CN-121995709-A - Method and device for computing lithography modeling and hot spot detection based on self-supervised learning

CN121995709ACN 121995709 ACN121995709 ACN 121995709ACN-121995709-A

Abstract

The invention relates to a self-supervised learning-based computational lithography modeling and hot spot detection method and device, comprising the steps of constructing a neural network model comprising a layout feature encoder, a geometric reconstruction branch and a lithography space image prediction branch, wherein a differentiable optical simulation layer is built in the lithography space image prediction branch so as to integrate into physical priori; based on the label-free mask layout data, masking the input layout, driving the model to simultaneously execute geometric reconstruction and pre-training of photoetching space image prediction, processing the original mask-free layout through the differentiable optical simulation layer to generate a pseudo-true value, providing a supervision signal for photoetching space image prediction, and performing fine adjustment by using a small amount of label-free data on the basis of the pre-training model to realize hot spot detection. By integrating the differentiable optical simulation layer into the self-supervision learning framework, the model can learn optical rules from massive non-tag data, so that dependence on labeling data is reduced, and the hot spot detection capability and accuracy of unknown graphics are improved.

Inventors

QIAO YI
NIU SHIWEN
WANG SHAOHUA
ZHANG GUIYU
YU HANYUAN
YU YUEYUE

Assignees

天津国瑞微电子科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260228

Claims (10)

1. The self-supervised learning-based computational lithography modeling and hot spot detection method is characterized by comprising the following steps of: Step one, an unmarked mask layout data set is constructed, wherein the unmarked mask layout data set comprises a plurality of mask layout data fragments extracted from a chip design layout database and divided into a plurality of mask layout data fragments; Step two, constructing a self-supervision pre-training neural network model based on physical perception, wherein the model comprises a layout feature encoder, a geometric reconstruction branch and a photoetching space image prediction branch embedded with a differentiable optical simulation layer; performing self-supervision pre-training, namely performing mask operation on the mask layout data fragment and inputting the mask layout data fragment into the layout feature encoder to obtain potential features, calculating geometric reconstruction loss based on the potential features by utilizing the geometric reconstruction branch, predicting a lithography aerial image based on the potential features by utilizing the lithography aerial image prediction branch, and calculating optical physical loss between the lithography aerial image and a pseudo-true value lithography aerial image generated based on an original mask layout by utilizing the differentiable optical simulation layer; step four, updating network parameters of the self-supervision pre-training neural network model according to the weighted sum of the geometric reconstruction loss and the optical physical loss; Fifthly, performing fine adjustment on the layout feature encoder after self-supervision pre-training by using sample data with hot spot labels to obtain a hot spot detection model; And step six, performing hot spot detection on the full-chip mask layout by applying the hot spot detection model.
2. The method of claim 1, wherein the layout feature encoder employs a backbone network of a attention-based transform architecture.
3. The method according to claim 1, wherein the step of masking the mask layout data segment in the step three specifically comprises: Calculating the gradient of the photoetching aerial image intensity relative to each pixel point of the mask layout data fragment by using the differentiable optical simulation layer, and generating a sensitivity map, wherein the sensitivity map quantifies the influence degree of the change of each pixel point on the mask on the photoetching imaging result; And setting probability distribution of masking operation according to the sensitivity map so that the higher numerical value area on the sensitivity map is higher in probability of being masked.
4. The method of claim 1, wherein the differentiable optical simulation layer is constructed based on a coherent system summation model and comprises a set of optical convolution kernels that are decomposed by the coherent system summation model based on a numerical aperture of a lithography machine, an exposure wavelength, and a light source shape.
5. The method of claim 1, further comprising, after the fourth step and before the fifth step, a step of performing optical equivalence-based contrast learning enhancement, the step comprising: applying optical equivalent disturbance to the mask layout data fragment to generate a disturbed mask layout data fragment, wherein the optical equivalent disturbance refers to adjusting the position or the size of a sub-resolution auxiliary graph in the mask layout data fragment on the premise of not violating the constraint of a design rule; taking the original mask layout data fragment and the perturbed mask layout data fragment as a pair of positive samples, and respectively inputting the positive samples into the layout feature encoder to obtain a pair of feature vectors; A contrast penalty is calculated and network parameters of the layout feature encoder are updated by decreasing the distance in feature space of the pair of feature vectors of the alignment sample while increasing the distance in feature space of the feature representations of the different mask layout data segments.
6. The method according to claim 5, wherein the adjustment of the position or size of the sub-resolution auxiliary pattern is performed, in particular by random fine tuning in the range of ±1nm to 5 nm.
7. The method according to claim 1, wherein the step of fine tuning using the sample data with the hot spot tag in the step five specifically comprises: removing the geometric reconstruction branch and the photoetching space image prediction branch, and accessing a classification module at the output end of the layout feature encoder; the network parameters of the layout feature encoder after the pre-training are reserved as initial values; the classification module is composed of at least one full-connection layer and is used for outputting and inputting the probability that the mask layout data fragment is a hot spot; And performing supervised training on network parameters of the layout feature encoder and the classification module by calculating the classification cross entropy loss by using a data set with a hot spot label.
8. The method of claim 1, wherein the step of constructing an unmarked mask layout data set in step one further comprises performing a data enhancement process on the segmented mask layout data segments, the data enhancement process comprising random rotation and flipping operations.
9. The method of claim 1, wherein the step of applying the hotspot detection model to perform full-chip hotspot detection in the step six specifically includes: dividing the full-chip mask layout to be detected into a plurality of data fragments with consistent sizes through a sliding window strategy with an overlapping area; inputting the data fragments into the hot spot detection model one by one, and outputting a hot spot probability value for each data fragment; And integrating the prediction results of all the data fragments to generate a hot spot distribution map covering the whole chip range.
10. A self-supervised learning based computational lithography modeling and hotspot detection apparatus, comprising: The data construction module is used for constructing an unmarked mask layout data set which comprises a plurality of mask layout data fragments extracted from a chip design layout database and segmented into the mask layout data fragments; the model construction module is used for constructing a self-supervision pre-training neural network model based on physical perception, and the model comprises a layout feature encoder, a geometric reconstruction branch and a photoetching space image prediction branch embedded with a differentiable optical simulation layer; the pre-training module is used for performing self-supervision pre-training, namely masking the mask layout data fragment and inputting the mask layout data fragment into the layout feature encoder to obtain potential features, calculating geometric reconstruction loss based on the potential features by utilizing the geometric reconstruction branch, predicting a lithography aerial image based on the potential features by utilizing the lithography aerial image prediction branch, calculating optical physical loss between the lithography aerial image and a pseudo-true value lithography aerial image generated based on an original mask layout by utilizing the differentiable optical simulation layer, and updating parameters of the neural network model according to weighted sum of the geometric reconstruction loss and the optical physical loss; The fine tuning module is used for carrying out fine tuning on the layout feature encoder after self-supervision pre-training by utilizing sample data with hot spot labels to obtain a hot spot detection model; And the detection application module is used for applying the hot spot detection model to carry out hot spot detection on the full-chip mask layout.

