Neural Nets Spruce Up Run-to-Run Control

Like the human brain, neural networks find order in chaos. First pioneered in the 1950s by Bernard Widrow of Stanford University, neural networks work by creating connections between processing elements, the computer equivalent of neurons, rather than digital computation that manipulates zeros and ones. The neural net's structured interconnection of computational nodes — neurons — contributes to the parallel computation of the network through a learning process. Knowledge is stored in inter-neuron connection strengths, or weights (Figure ), and the weighted sum of the inputs can be filtered by a neuron activation function, which is part of what differentiates neural networks from other regression systems.

Because of their ability to learn complex nonlinear mapping, neural networks have been applied to semiconductor process modeling, optimization and control applications. In modular neural networks, which are under investigation for fault detection and yield optimization, two or more subsystems operate on distinct inputs, and their outputs are mediated by an integrator, but there is no feedback to the modules. Diagnosis of reactive ion etching (RIE) processes is one application that can benefit from modular neural network tools by reducing misprocessing and increasing productivity.

A group of researchers from the Georgia Institute of Technology(Atlanta) and Myunghi University (Yongin, Korea) recently showed that modular neural networks can be used as an effective fault diagnostic tool for RIE processes, with the potential for reducing wafer misprocessing, improving process control and increasing yield. Gary May and coworkers presented their findings at the latest Advanced Semiconductor Manufacturing Conference, held in Boston in May.

The neural net (left) is a structured interconnection of computational nodes (neurons) that contribute to parallel computation in a manner similar to the human brain. Acquired knowledge is stored in interconnection strength (weights). The modular NN has local expert networks that operate on distinct inputs without communcating with each other for faster learning speed and decomposition of a complex function into simpler functions.

The group used in situ optical emission spectroscopy (OES), which is commonly used in plasma processing for endpoint detection, process modeling and plasma diagnostics. RIE of SiLK low-k dielectric from Dow Chemical (Midland, Mich.) was performed using a Plasma Therm 700 etcher in 25 baseline runs and 25 faulty runs, with intentionally modified RF power. RF settings of 220-280 W in 10 W increments were used, and one data set was run with fluctuating RF power. OES data was taken every 10 seconds. From the OES data, 40 significant wavelengths had to be selected (among >2000), and out of 49 runs, the modular neural network was trained with data from 39 randomly selected runs and 10 runs were used to verify the model. Model accuracy (as measured by root-mean-squared error) was 0.56% in the training data set and 2.81% in the testing data set.

To apply modular neural networks to fault diagnosis, the researchers took four consecutive baseline runs and stepwise excursions by run. An off-line simulation tested trained modular neural networks to detect excursions as soon as a process run was completed. With both increasing and decreasing RF power, the modular neural network detected process excursions as low as 5% from baseline. It was able to rapidly track the trends of a shifted process.

The researchers noted that future work would focus on run-to-run prognosis of RIE using actual in situ metrology and tool data. They will use each expert network to forecast the results of each process run in a manner similar to time series modeling.

For additional information on yield management, go to www.semiconductor.net/yield