Achieving high processing quality for chemical mechanical planarization (CMP) in semiconductor manufacturing is difficult due to the distinct process variations associated with this method, such as drift and shift. Run-to-run control aims to maintain the targeted process quality by reducing the effect of process variations. The goal of controller learning is to infer an underlying output–input reverse mapping based on input–output samples considering the process variations. Existing controllers learn reverse mapping by minimizing the total mapping error for sample data. However, this approach often fails to generate inputs for unseen target outputs because conditional input distributions on target outputs are not captured in the learning. In this study, we propose a controller based on a least squares generative adversarial network (LSGAN) that can capture the input distributions. GANs are deep-learning architectures composed of two neural nets: a generator and a discriminator. In the proposed model, the generator attempts to produce fake input distributions that are similar to the real input distributions considering the process variation features extracted using convolutional layers, while the discriminator attempts to detect the fake distributions. Competition in this game drives both networks to improve their performance until the generated input distributions are indistinguishable from the real distributions. An experiment using the data obtained from a work-site CMP tool verified that the proposed model outperformed the comparison models in terms of control accuracy and computation time.

由于与该方法相关的不同工艺变化（例如漂移和偏移），很难在半导体制造中实现化学机械平面化（CMP）的高处理质量。运行间控制旨在通过减少过程变化的影响来维持目标过程质量。控制器学习的目的是在考虑过程变化的基础上，根据输入-输出样本来推断潜在的输出-输入反向映射。现有的控制器通过最小化样本数据的总映射误差来学习反向映射。但是，这种方法通常无法为看不见的目标输出生成输入，因为在学习中未捕获目标输出上的条件输入分布。在这个研究中，我们提出了一种基于最小二乘法生成对抗网络（LSGAN）的控制器，该控制器可以捕获输入分布。GAN是由两个神经网络组成的深度学习架构：生成器和鉴别器。在提出的模型中，考虑到使用卷积层提取的过程变化特征，生成器尝试生成与实际输入分布相似的伪输入分布，而鉴别器尝试检测伪分布。在该游戏中，竞争促使两个网络提高其性能，直到生成的输入分布与实际分布无法区分为止。使用从现场CMP工具获得的数据进行的实验证明，在控制精度和计算时间方面，所提出的模型优于比较模型。