Gao Guohan, Wang Jiong, Liu Xin, Du Junfeng, Bian Jiang, Yang Hu
Institute of Optics and Electronics, Chinese Academy of Sciences, Chengdu 610209, China.
National Key Laboratory of Optical Field Manipulation Science and Technology, Chinese Academy of Sciences, Chengdu 610209, China.
Micromachines (Basel). 2025 Jul 28;16(8):871. doi: 10.3390/mi16080871.
This study proposes a deep learning-based method to improve overlay alignment precision in laser direct writing systems. Alignment errors arise from multiple sources in nanoscale processes, including optical aberrations, mechanical drift, and fiducial mark imperfections. A significant portion of the residual alignment error stems from the interpretation of mark coordinates by the vision system and algorithms. Here, we developed a convolutional neural network (CNN) model to predict the coordinates calculation error of 66,000 sets of computer-generated defective crosshair marks (simulating real fiducial mark imperfections). We compared 14 neural network architectures (8 CNN variants and 6 feedforward neural network (FNN) configurations) and found a well-performing, simple CNN structure achieving a mean squared error (MSE) of 0.0011 on the training sets and 0.0016 on the validation sets, demonstrating 90% error reduction compared to the FNN structure. Experimental results on test datasets showed the CNN's capability to maintain prediction errors below 100 nm in both X/Y coordinates, significantly outperforming traditional FNN approaches. The proposed method's success stems from the CNN's inherent advantages in local feature extraction and translation invariance, combined with a simplified network architecture that prevents overfitting while maintaining computational efficiency. This breakthrough establishes a new paradigm for precision enhancement in micro-nano optical device fabrication.
本研究提出了一种基于深度学习的方法，以提高激光直写系统中的套刻对准精度。在纳米级工艺中，对准误差源于多个来源，包括光学像差、机械漂移和基准标记缺陷。残余对准误差的很大一部分源于视觉系统和算法对标记坐标的解读。在此，我们开发了一种卷积神经网络（CNN）模型，用于预测66000组计算机生成的有缺陷十字准线标记（模拟实际基准标记缺陷）的坐标计算误差。我们比较了14种神经网络架构（8种CNN变体和6种前馈神经网络（FNN）配置），发现一种性能良好的简单CNN结构在训练集上的均方误差（MSE）为0.0011，在验证集上为0.0016，与FNN结构相比，误差降低了90%。测试数据集的实验结果表明，CNN能够在X/Y坐标中均将预测误差保持在100 nm以下，显著优于传统的FNN方法。所提出方法的成功源于CNN在局部特征提取和平移不变性方面的固有优势，以及简化的网络架构，该架构可防止过拟合同时保持计算效率。这一突破为微纳光学器件制造中的精度提升建立了新的范例。
