Modeling and optimization of integrated circuit production process based on neural network

The semiconductor production process is a very complicated production process, during which many unpredictable factors are at work. Taking environmental factors as an example, the factors that affect the yield are air cleanliness, temperature, humidity and noise. Through the analysis of the process parameters of the previous semiconductor production line, it can be seen that the changes of the semiconductor process parameters show a strong nonlinearity and uncertainty. It is difficult to obtain an accurate model using the traditional mathematical regression method for this change process.

Artificial neural networks have excellent performance for identifying and approximating complex nonlinear systems, and have been widely and successfully applied in the field of control. However, although many research institutions at home and abroad have tried to use artificial neural networks for modeling optimization and monitoring of a specific processing step in the semiconductor manufacturing process, these methods have not been used by semiconductor manufacturers in the actual integrated circuit industrial production process. Used in the process, especially relying on the neural network method to successfully optimize the entire process of integrated circuit chip production, no matter in production practice or scientific experiment. Mod and optimization ask.

The neural network method is used to build a model for the semiconductor chip production line, determine the mapping relationship between the process parameters and the production rate on the production line, and create a multi-dimensional function curve with the art parameters as the input and the product rate as the output.

Search the above-mentioned multi-dimensional function surface, search for the best advantage with the highest yield, optimize the process parameter specification with the best parameters, practice it in the actual production process, and repeat it until the purpose of improving the yield is achieved.

Production practice 1 Ming. According to the optimization suggestions made by Shenling network analysis, the yield of the semi-conducting production line has been effectively improved.

2 The process of the half-good body production process constitutes the telephone voice network circuit chip produced by Huajing Company. It is a bipolar analog circuit, and its production process is quite complicated. Phosphorus re-expansion sub-lithography and sub-etching to say 1 isolation and re-expansion to remove the oxide film thin from the oxide-based lithography test strip width, boron into the upper-based anti-base diffusion photo-moxibustion test strip width and into the glue 2. silicon, expand 1 target Fire time lithography 1 time corrosion etching emitter area pre-deposition, 5 coating over emitter area re-diffusion capacitance corrosion galvanized silicon deposition silicon nitride photo-hai time lithography time corrosion sputtering wrong time lithography time corrosion 51 check alloy passivation Passivated Guanghai passivation corrosion alloy test test.

3 Modeling and Optimization 3.1 Static Modeling and Structural Design of Neural Networks In the process of analyzing and optimizing the production process, the suitability of the neural network algorithm used in modeling is closely related to whether the work can be carried out smoothly. In view of the fact that the production process of semiconductor chips is a complex nonlinear process, and there are always certain differences between the measured Ding Yi parameters and the actual situation, the neural network used to model the production process must have the following characteristics: Fault tolerance.

Conditions 1 and 2 are not difficult to meet, and currently the most widely used feedforward network 8 network has the above characteristics. A typical 8-network optimizes network parameters according to the direction of the negative gradient, and the excitation function usually uses the 53 function. In addition to this, in order to conduct the search for the best advantage after modeling and to ensure the physical significance of the optimal parameters, the neural network algorithm should also have the following feature mapping function surface. The best advantage should be close to the actual process parameter values.

For the above conditions, we use a feed-forward neural network, and take the Gaussian function with true local response characteristics as the excitation function of the hidden layer neurons of the network. The feed-forward network of this structure is called the radial basis function ruler 6 dipper. The output neurons of the network are still used, and the function is used as the excitation function. Similar to the 6 network, the training algorithm of this network is essentially a gradient descent algorithm based on the minimum mean square error criterion, and optimizes the network parameters according to the direction of the negative gradient. In the production process of integrated circuits in the network, the test data of the feed-forward network structure collected from the production line must contain certain noise. In fact, due to technical or management reasons, the influence of noise in some samples is very prominent, which has made the samples seriously deviate from the actual situation. This situation has brought great difficulties to the learning of neural networks. How to solve the problem of learning noisy samples in feedforward networks is the key to the success or failure of modeling integrated circuit production processes.

When learning using neural networks, the learning of the training sample set by the plastic neural network is to learn the regular information it contains, while the irregular information contained in the samples, that is, noise, is discarded. If the training error of the neural network for the sample set containing noise approaches, it means that the neural network has not only learned the regular information contained in the sample, but also learned the noise in the sample. This shows that when learning against noisy samples, the learning effect of the neural network should only be based on the training error of the network. Therefore, we used an independent test sample set to verify the learning effect of the network.

