Byungwhan Kim

An optimal neural network plasma model: a case study

Abstract Artificial neural networks, particularly backpropagation neural network (BPNN), have recently been applied to model various plasma processes. Developing BPNN model, however, is complicated by the presence of several adjustable factors whose optimal values are initially unknown. These may include initial weight distribution, hidden neurons, gradient of neuron activation function, and training tolerance. A methodology is presented to optimize various factor effects, which was accomplished by implementing genetic algorithm (GA) on the best models. Particular emphasis was placed on a qualitative measure of initial weight distribution, whose magnitude and directionality were varied. Interactions between factors were examined by means of a 2 4 factorial experiment. Parametric effect analysis revealed the dissimilarity between the best and average prediction characteristics. Both gradient and initial weight distribution exerted a conflicting effect on both average and best performances. GA-optimized models exhibited about 20% improvement over the experimentally chosen best models. Further improvement of more than 30% was achieved with respect to statistical response surface models. Plasma modeled is an inductively coupled plasma, whose experimental data were collected with Langmuir probe from an etch equipment capable of processing 200-mm wafers.

Modelling of plasma etching using a generalized regression neural network

Plasma etching was modelled by using a generalized regression neural network (GRNN). The etching process was characterized with a statistical experimental design. Three etch responses were modelled, which include two etch rates of aluminium and silica and etching profile. GRNN prediction ability was optimized as a function of training factor. Three types of models were constructed depending on the type of prepared data. Type I model corresponds to the model constructed with the original, non-classified data. Type II and III models were built for the classified data without and with the control of data interface, respectively. Compared to type I models, type II models for two etch rates demonstrated more than 25% improvement. By the control of data interface, type III models exhibited more than 15% improvement over type II models. Classification-based models in conjunction with data control thus illustrated much improved prediction of GRNN over those for non-classified models.

Modeling SiC etching in C2F6/O2 inductively coupled plasma using neural networks

Silicon carbide (SiC) has been etched in a C2F6/O2 inductively coupled plasma and modeled using neural networks. A 25 full factorial experiment was used to characterize the relationships between input process factors and etch response. The factors that were varied include source power, bias power, pressure, O2 fraction, and gap between the chuck holder and coil antenna. Neural networks were trained on the resultant 32 experiments and then tested on 18 additional experiments to evaluate prediction accuracy. Due to little variations in etch anisotropy, etch rate was only modeled and its root-mean-squared prediction error was 23.9 nm/min. Etch rate was found to be a strong function of source power. Increasing etch rate with pressure may partly be attributed to increased ion density and ion energy. Placing the chuck holder closer to the source antenna coil increased the etch rate. At higher bias powers, increasing the O2 fraction resulted in a crossover. This crossover seems to be weakened significantly with ...

Qualitative modeling of silica plasma etching using neural network

An etching of silica thin film is qualitatively modeled by using a neural network. The process was characterized by a 23 full factorial experiment plus one center point, in which the experimental factors and ranges include 100–800 W radio-frequency source power, 100–400 W bias power and gas flow rate ratio CHF3/CF4. The gas flow rate ratio varied from 0.2 to 5.0. The backpropagation neural network (BPNN) was trained on nine experiments and tested on six experiments, not pertaining to the original training data. The prediction ability of the BPNN was optimized as a function of the training parameters. Prediction errors are 180 A/min and 1.33, for the etch rate and anisotropy models, respectively. Physical etch mechanisms were estimated from the three-dimensional plots generated from the optimized models. Predicted response surfaces were consistent with experimentally measured etch data. The dc bias was correlated to the etch responses to evaluate its contribution. Both the source power (plasma density) and...

Use of neural networks to model low-temperature tungsten etch characteristics in high density SF6 plasma

A tungsten (W) etch process in a SF6 helicon plasma has been modeled using neural networks. The process was characterized by a 24−1 fractional factorial experimental design. The design factors that were varied include source power, bias power, chuck holder temperature, and SF6 flow rate. The responses modeled include etch rate, selectivity, anisotropy, and nonuniformity. With optical emission spectroscopy, spectra of radical F intensity were collected to investigate the etch mechanisms. High prediction accuracy was achieved in the etch models. The root mean-square prediction errors were 249 A/min, and 0.41, 0.16 and 0.83 for the etch rate, selectivity, anisotropy, and uniformity models, respectively. While exerting little impact on the selectivity, the temperature greatly affected the etch rate and anisotropy. In particular, the etch nonuniformity was improved at low temperature. Both the selectivity and nonuniformity were predominantly determined by the bias power. The anisotropy was inversely related to...

