Using A Neural Network For Evaluating Semiconductor Deterioration

Key takeaways

Advancements in semiconductor technology require enhanced precision in design and production processes.
Traditional information systems alone cannot provide complete real-time insights into semiconductor manufacturing variables.
Artificial intelligence, particularly neural networks, is being explored to predict machine wear and tear in semiconductor production.
A recent study from China demonstrated the effectiveness of a backpropagation neural network in evaluating semiconductor equipment deterioration.
Integrating AI with existing systems can improve the accuracy of predicting equipment status and enhance production efficiency.

We always talk about advances in computing technologies and hardware and the endless possibilities that it can bring to the future. Semiconductors that make up for powerful components in these computing devices, are also getting revamped greatly in terms of size and affordability. This means that higher precision is necessary in using machines designing for accurate, efficient semiconductors. For this to happen correctly, companies need to include elements such as state-of-the-art information systems for semiconductor processes. Though information systems provide timely data, it alone cannot be sufficient to give out the real-time status of a large number of variables involved in the production processes. With AI and sub-fields already taking the mantle in futuristic technologies, it is now being experimented in detecting machine parameters associated with semiconductor manufacturing in time. In this article, we will explore one such study that has used a neural network with backpropagation to analyse data from an information system and help predict a machine’s wear-and-tear status accurately.

The Experimental Setup

A group of researchers from China have come up with a study that predicts a semiconductor equipment deterioration. They use a backpropagation neural network model along with another method called grey relational analysis to analyse data coming from the information system. Fault detection and classification (FDC) is the information system used here. This system actually helps identify extreme parameters in a machine that is used for semiconductor processes. Keng-Chieh Yang, the lead researcher in the study, describes the experimental setup, “We use a semiconductor machine as the experiment tool to analyse the results in the manufacturing process. The return value of this machine is as a network training inputting data for the analysis. This study uses Novellus Vector Machine and its Remote Process Controller (RPC) function to collect the data. The data collection period is between April 2008 and December 2011.” MATLAB Neural Network Toolbox is used to analyse data coming from the machine. For the study, a number of required tasks are performed such as data preprocessing, network variables setting, hidden layer neuron determination, choosing input variables, network output results determining principles, and sensitivity analysis. As mentioned earlier, the grey relational analysis is also used to detect large changes in the machine parameters. Therefore, gas transmission pressure is considered as the core parameter as well as being the prime cause for changes in the semiconductor wafers of the Novellus Vector machine during a process.

Backpropagation Neural Network And Its Model

The researchers create the artificial neural network by considering all the relevant factors such as learning rate, learning trials and momentum correction coefficient. Since backpropagation mostly uses mean squared error(MSE) as a performance indicator for good measurement, the same is considered in this study. Keng-Chieh and associates use a learning rate of 0.1 to 0.3 for dynamic changes and momentum correction coefficient is set to 0.01 to 0.02. This is to accommodate frequent variations in the neural network. As the MSE decreases in a learning trial by less than 3,000 epochs, the network is said to have learnt the process completely. This means, the machine encounters an average of 1,000 to 3,000 learning instances. Apart from this, the number of hidden layers in the network is determined using the re-selection of previous variables obtained in the experiment. This is to achieve the forward selection technique properly in the network. Once the neural model is set, it is subjected to training along three instances:

One-month network
Three-month network
Six-month network

The researchers elucidate the model for training: “Using technology of the process and monitor recipe as units, each of them has its own training dataset and artificial neural backpropagation network model. The next step would be to analyse the characteristics of the data input. Specifically, the analysis should emphasise the column input and the column to see if it is global or partial. This means that it should be checked if the input space is mainly centred on specific areas and return its relationship with time.” The results obtained from training the neural network model shows satisfactory results and the performance was significantly better than without the network in the model and predicted extreme values in the semiconductor deterioration. The performance is measured in MSE and the correlation coefficient for all the periods. The results can be found here.

Conclusion

The study was restricted to only one machine. However, it has provided an avenue for future research in neural network predictions. If predictions are made more accurate, observing the effect of wear in semiconductors could greatly see perfection. Machine maintenance can also be minimised over time. On top of this, semiconductor materials could be gauged based on wear over time. Thus, neural network methods can provide an effective way to analyse the extreme properties of semiconductors.

Key takeaways

The Experimental Setup

Backpropagation Neural Network And Its Model

Conclusion

Stay ahead with AI insights.