PaperlayerView exact revision on arXiv

Source-linked AI summary

Image resolution enhancement for advanced semiconductor nodes

Lucas Rencker, Omid Tajalizadehkhoob, Khalid Elsayed, Artem Khachaturiants, Helda Pahlavani, Yan Guo, Jakob van de Laar, Erik Simons, Niranjan Saikumar, Hamed Sadeghian

arXiv:2610.05809v1cs.LG

TL;DR

Faster SPM scanning can reduce image resolution and the accuracy and precision of extracted parameters, despite advanced-node metrology’s need for high throughput and sub-nm resolution. The paper applies a Res-UNet enhancer to SPM images and uses transfer learning for differing structures. It reports recovery of fine image details and monitored-parameter accuracy, with transfer learning reducing data and training requirements.

  • Problem

    Advanced-node process control needs sub-nm metrology and higher throughput, but increasing SPM scanning speed can reduce image resolution and extracted-parameter accuracy and precision.

    Source: INTRODUCTION · p. 1

  • Method

    The paper uses a Res-UNet to enhance SPM image resolution and transfer learning to retrain the model for structures with differing dimensions.

    Source: ML FRAMEWORK FOR RESOLUTION ENHANCEMENT · p. 3, RESULTS AND TRANSFER LEARNING · p. 5

  • Results

    The model recovers fine details and monitored-parameter accuracy, while transfer learning produces expected results on structures of differing dimensions.

    Source: RESULTS AND TRANSFER LEARNING · p. 5, CONCLUSIONS · p. 7

  • Takeaways & Limitations

    The framework provides inputs toward increasing SPM-metrology throughput, and transfer learning reduces the data set and training period for different structure dimensions.

    Source: CONCLUSIONS · p. 7

Abstract

from arXiv · show

Advanced semiconductor nodes are pushing the limits of feature sizes and require metrology with sub-nm resolution without compromising on the throughput as needed for in-line process control. Recently, high-throughput scanning probe microscopy (SPM) based metrology and inspection tools capable of meeting these needs have been introduced to the market and qualified for use in HVM. While innovative measurement methods and tool architecture have allowed for a leap of improvement in throughput, the next step in further reducing imaging time can be obtained through the application of machine learning for enhancing the resolution of measured images for extraction of relevant parameters. In this work, we provide the general framework under which a neural network-based resolution enhancer is designed and used for SPM images. We showcase the effectiveness of this framework using measurements performed on Line/Space structures with a pitch of 200 nm. For the reusability of a pre-developed pre-trained model, we additionally leverage transfer learning and show that a new model for slightly differing structures can be re-trained and calibrated with a smaller data set of measurements performed on Line/Space structures with a pitch of 100 nm.

INTRODUCTION

Advanced-node metrology needs sub-nm resolution and higher throughput, but faster SPM scanning can reduce image resolution and parameter accuracy. The paper presents a neural-network resolution enhancer for semiconductor SPM images and extends it with transfer learning for structures of differing dimensions.

INTRODUCTION

  • Advanced nodes require sub-nm metrology for full 3D profiles and parameter extraction, while high-throughput SPM tools provide short imaging times.Source: INTRODUCTION · p. 1
  • Faster SPM scanning can sacrifice image resolution, directly affecting the accuracy and precision of extracted parameters.This trade-off matters as increased stochastic defects motivate measurements at more wafer locations.Source: INTRODUCTION · p. 1
  • The framework applies neural-network resolution enhancement to semiconductor SPM images, compensating for reduced measured pixels as scanning speed increases.The work demonstrates the approach on Line/Space structures.Source: INTRODUCTION · p. 1
  • Transfer learning is used to retrain the model for similar structures with differing dimensions using less training data, without compromising extracted-parameter accuracy or precision.Source: INTRODUCTION · p. 1

SPM MEASUREMENTS OF SEMICONDUCTOR STRUCTURES

Measurements on 100 nm-pitch Line/Space structures compare an optimized nominal scan speed with faster scans. Line height remains unaffected under a stated sampling condition, while lateral parameters lose accuracy and precision as resolution falls.

SPM MEASUREMENTS OF SEMICONDUCTOR STRUCTURES

  • Measurements used 100 nm-pitch, 50%-duty-cycle Line/Space structures, with five repeats at nominal speed and at 5× and 7× nominal speed.The nominal speed was optimized until extracted-parameter accuracy and precision across five static repeats were within the acceptable uncertainty window.Source: SPM MEASUREMENTS OF SEMICONDUCTOR… · p. 2
  • The measured parameters comprise line height, line width, LER, and LWR, with Figure 2 depicting their definitions.Source: SPM MEASUREMENTS OF SEMICONDUCTOR… · p. 2, Figure 2 · p. 2
  • At 5× nominal speed, the low-resolution image’s cross-section differs from the optimized nominal-speed image.The figure links the profile change to inaccuracy in extracted parameters.Source: Figure 1 · p. 2
  • Line height is unaffected when at least one pixel from each structure’s bottom is captured, but lateral parameters suffer in accuracy and precision as scanning speed rises.The lateral parameters include line width, line edge roughness (LER), and line width roughness (LWR).Source: SPM MEASUREMENTS OF SEMICONDUCTOR… · p. 2, Figure 3 · p. 3, Figure 4 · p. 3

ML FRAMEWORK FOR RESOLUTION ENHANCEMENT

The image-enhancement framework uses a Res-UNet architecture that combines an encoder-decoder structure with residual connections. The described implementation uses small filters, with training parameters listed in Table 1.

ML FRAMEWORK FOR RESOLUTION ENHANCEMENT

RESULTS AND TRANSFER LEARNING

The Res-UNet restores lost image pixels, and the extracted lateral parameters are compared across nominal-speed images, faster raw images, and resolution-enhanced images. Transfer learning adapts the model from 200 nm- to 100 nm-pitch structures with a substantially smaller dataset.

RESULTS AND TRANSFER LEARNING

  • Transfer learning retrains the 200 nm-pitch model for 100 nm-pitch structures with a dataset reduced by a factor of four.The retraining uses a 1e-4 learning rate and 50 epochs; Figure 8 shows recovery of lateral-parameter accuracy and precision.Source: RESULTS AND TRANSFER LEARNING · p. 5, Figure 8 · p. 6
  • Figure 6 shows recovery of lost pixels in the resolution-enhanced image and its cross-section.The middle and right images are the ML model’s input and output, respectively.Source: Figure 6 · p. 5, RESULTS AND TRANSFER LEARNING · p. 5
  • Figure 7 compares lateral-parameter accuracy and precision from nominal-speed, faster raw, and resolution-enhanced images, supporting the approach for further increasing SPM metrology throughput.The prose reports that the comparison indicates success, but the supplied passages give no numerical values.Source: RESULTS AND TRANSFER LEARNING · p. 5, Figure 7 · p. 6

CONCLUSIONS

The Res-UNet framework enhances SPM images to recover fine details and restore the accuracy of process-control parameters. Transfer learning reduces the dataset and training period needed to adapt the model to structures of different dimensions.

CONCLUSIONS

  • SPM tools face continued throughput demands as semiconductor fabs move toward massive metrology.The conclusion identifies ML image enhancement as an input for further increasing SPM-based metrology throughput.Source: CONCLUSIONS · pp. 6–7
  • The resolution enhancer recovers fine image details and returns accuracy to parameters monitored in process control.The conclusion also reports expected results when applying transfer learning to structures of different dimensions.Source: CONCLUSIONS · pp. 6–7
Loading 2610.05809v1…