Circuit Lithography Hotspot Detection Technology Based On Deep Learning

With the rapid development of VLSI technology,the feature size of transistors has become smaller and smaller,and circuit layouts have become more and more complex,which has brought great challenges to circuit lithography.The wavelength of the current light has reached the 193 nm limit,which is much larger than the existing feature size of the transistor.When we burn the standard circuit design layout onto the silicon wafer,some defect patterns,which are called hotspots,will be generated on the silicon wafer by the diffraction of light.These hotspots layout are very likely to cause an open circuit or short circuit,which could burn the circuit,result in a reduction of chip yield and huge economic losses.Although the chip manufacturing industry has proposed a variety of production design techniques,such as design rule check,optical proximity correction,multi-pattern lithography,etc.,hotspots layout cannot be eliminated.Therefore,before the circuit layout is transferred to the silicon chip,the entire circuit layout must be tested to find out the hotspots layout and correct them.The traditional detection methods include simulation,pattern matching,and machine learning.However,as the scale of integrated circuit becomes larger and larger,and th e design circuit becomes more and more complex,the traditional detection methods are no longer applicable.In this paper,a network model based on FFT feature extraction and deep learning is proposed,which can effectively solve the problems of large circ uit size and complex design,and it also has good scalability to meet the needs of future development.F rom the perspective of detection performance,it is also greatly improved compared with traditional methods.The main work of this paper is as follows:(1)A feature extraction method based on Fast Fourier Transform is proposed,which converts the circuit layout from the spatial domain to the frequency domain,and concentrates the feature information of the layout on the four corners of the frequency domain,which not only facilitates feature extraction,but also facilitates feature extraction.It can effectively filter the useless information of the layout,retain the key feature information of the circuit layout to the maximum extent,and improve the detection accuracy and training efficiency of the training model.(2)A batch learning algorithm based on preference is proposed,which uses batch training methods for unbalanced training samples,ensuring that the distribution of different categories of samples in each batch of training samples is balanced.At the same time,in order to reduce the false positive rate of detection models,we give a certain weight penalty for misjudgment of different types of samples.And in order to make the detection model reaching a better balance between accuracy and false positive rate,we dynamically embed this preferred learning algorithm into the training process.(3)We have implemented our hotspot detection model based on deep learning,the experimental results show that our proposed hotspot detection model not only improves the detection accuracy,but also reduces the false positive rate by 40%-50%.