If we intuitively describe the influence of noise in the sample on the input and output function surface of the neural network, it can be said that the neural network learning of the noise contained in the sample causes many burrs on the surface, and the smoothness of the function surface is reduced. We believe that if the Greek plastic neural network only learns the regularity information contained in the sample and does not learn the smoothness of the input and output function surfaces of the feedforward network in the sample, the smoothness of the surface is related to the size of the network. The greater the number of network parameters, the more complex the shape of the function surface and the lower the smoothness of the surface. Therefore, in the design of the network, we limit the size of the network on the premise of ensuring the network learning ability.

The steps to determine the size of the network are as follows: the samples collected from the production line are divided into two subsets of training samples and test samples; use a small-scale neural network such as a network with only one hidden layer neuron, and train the network according to the training sample set until the end Converge; then, input the test sample into the neural network, and calculate the error of the output of the network and the ideal output as the test error; select different network initial states to repeat the learning and testing of the network many times, and calculate the expected value of the learning error and test error. Network neurons in the hidden layer, re-learning and testing, so gradually expand the network size until the network size can no longer bring about a significant reduction in learning error; the network learning effect is the best when the expected value of the verification error of the network is determined to be the smallest, choose The scale of the network at this time is modeled by the production line.

3.2 The dynamic rolling semiconductor production process of neural network model is a very complicated production process, in this process, there are many unpredictable micro factors in play. The instability of the production process itself, as well as the influence of various subjective and objective factors when collecting process parameters, make the training samples of the neural network model show considerable chance, which may not reflect the laws of chip production, and some of them There may also be major discrepancies with the actual situation. In this case, it is very difficult to use traditional neural network training methods to establish a model that conforms to semiconductor production laws.

In response to the specific situation of semiconductor production, we proposed a dynamic rolling method of neural network model to divide the collected process parameter samples into training samples and test samples. After training according to the training samples, the neural network was established and the test samples were input into the neural network for inspection. The output of the Jingjing network is true; if the output of the network deviates from the actual situation, remove some samples from the training sample set and retrain; compare the results of the two trainings, if the retraining results are more than the first training The result is closer to the actual situation, it is determined that the removed samples cannot reflect the regularity of semiconductor chip production, and the remaining training samples after the removal constitute a new training sample set; repeat the above process until the training sample set can no longer be removed, according to this time The training sample set is trained to establish the final neural network model.

3.3 The dynamic search algorithm for the mapping function surface After using a multi-layer feedforward network to construct the mapping function surface, the best dynamic search algorithm for the surface is as follows. The sample with the highest yield in the sample set is the starting point for the search 4; for each process of the sample The parameters are perturbed to construct a new sample, and the neural network model is input. The resulting network output is called a new sample, and the pseudo-yield rate; observe the impact of various process parameters on the simulated yield rate after perturbation, and determine which process parameters are Sensitive factors affecting yield. Find the optimization direction of the process parameters that increase the yield rate. Correct the process parameters in the direction of the increase of the yield rate, get a new sample, and then use the new sample as the starting point to find the optimization direction again; if iterate repeatedly, no matter how small the sample is Disturbance, the simulated finished product rate of the new sample will all sway at the mouth 1, Chungang meets 8 goods Huo Ru 181 ±, which. 1 beer play 3 generation 361.1 4 practical application The production of the voice network circuit chip of the phone is the analysis object.

The testing cost of semiconductor chips is relatively high and it takes a long time. In order not to increase the workload of the production line, and at the same time, we can use the existing historical data of the process parameters as much as possible. We use the conventional detection parameters in the parts list on the production line as the input of the neural network, and the finished product rate is the network output. The bipolar production line has provided us with 3 pieces of pick-up slips in 1997, 46 pieces of pick-up slips at the beginning of 1998. Each piece of slip records the process parameters and yield of each batch of chips. The number of good products should be about 20,000.

4.1 Selection of process parameters and pretreatment of semiconductor chip production is a very complicated process, and the process parameters collected on the parts list are also quite complicated. It is necessary and beneficial to select and pre-process the process parameters.

We deleted some parameters that could not be optimized, such as the start date and mid-test date; it is difficult to quantify the thickness and isolate 0 thickness. Other deleted parameters are isolation and re-diffusion 8 times of photo-oxidized silicon oxide, thickness and coating thickness of phosphorous and arsenic.

Some of the retained parameters have also been pre-processed, such as the resistivity and thickness of the epitaxial layer, which were originally based on their upper and lower limits, but are now levied on the yield rate. Finally, all parameters have been normalized.

