Search

A new back propagation neural network (BPNN) model is presented to construct a plasma etch process. This is accomplished by optimizing multi-parameterized neuron gradients using genetic algorithm. The technique was evaluated with experimental data collected during the etching of silicon carbide films in a NF 3 / CH 4 inductively coupled plasma. The etch process was characterized by a 2 4 full factorial experiment plus one center point. The trained model with the resulting 17 experiments was tested with the test data consisted of 16 experiments. The etch outputs to model include etch rate, profile angle, dc bias, and surface roughness. Compared to conventional BPNN models, for all etch outputs, GA-BPNN models demonstrated improvements ranging between 26% and 83%. For another smaller size of data, the improvement was more conspicuous and it ranged between 26% and 59%. These results reveal that the proposed technique can contribute to building accurate plasma models. Moreover, the technique is general in that it can be applied to any other complex plasma data.

A low pressure etching of silicon carbide is qualitatively characterized by using a neural network. To construct a predictive model, the etch process was characterized by means of a 25 full factorial experiment. Experimental factors that were varied include radio frequency (rf) source power, bias power, pressure, O2 fraction, and gap between the plasma source and wafer. An additional 15 experiments were conducted to test the appropriateness of the trained model. An optimized etch rate model has a root mean-squared error of 12.78 nm/min. Model response surface behaviors were certified by actual measurements. Several noticeable features at lower pressure etching include a lower etch rate, inverse relationship between the source power level and the dc bias, and a smaller etch rate variation with the source power. The effect of the bias power on the etch rate or dc bias was affected little by the pressure level. Etch mechanisms for the gap variations were quite different depending on the bias powers. Several etching aspects useful for plasma control were revealed.

Uniformity of plasma etching has been conventionally examined only in terms of etch rate. A uniformity of etching profile surface is increasingly demanded to improve process quality. This is accomplished by applying a discrete wavelet transformation (DWT) to profile images obtained with a scanning electron microscopy. Applicability of DWT-based profile uniformity was evaluated with a tungsten etch experiment, conducted in an SF/sub 6/ helicon plasma. Its suitability was investigated as a function of process parameters and scale levels. Compared to a conventional metric, the wavelet-based one characterized more effectively the uniformity of profile variations. The proposed metric can be applied to any other plasma-processed surfaces.

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.

By controlling the proximity relative to plasma source, silicon carbide (SiC) has been etched in a C2F6 inductively coupled plasma. This was accomplished by adjusting the gap between the wafer substrate and the plasma source. Gap effects on SiC etching were extensively examined as a function of process factors as well as at particular two conditions, high source power (plasma density) and high bias power (ion bombardment). Decreasing the gap increased the etch rates for all conditions mainly due to the increased plasma density. The DC bias induced with the gap variation played a little role in affecting the etch rate. Gap effect was more enhanced as the wafer electrode was powered with relatively larger DC bias. Factor effects on the etch rates were quite different depending on the level of plasma density or ion bombardment. Decreasing the gap resulted in a microtrench at the base of the sidewall, which was little affected by the DC bias.

Summary form only given, as follows. Profiles of plasma etching have conventionally been characterized by approximating the slope with an angle or anisotropy. This is critically limited in that detailed variations on the profile surface are inevitably neglected. In current high density plasma etching, this becomes more serious since unexpected microfeatures such as bowing or microtrenching are frequently formed along the profile surface.

A prediction model for surface roughness was constructed using a neural network and atomic force microscopy. The silicon carbide etch process was characterized by a 25 full factorial experiment. The experimental ranges of process parameters were 600–900W source power, 50–150W bias power, 4–16mTorr pressure, 0–80% O2 percentage, and 6–12cm gap. The model factors were optimized by means of a genetic algorithm. The optimized model had a root mean-squared error of 0.11nm. From the model, various plots were predicted while being supported by actual measurements. The dc bias induced by each process parameter was correlated to the surface roughness. Increasing the bias power increased the surface roughness. In contrast, the surface roughness decreased as the dc bias was larger than about 600V. The surface roughness was strongly correlated to the source power-induced dc bias only at low bias powers. The pressure effect was clear only as the dc bias was maintained at 480V. For the variations in the O2 percentage, ...