一种基于人工神经网络的反馈式神经元数优化方法

摘要

针对纳米级Cu薄膜电阻率,基于BP神经网络模型,本文提出了一种反馈式神经网络优化方法,利用蒙特卡洛分析方法对隐含层神经元数进行了优化,并基于随机样本集进行网络训练,建立了反馈式BP神经网络的电阻率预测模型.通过100组学习样本训练后的神经网络模型,与50组测试样本进行测试,结果表明,所提方法能够实现电学参数值与金属Cu电阻率较好的非线性映射,预测结果与Marom模型相比较,最大误差不超过4％,并且训练范围外的预测结果与测试样本吻合较好,验证了该方法的精度和泛化能力,为超薄金属互连电阻率模型估算提供了重要参考.%Based on BP neural network model,a feedback neural network optimization method is presented for the electrical resistivity of Cu in nanometric dimensions.The number of neurons in hidden layers is optimized by means of Monte Carlo analysis method.Random sample set is trained and a resistivity prediction model is established by the developed method.50 sets of test samples are used to validate the neural network model trained by 100 random samples.The findings indicate that a good nonlinear mapping can be obtained between electrical parameters and resistivity of Cu film.Maximum error between Marom model and the proposed model is less than 4％ and the outside trained results are in a good accordance with the test samples,which verify precision and generalization capability of this method.And it can provide a practical reference for the resistivity performance estimation with ultra-thin metal interconnect.