4.2 The most superior search results and process parameter specification optimization of the mapping function surface After parameter selection and preprocessing, the final process parameters obtained as network input count 21 items, as new, and the network output is only 1. Yield. This determines that the feedforward network should have 21 input parameters. The original parameter specification value is searched for. The optimal value of the parameter is epitaxial resistivity. Sub-oxidized thickness. Deep phosphorus predeposition resistivity. Thin oxygen oxidation thickness. The lithography test strip has a wide-area pre-deposited resistivity silicon nitride deposition thickness emitter area electrical parameter, alloy characteristic check with a value of 1, V monitors the resistance element, and one output passes through nine. Use the independent sample set test method in 3.1. We decided that the number of neurons in the hidden layer of the network is 4. First, we use the sample set composed of 3 sets of process parameter data in 1997 to train the neural network for the first time, and use the network state after the initial training as the initial state of the network. The first batch of 46 sets of process parameters was used as the training sample set of the neural network to retrain the neural network to complete the dynamic correction of the neural network.

Then we conducted the most superior search on the mapping function surface of the neural network model, and obtained the ideal value of the process parameters.

4.3 The final optimization proposal is divided and optimized by the neural network to obtain the cardinal value of the widely optimized parameter specification. However, after the optimization results are obtained, if the specification values ​​of all 21 process parameters are improved at the same time, it will be difficult to determine which parameter improvement has a practical effect, which is detrimental to information feedback and dynamic correction of the model. In fact, there is no need to improve all the process parameters, because the control center value of some processes is in good agreement with the optimized center value. For example, the center value of the epitaxial thickness and the optimization result are almost the same, indicating that the original process specifications are quite reasonable. Therefore, we selected three controllable parameters with a large gap between the optimal value of the parameter and the original value of the original process specification, and proposed optimization suggestions such as 2.

Parameter Name Parameter Optimization Center Value Optimization Parameter Range Unit Deep Phosphorus Predeposition Resistivity Monitoring Resistor K 44 Optimization Results Practicability Verification Before making optimization suggestions, the average yield of 133 batches of telephone network voice circuit chips produced by Huajing Company 51. View. After making optimization suggestions, the 26 batches of chips produced according to the new process specifications are brand new. The number of good products for each batch of chips is about 20,000. The results of the practice show that the modeling and optimization based on the neural network are reasonable and have a practical and effective effect on improving the yield.

5 Conclusions and Discussion This project takes the optimization of the main processes in the large-scale integrated circuit industrial production as the specific research object. Its research results not only have scientific significance for the models and algorithms used by the neural network for optimization, but also have a larger industry. The industrial production also has potential application value.

Acknowledgements Jiang Cheng thanked Ping Jing for its strong support and the technicians of the bipolar production line for their active cooperation.

Wang Xiangdong 197, born annually, Ph.D., is currently a postdoctoral fellow at the Institute of Automation, Chinese Academy of Sciences. His current research interests include neural network model Wang Shoujue born in 1925, graduated from Tongji University, academician of the Chinese Academy of Sciences, deputy chairman of the Chinese Institute of Electronics, and editor-in-chief of the Journal of Electronics. He is the founder of the Chinese semiconductor discipline, and is currently engaged in the research on the hardware and application of semiconductor ultra-high-speed circuits and neural network algorithm models.

In 2001, the National Academic Annual Conference of the Chinese Society of Artificial Intelligence, 19 essays informed the 1990s that intelligence theory and technology have achieved unprecedented progress, and the application of intelligent technology has shown unprecedented prosperity. The intersection of field interest-bearing science and biology, the field of intelligence must become the focus of the entire science and technology.

In order to summarize the achievements made in the field of intelligence in China, exchange the latest progress, and meet the greater challenges in the field of intelligence in the 21st century. The Society is scheduled to hold the Ninth National Academic Annual Conference in Beijing on October 810, 2001. Welcome academicians from universities, scientific research institutes and enterprises, as well as doctoral and master students to participate actively in the annual meeting of the first academic event in the field of intelligence in China in the new century. New directors, all professional committees of the board of directors and colleagues who are concerned about the development of the society will attend the meeting.

The main areas of intelligence and network essays for this academic annual meeting include but are not limited to basic theoretical research, intelligence, physiology, basic intelligence, cognitive science, discrete mathematics and logic theory, intelligent research, philosophy and methodology, and basic technology research. Knowledge theory search theory method new model and algorithm of neural network network intelligence distributed intelligence and swarm intelligence machine learning human, life