Overlay Error Measurement Method Based on Through-Focus Scanning Optical Microscopy
Objective Over the past decade,integrated circuit(IC)technology has experienced remarkable advancements through the use of complex three-dimensional(3D)device structures,including new materials,patterning techniques,and processes that provide higher device performance with reduced feature sizes.These nanoscale 3D structures present significant challenges to the field of measurement.Lithography technology plays an important role in IC manufacturing,and its quality directly affects the yield of products.As an important factor affecting the quality of lithography,the precision requirements of overlay error have become increasingly stringent with the continuous breakthrough of the IC manufacturing process and the continuous reduction of advanced node sizes.According to the rule of thumb,the overlay accuracy should be better than 20％-30％of the critical dimension(CD).Currently,the measurement methods of overlay error are mainly divided into two categories:image-based overlay(IBO)and diffraction-based overlay(DBO).The IBO method is limited by the resolution of the optical microscope with the continuous breakthrough of the node,and the focal length and the laser wavelength need to be adjusted to enhance image contrast.There is no research on the traceability of the DBO method,which may result in large measurement errors.Through-focus scanning optical microscope(TSOM)is a fast,non-destructive,and highly reliable measurement technique.In order to realize the rapid non-destructive detection of the overlay error,a novel method for detecting the overlay error using the TSOM was proposed and explored in detail.This innovative approach aims to enhance the accuracy and efficiency of overlay error measurements,ultimately contributing to the advancement of IC technology.Methods The sample was placed on a microscope sample stage,which was driven by a piezoelectric transducer(PZT)to facilitate scanning along the Z-axis and through the focus point.During the scanning process,a series of sample patterns at different focus positions were captured by a charge-coupled device(CCD)camera.These images were then stacked based on their spatial positions to obtain a TSOM 3D light field.By intercepting the 3D light field along the Z-axis direction,a TSOM map containing the structural information of the sample was generated.After the TSOM map was processed,the train set and test set were established,and a convolutional neural network(CNN)model was constructed.The mean square error(MSE)loss function and adaptive moment estimation(Adam)optimizer were used to evaluate the prediction performance of the model through the test set.If the evaluation results did not meet the desired criteria,it was necessary to retrain the hyperparameters,such as the optimizer parameters and the number of convolutional network layers,by changing the activation function to determine the final model parameters.The information of the parameters to be tested was extracted by deep learning model prediction.Results and Discussions This method enables accurate prediction of overlay errors.For the model trained by different offset samples,the predicted result curve closely resembles the true value curve(Fig.6).When the sample offset interval is 200 nm,the mean absolute error(MAE)and root mean square error(RMSE)values are 4.2 nm and 5.3 nm,respectively.As the offset interval decreases,both the MAE and RMSE values decrease linearly(Fig.7).When the offset interval is reduced to 20 nm,the MAE and RMSE values of the offset prediction results are further reduced to 0.05 nm and 0.12 nm,respectively.This can be explained by the fact that as the offset interval of training samples decreases,the interval of deep learning label value decreases,and measurement resolution is enhanced.The standard deviation(STD)values of the four groups of samples are all below 6 nm and show a significant downward trend with the decrease in the offset interval of samples,which indicates that the offset interval of samples will also affect the repeatability of the prediction results.When the offset interval is reduced to 20 nm,the STD values of the overlay error measurement results are better than 0.083 nm(Fig.8),and the corresponding repeatability accuracy(3σ)is 0.25 nm.By using a smaller interval of experimental samples,higher measurement accuracy can be achieved.Conclusions The measurement of Bar-in-Bar marking offset is realized based on TSOM combined with a deep learning model.Unlike the traditional IBO method,we utilize an optical microscope to capture a series of images at different focal points,generating a TSOM atlas.A convolutional neural network model is then established for training and verification,which saves measurement simulation time and enables the regression prediction of the overlay marking offset.The sample with an offset interval of 20 nm is used for model training.The measurement accuracy is superior to 0.1 nm,and the repeatability accuracy(3σ)is superior to 0.25 nm.The experimental results show that this method is capable of measuring sub-nanometer overlay errors and is suitable for various types of overlay markings.In addition,it has a simple structure and low cost,which serves as a novel measurement method for overlay error measurement.
NETL
NSTL
万方数据