Characterizing metal-masked silica etch process in a CHF3/CF4 inductively coupled plasma

A silica etch process conducted using a CHF3/CF4 inductively coupled plasma is characterized. This was accomplished by employing a statistical experimental design in conjunction with neural network process models. As a mask layer, patterned AlSi (1%) metal was used. Parameters varied in the design includes source power, bias power, and gas ratio. Besides those conventional etch responses including etch rate, selectivity, and profile, sidewall roughness of the etched pattern is first modeled. Etch rate and sidewall roughness were found to be predominantly influenced by source power with a trade-off between them. Bias power significantly affected selectivity while controlling a trade-off against etch rate. A decrease in profile angle with increasing bias power was attributed to AlSi (1%) film expansion induced by ion bombardment effects. As gas ratio was varied, profile angle remained almost constant due to nearly same chemical reaction of its plasma on the silica surface. The roughness was little affected ...

Prediction of profile surface roughness in CHF3/CF4 plasma using neural network

Abstract Using a neural network, a profile roughness of plasma etching is characterized. The etching was conducted in a CHF3/CF4 inductively coupled plasma. The etch process was characterized by a 23 full factorial experiment. The process parameters that were varied in the design include radio frequency source and bias powers, and gas ratio. Relationships between the parameters and profile roughness were captured by training neural network with eight experiments plus one center experiment. Model appropriateness was tested with six experiments not pertaining to the training data. Model prediction capability was optimized by means of a genetic algorithm (GA). Compared to a conventional model, GA-optimized model demonstrated a drastic improvement of about 54% in predicting profile roughness. From the optimized model, several plots were generated to examine parameter effects on the profile roughness. Increasing the source power (or bias power) under high bias power (or source power) increased the profile roughness. More significant effect of the bias power was revealed. The profile roughness decreased with increasing the gas ratio was strongly correlated to the dc bias. The little variation in the profile roughness was ascribed to chamber plasma condition maintained at relatively low dc bias.

Surface roughness of silicon carbide etched in a C2F6/O2 inductively coupled plasma

Investigation of the surface roughness of etched SiC films is of great importance when attempting to improve electrical properties of SiC active devices. The roughness of SiC etched in a C2F6/O2 inductively coupled plasma has been examined as a function of process parameters. Experimental ranges of parameters that were varied include 600–900 W source power, 50–150 W bias power, 4–16 mTorr pressure, 0%–80% O2 fraction, and 6–12 cm gap between the wafer and plasma source, respectively. Surface roughness was characterized using atomic force microscopy. dc bias voltage was correlated to the roughness. At medium parameter levels the roughness was reduced with an increase in source power while the roughness was increased with increasing bias power. With variations in pressure or O2 fraction, the roughness varied nonlinearly. dc bias was strongly correlated to the roughness for variations in source power and bias power. Interestingly, the bias was inversely related to the roughness for variations in O2 fraction ...

Relationships between etch rate and roughness of plasma etched surface

Relationships between etch rate and surface roughness are examined as a function of process factors, including source power, bias power, pressure, and O/sub 2/ fraction. Experimental factor ranges are 600-900 W source power, 50-150-W bias power, 4-16-mtorr pressure, and 0%-80%-O/sub 2/ fraction. Atomic force microscopy was used to quantify surface roughness of silicon carbide etched in a C/sub 2/F/sub 6/ inductively coupled plasma. The impact of ion energy was estimated by means of DC bias. The etch rate was inversely related to surface roughness as source power (plasma density) varied. The source power-induced DC bias was strongly correlated to surface roughness. For variations in bias power (ion bombardment energy), the etch rate was almost linearly correlated to both surface roughness and DC bias. For variations in pressure or O/sub 2/ fraction, the etch rate was related to surface roughness in a complex way.

Characterization of inductively coupled plasma using neural networks

Hemispherical inductively coupled plasma (HICP) in a chlorine (Cl/sub 2/) discharge is qualitatively characterized using neural networks. Plasma attributes collected with Langmuir probe from a HICP etch system include electron density, electron temperature, and plasma potential. Process factors that were varied in a 2/sup 4/ full-factorial experiment include RF power, bias power, pressure, and Cl/sub 2/ flow rate. Their experimental ranges are 700-900 W, 5-10 mtorr, 20-80 W, and 60-120 sccm, for source power, pressure, bias power, and Cl/sub 2/ flow rate, respectively. To validate models, eight experiments were additionally conducted. Root mean-squared prediction errors of optimized models are 0.288 (10/sup 11//cm/sup 3/), 0.301 (eV), and 0.520 (V), for electron density, electron temperature, and plasma potential, respectively. Model behaviors were in good agreement with experimental data and reports. For electron temperature and plasma potential, interaction effects between factors were observed to be highly complex, depending on the factors as well as on their levels. A close match was observed between the models of electron temperature and plasma potential.