Mark Kiehlbauch

Neural Network Modeling for Advanced Process Control Using Production Data

The fabrication of integrated circuits involves many unit processes, some linear and some non-linear, and each with multiple inputs and outputs. These complexities suggest that benefits could be derived from the development and implementation of advanced process control tools and strategies. Empirical process models are one of these tools. In this research, sequential neural network models are developed to characterize critical steps in a fabrication process. The data used was collected from an industrial process. The data comes from experiments related to the processes under investigation, but not systematically designed to generate data for modeling. The models performed well, with an average prediction error of 3.3%, demonstrating the flexibility of the sequential neural network modeling process. Additionally, the models are used in a sensitivity analysis to study the output response to the various inputs. The methodologies presented are currently being ported to a similar manufacturing process with a larger database. Future work includes using the models for process optimization and as part of a model-based supervisory control system.

Etching mechanism of the single-step through-silicon-via dry etch using SF6/C4F8 chemistry

A single-step etching method using the SF6/C4F8 chemistry is developed in this study as an alternative through-silicon-via (TSV) etching approach of the traditional Bosch process to realize ultrasmooth and vertical TSV profiles. Experimental results show that there is a profile discontinuity, or a “transition,” on the TSV profile produced by the single-step etching method at high bias voltages and high SF6 flow rates. Comparison between the intensity of the species generated in a pure SF6 or a pure C4F8 plasma and in a SF6/C4F8 plasma is investigated for better understanding interactions between SF6 and C4F8. The densities of all positive ions are reduced in the SF6/C4F8 plasma compared to a pure SF6 plasma and a pure C4F8 plasma at the same partial pressure, indicating a change of plasma chemistry when SF6 and C4F8 fluxes are mixed. The formation mechanism of the transition is proposed as a chemistry discontinuity caused by large-angle ion sputtering at the top part of the sidewalls and the polymer accumulation at the bottom part of the sidewalls. The formation of the transition has found to have an effect of improving the sidewall smoothness below the position where it is formed. Parameter study has shown that a decreased bias voltage and a reduced SF6/C4F8 ratio can help to improve the sidewall smoothness and eliminate the transition on the TSV profiles.

Finite-element simulation models and experimental verification for through-silicon-via etching: Bosch process and single-step etching

In this study, time-dependent simulation models are established for both the Bosch process and single-step through-silicon-via (TSV) etching using SF6 and C4F8 chemistry by employing a finite-element-method method. The simulation models take into account the thermal etching of F radicals, ion-enhanced etching, neutral deposition and ion-enhanced deposition mechanisms, as well as the angular dependence of the ion sputtering with aspect to a surface element. Comparison between the simulation results and experiments suggests that consideration of two ion fluxes (high-energy and low-energy) is critical for matching the simulation etch profile with the experiments. It is found that the underlying reason for the transition formed on the TSVs using the single-step etching originates from the difference of the ion angular distributions of etching species and depositing species. The Bosch process model successfully predicted profile details, such as the top scallops of the TSV profile, and the model established for single-step etching can be used to predict the transition position shown on the sidewalls. The simulation models can be used to study the individual effects of low-energy ions and the high-energy ions in the etching and passivation mechanisms for TSV etching in both Bosch process and single-step etching techniques.

Neural network modeling of fabrication yield using manufacturing data

This paper describes the creation of artificial neural network models using production line data and illustrates their usefulness for process control in semiconductor manufacturing. Three artificial neural network models are created. The first models a high aspect ratio etch process. The other two are created to predict yield metrics from inline critical dimension (CD) measurements. One model predicts the number of faults on a die, and the other predicts the probability of die failure at probe. The high aspect ratio etch model has an average prediction error of 3.9%. The average prediction error for the number of faults on a die is 14.9%, and the average prediction error for probability of die failure at probe is 21.8%. A sensitivity analysis is performed on each model to illustrate how they can be used to judge the relative impact of each input.