As the technology node in semiconductor manufacturing continuously shrinks its feature size and boosts the transistor density, etch bias is facing great challenges that require much better control of the edge placement error (EPE). The traditional applications of etch bias either by rule or by model are sometimes of lower precision or too much time consuming. We propose and demonstrate several innovative EPE based machine learning models for etch bias that have successfully achieved satisfying accuracy and time cost for one of the latest advanced tech nodes in industry. In addition, we propose a novel methodology for massive EPE measurement on wafer that is based on automatic image processing. Three types of machine learning models (single neural network, ensemble neural networks, and random forest) and a novel feature vector used for the machine learning have been studied here. A comparison with the commercial etch-model software, Variable Etch Bias (VEB) from Mentor Graphics, has also been taken. As a result, our proposed machine learning models achieved better accuracy within greatly shortened time compared to the VEB model in our test case.

中文翻译：

---

基于边缘放置误差的蚀刻偏差的机器学习模型


随着半导体制造中的技术节点不断缩小其特征尺寸并提高晶体管密度，蚀刻偏差面临着巨大的挑战，需要更好地控制边缘放置误差（EPE）。传统的蚀刻偏差应用（无论是通过规则还是通过模型）有时精度较低或耗时过多。我们提出并演示了几种基于 EPE 的创新型蚀刻偏差机器学习模型，这些模型已成功实现了业界最新先进技术节点之一令人满意的准确性和时间成本。此外，我们提出了一种基于自动图像处理的晶圆上大规模 EPE 测量的新颖方法。这里研究了三种类型的机器学习模型（单一神经网络、集成神经网络和随机森林）和用于机器学习的新颖特征向量。还与 Mentor Graphics 的商业蚀刻模型软件 Variable Etch Bias (VEB) 进行了比较。因此，与测试用例中的 VEB 模型相比，我们提出的机器学习模型在大大缩短的时间内实现了更高的准确性。