Evaluating the manufacturability of GaAs/AlGaAs multiple quantum well avalanche photodiodes using neural networks

This paper presents a novel methodology for the parametric yield prediction of GaAs/AlGaAs multiple quantum well (MQW) avalanche photodiodes (APDs). Even in a defect-free manufacturing environment, random variations in the APD fabrication process lead to varying levels of device performance. Accurate performance prediction requires precise characterization of these variations. The approach described herein requires a model of the probability distribution of each of the relevant process variables, as well as a model to account for the correlation between this measured process data and device performance metrics. Neural networks are proposed as a tool for generating these models, which enable the computation of the joint density function required for predicting performance using Jacobian transformation method. The resulting density function can then be numerically integrated to determine parametric yield. In applying this methodology to MQW APDs, using a small number of test devices enables accurate prediction of the expected performance variation of APD gain and noise in larger populations of devices. This approach potentially allows yield estimation prior to high volume manufacturing.