稳健估计用于氮化铝金属化工艺的研究

应用化学镀镍的方法实现了氮化铝的金属化。为得到较大的氮化铝金属化层粘附力，运用基于稳健估计的神经网络研究氮化铝金属化学镀镍的反应参数与金属层粘附力的关系。为使神经网络更加稳健，本文根据统计学原理，在前馈神经网络基础上，采取稳健估计方法改进神经网络。建立了定量预测粘附力性能的模型，并进行实验验证。确定金属化工艺中稳定的优化工作区域。结果表明，稳健估计方法既有传统神经网络的优点，又有较强的抵抗异常值的能力，具有较广泛的实用性。

Study on Metallization of Aluminum Nitride Using Robust Estimation Method

Electroless plating method is applied to metallize the AlN substrate. The improved neural network is applied in study of the relationship between reaction parameters of electroless plating nickel in the process of the metallization of aluminum nitride and the adhesion strength between the metal layer and the substrate. To make neural network more robust, a novel robust estimation function is proposed to improve the feedforward neural network according to the theorem of statistics. The quantitative prediction model is developed. The optimum operation region of the process parameters is found in the new neural network. The results show merits of robust estimation,which not only holds the merit of traditional neural network, but also has robustness against outliers.