IC metrology is a necessary means for measuring the fabrication performance in the semiconductor industry. It is significant for yield enhancement and process control. However, real-time monitoring of wafer production is required in recent years especially for the 300mm semiconductor manufacturing. Therefore, virtual metrology (VM) is developed for the tide of demand. It is a novel technology to predict the process results based on the previous metrology measurements, instead of measuring practically. Consequently it can assist in achieving total quality management and enable run-to-run control. In this paper a systematic methodology for virtual metrology is proposed. This VM system which is mainly designed for the process subject to linear process drift consists of a piecewise linear neural network and a fuzzy neural network. Because many semiconductor processes exhibit inevitable steady drifts in nature, the design of piecewise linear neural network is to approximate the drift trend. In addition, the influence of process recipe on fabrication outcome is learned using the fuzzy neural network. The system has good generalization capability and performance. Thus, it provides an effective and economical solution for metrology prediction.