown to offer high resolution (approximately 1300 Hz/um), potentially permitting accurate in situ monitoring of etch rate and uniformity at a nominal cost. Index Terms - Micromachining, plasma etching, process monitoring, sensors.

This paper presents a novel procedure for predicting integrated circuit parametric performance and yield when provided with sample transistor test results and a circuit schematic. Two enhancements to the existing Monte Carlo simulation procedures are described: 1) a multivariate nested model is used to reproduce random process-induced device variations, rather than the multivariate multinormal model typically used, and 2) the stochastic Monte Carlo method for mapping process variability into a performance distribution is replaced with a deterministic mapping technique. The use of multivariate nested distributions allows estimation not only of correlation between various model parameters, but also allows each of those variations to be apportioned among the various stages of the process (i.e., wafer to wafer, lot to lot, etc.). This allows matched devices to be more accurately simulated, without having to develop customized models for each configuration of matching, and provides focus for process improvement efforts into those areas with the maximum potential reward. The use of deterministic mapping provides simulation results which are repeatable and do not rely on chance to insure that the process parameter space has been evenly explored. A software package which implements the entire procedure has been written in C++. Index Terms - Monte Carlo simulation, multivariate statistics, parametric yield.

Silicon oxide (Si[O.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 Si[O.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 Si[O.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, Powell's 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.

Metal-organic chemical vapor deposition (MOCVD) is an important technique for growing thin films with various applications in electronics and optics. The development of accurate and efficient MOCVD process models is therefore desirable, since such models can be instrumental in improving process control in a manufacturing environment. This paper presents a semi-empirical MOCVD model based on 'hybrid' neural networks. The model is constructed by characterizing the MOCVD of titanium dioxide (Ti[O.sub.2]) films through the measurement of deposition rate over a range of deposition conditions by a statistically designed experiment in which susceptor and source temperature, flow rate of the carrier gas for the precursor and chamber pressure are varied. A modified backpropagation neural network is then trained on the experimental data to determine the value of the adjustable parameters in an analytical expression for the Ti[O.sub.2] deposition rate. In so doing, a general purpose methodology for deriving semi-empirical neural process models which take into account prior knowledge of the underlying process physics is developed.

Metal-organic chemical vapor deposition (MOCVD) performance is optimized for growing titanium dioxide (Ti[O.sub.2]) thin films. Different gas flow directions and susceptor rotation, along with reactor geometry and shape variations are considered. Gravity proves to be an important parameter in changing the flow pattern in the reaction chamber. However, since film uniformity is not improved by changing the flow direction, modifying the reactor geometry is also proposed. Among the different geometrical parameters, the susceptor-inlet distance, inlet tube diameter, and susceptor size are considered. To minimize the occurrence of recirculation cells in the reaction chamber, modifications in the reactor shape are also suggested. Acceptable results are achieved by changing the cylindrical reactor to a diamond shape.

The impact of different doping techniques on the reliability of gallium arsenide/aluminum gallium arsenide multiple quantum well avalanche photodiode (APD) structures constructed by molecular beam epitaxy was studied. Analysis of variance revealed that dark current is affected mainly by device diameter, temperature and stressing time, and breakdown voltage depends on the diameter, stressing time, and APD type. Findings indicated that the undoped APD has the highest reliability, followed by the doped-well and doped-barrier devices, respectively.

Silicon nitride films grown by plasma-enhanced chemical vapor deposition (PECVD) are useful for a variety of applications, including anti-reflection coatings in solar cells, passivation layers, dielectric layers in metal/insulator structures, and diffusion masks. PECVD nitride films are known to contain hydrogen, and defect passivation by hydrogenation enhances efficiency in polycrystalline silicon solar cells. PECVD systems are controlled by many operating variables, including RF power, pressure, gas flow rate, reactant composition, and substrate temperature. The wide variety of processing conditions, as well as the complex nature of particle dynamics within a plasma, makes tailoring [Si.sub.3][N.sub.4] film properties very challenging, since it is difficult to determine the exact relationship between desired film properties and controllable deposition conditions. In this study, silicon nitride PECVD modeling using neural networks has been investigated. The deposition of [Si.sub.3][N.sub.4] was characterized via a central composite experimental design, and data from this experiment was used to train optimized feed-forward neural networks using the back-propagation algorithm. From these neural process models, the effect of deposition conditions on film properties, has been studied. It was found that the process parameters critical to increasing hydrogenation and therefore enhancing carrier lifetime in polysilicon solar cells are temperature, silane, and ammonia flow rate. The deposition experiments were carried out in a Plasma Therm 700 series PECVD system.

