In LED semiconductor manufacturing, it is important to evaluate the wafer reflectance in order to validate the quality of wafers. In this work, a learning model based on ${K}$ -means clustering and neural networks is proposed to analyze the relationship between etching parameters and wafer reflectance under a complex etching environment. The implemented clustering algorithm is used to cluster historical data and reduce the effect caused by different processing environments. Then, for each obtained cluster, a neural network is developed to learn the relationship between etching parameters and wafer reflectance. Finally, a real case study from an LED semiconductor fab is conducted to show the application of the proposed method. Experiments show that the proposed method achieves much better performance in comparison with support vector machine for mapping the relationship between etching parameters and wafer reflectance. Also, by the proposed method, the average prediction accuracy of the wafer reflectance can be improved by up to 9.38%, and the mean square error is reduced by 21.64% compared with the methods without clustering.

中文翻译：

---

基于复杂刻蚀的晶圆反射率预测。 ķ-均值聚类和神经网络

在LED半导体制造中，重要的是评估晶片的反射率，以验证晶片的质量。在这项工作中，基于 $ {K} $ 提出了基于均值聚类和神经网络的方法，以分析复杂刻蚀环境下刻蚀参数与晶圆反射率之间的关系。所实现的聚类算法用于对历史数据进行聚类，并减少不同处理环境所造成的影响。然后，对于每个获得的簇，开发神经网络以学习蚀刻参数和晶片反射率之间的关系。最后，从LED半导体晶圆厂进行了实际案例研究，以展示所提出方法的应用。实验表明，与支持向量机相比，该方法在刻蚀参数与晶圆反射率之间的映射关系上具有更好的性能。而且，通过提出的方法，