Improving Laser Direct Writing Overlay Precision Based on a Deep Learning Method.

Area of Science:

• Optics and Photonics
• Machine Learning
• Nanotechnology

Background:

• Overlay alignment precision is critical in laser direct writing systems for micro-nano fabrication.
• Alignment errors originate from optical aberrations, mechanical drift, and fiducial mark imperfections.
• Vision systems' interpretation of fiducial mark coordinates contributes significantly to residual alignment error.

Purpose of the Study:

• To develop a deep learning-based method to improve overlay alignment precision.
• To create a convolutional neural network (CNN) model for predicting coordinate calculation errors in fiducial marks.
• To compare the performance of CNNs against traditional feedforward neural networks (FNNs) for this task.

Main Methods:

• Generated 66,000 computer-simulated defective crosshair marks to mimic real-world fiducial mark imperfections.
• Developed and compared 14 neural network architectures, including 8 CNN variants and 6 FNN configurations.
• Trained and validated the models using the generated datasets to evaluate prediction accuracy.

Main Results:

• A simple CNN architecture achieved a mean squared error (MSE) of 0.0011 (training) and 0.0016 (validation).
• The CNN demonstrated a 90% error reduction compared to FNN structures.
• Experimental results showed CNN prediction errors below 100 nm in X/Y coordinates, outperforming FNNs.

Conclusions:

• Deep learning, specifically CNNs, offers a powerful approach to enhance overlay alignment precision in laser direct writing.
• The CNN's ability in local feature extraction and translation invariance is key to its superior performance.
• This method establishes a new paradigm for precision enhancement in micro-nano optical device fabrication.