Description

Method and device for computing lithography modeling and hot spot detection based on self-supervised learning Technical Field The application relates to the technical field of artificial intelligence and semiconductor manufacturing, in particular to a method and a device for computational lithography modeling and hot spot detection based on self-supervision learning. Background In the context of integrated circuit fabrication processes entering 5 nm and more advanced technology nodes, photolithography is a key element that faces the challenges presented by the optical diffraction limit. In order to ensure that the pattern on the wafer can be accurately re-engraved into the design layout, resolution enhancement techniques, such as inversion lithography and light source-mask co-optimization, are commonly employed in the industry. The application of these techniques allows the pattern on the reticle to exhibit a highly complex curvilinear shape, rather than the traditional rectangular or polygonal shape. Such complex patterns are very prone to pattern defects, such as bridging, open circuit or line width narrowing, in certain areas during lithographic imaging due to optical proximity effects and process window variations, and these defect prone areas are known as lithographic hotspots. The accurate and rapid detection of the photoetching hot spot in the whole chip range is a key for guaranteeing and improving the chip manufacturing yield under advanced process nodes. The traditional detection method depends on physical lithography simulation, and although the method can provide a high-precision imaging prediction result, the calculation process is extremely time-consuming and cannot meet the requirement of rapid verification of a full-chip layout in mass production. In recent years, deep learning-based methods have been introduced into hotspot detection tasks due to their efficient reasoning capabilities. However, the existing deep learning method mostly adopts a supervised learning paradigm, and the performance of the method is highly dependent on large-scale and high-quality labeling data. In the field of lithography, the cost of obtaining samples with hot spot labels is extremely high, and the samples are usually obtained through expensive wafer exposure experiments or time-consuming accurate physical simulation, so that the training data set is limited in scale and has serious class imbalance problems. A feasible solving path is provided for utilizing massive unmarked layout data and a self-supervision learning technology. However, when the self-supervision learning framework in the field of general computer vision is directly applied to lithography layout data, the problem of characterization mismatch exists. Common self-supervision methods typically construct pre-training tasks based on geometric similarity, such as data enhancement by rotation, clipping, or mask reconstruction, for model learning geometric features. However, in the lithographic physics process, the geometry of the mask pattern is a highly nonlinear optical transformation relationship with the imaging result on the wafer. The mask patterns that are visually very different may produce the same imaging effect due to optical equivalence, whereas the geometrically very similar patterns may cause a large difference in imaging quality due to only small variations in the sub-resolution assist patterns. The existing self-supervision method cannot capture the implicit photoetching physical rule, the feature learned by the pre-training model of the method represents the geometric shape of the layout more rather than the printability of the layout, so that the generalization capability of the hot spot detection model is insufficient and the false positive rate is higher when the mask with an unseen complex curve is faced. Therefore, how to integrate the physical mechanism of lithography imaging into the self-supervision learning process, so that the model can learn the characteristic representation with the optical perception capability on the label-free data is a problem to be solved in the current technical field. Disclosure of Invention The invention provides a calculation photoetching modeling and hot spot detection method based on self-supervision learning, which mainly solves the technical problems that a deep learning hot spot detection method in the prior art depends on a large amount of expensive labeling data, and a general self-supervision learning method cannot capture photoetching physical rules, so that model generalization capability is poor, and detection precision is low. According to one aspect of the present invention, there is provided a method for computational lithography modeling and hotspot detection based on self-supervised learning, comprising the steps of: Step one, an unmarked mask layout data set is constructed, wherein the unmarked mask layout data set comprises a plurality of mask layout data fragments extracted from