A reactive ion etching system uses a back-propagation neural network-based system for predicting the etch endpoint of a silicon dioxide film in a CHF3/O2 plasma. The neural network correlates the etch rate with the dc bias to determine the remaining time necessary to etch a known thickness of silicon dioxide. Monitoring of the dc bias is done using a real time data acquisition system, where process conditions are sent from a Plasma Therm 700 series reactive ion etching system. Results obtained using the network are discussed.

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 back-propagation 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.

Neural network models of semiconductor processes have recently been shown to offer advantages in both accuracy and predictive ability over traditional statistical methods (1)-(4). 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 CCl4-based plasma (5). 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.

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.

In-line measurements and electrical test data have historically been used to detect process fluctuations in an integrated circuit manufacturing facility. Unfortunately, these methods alone cannot rapidly identify possible problems in processes with a very narrow range of acceptable performance. When unreliable equipment performance causes processes to vary beyond desired limits, product quality is jeopardized. It is therefore essential that root causes for the malfunctions be diagnosed and corrected quickly to prevent expensive misprocessing. This paper presents a general methodology for the automated diagnosis of integrated circuit fabrication equipment. This technique combines the best aspects of quantitative algorithmic diagnosis and qualitative knowledge-based approaches. Evidence from equipment maintenance history, real-time tool data and in-line measurements are integrated using evidential reasoning (1). This methodology is applied to the identification of faults in the Lam Research Autoetch 490 automated plasma etching system located in the Berkeley Microfabrication Laboratory.

The modeling of semiconductor manufacturing processes has been the subject of intensive research efforts for years. Physical-based (first-principle) models have been shown to be difficult to develop for processes such as plasma etching and plasma deposition, which exhibit highly nonlinear and complex multidimensional relationships between input and output process variables. As a result, many researchers have turned to empirical techniques to model many semiconductor processes. This paper presents a neurofuzzy approach as a general tool for modeling chemical vapor deposition (CVD) processes. A five-layer feed-forward neural network is proposed to model the input--output relationships of a plasma-enhanced CVD deposition of a SiN film. The proposed five-layer network is constructed from a set of input--output training data using unsupervised and supervised neural learning techniques. Product space data clustering is used to perform the partitioning of the input and output spaces. Fuzzy logic rules that describe the input-output relationships are then determined using competitive learning algorithms. Finally, the fuzzy membership functions of the input and output variables are optimally adjusted using the backpropagation learning algorithm. A salient feature of the proposed neurofuzzy network is that after the training process, the internal units are transparent to the user, and the input--output relationship of the CVD process can be described linguistically in terms of IF-THEN fuzzy rules. Computer simulations are conducted to verify the validity and the performance of the proposed neurofuzzy network for modeling CVD processes. Index Terms--Chemical vapor deposition (CVD), fuzzy logic systems, neurofuzzy system, neural networks, supervised learning.

This paper presents the systematic characterization of the molecular beam epitaxy (MBE) process to quantitatively model the effects of process conditions on film qualities. A five-layer, undoped AIGaAs and InGaAs single quantum well structure grown on a GaAs substrate is designed and fabricated. Six input factors (time and temperature for oxide removal, substrate temperatures for AIGaAs and InGaAs layer growth, beam equivalent pressure of the As source and quantum well interrupt time) are examined by means of a fractional factorial experiment. Defect density, X-ray diffraction, and photoluminescence are characterized by a static response model developed by training back-propagation neural networks. In addition, two novel approaches for characterizing reflection high-energy electron diffraction (RHEED) signals used in the real-time monitoring of MBE are developed. In the first technique, principal component analysis is used to reduce the dimensionality of the RHEED data set, and the reduced RHEED data set is used to train neural nets to model the process responses. A second technique uses neural nets to model RHEED intensity signals as time series, and matches specific RHEED patterns to ambient process conditions. In each case, the neural process models exhibit good agreement with experimental results. Index Terms--Molecular beam epitaxy (MBE), neural networks, process modeling.

Via formation is a critical process sequence in multichip module (MCM) manufacturing, as it greatly impacts yield, density, and reliability. To achieve low-cost manufacturing, modeling, optimization, and control of via formation are crucial. In this paper, a model-based supervisory control algorithm is developed and applied to reduce undesirable behavior resulting from various process disturbances. A series of designed experiments are used to characterize the via formation workcell (which consists of the spin coat, soft bake, expose, develop, cure, and plasma descum unit process steps). The output characteristics considered are film thickness, uniformity, film retention, and via yield. Sequential neural network process models are used for system identification, and hybrid genetic algorithms are applied to synthesize process recipes. Computer simulation results show excellent control of output response shift and drift, resulting in a reduction of process variation. The performance limits of the supervisory control system are investigated based on these simulation results. The control algorithm is verified experimentally, and the results show 82.6, 64.4, and 17.3% improvements in maintaining target via yield, film thickness, and film uniformity, respectively, as compared to open loop operation. Index Terms--Genetic algorithms, multichip modules, neural networks, supervisory control, via formation.