Abstract: Wafer manufacturing cycle time forecasting is the core problem of semiconductor wafer fabrication system operation optimization, which is the key to guaranteeing the on-time delivery of wafer products. Deep learning methods learn the data fluctuation laws from massive data, construct black-box prediction models of complex systems, and achieve impressive prediction accuracy in static environments. However, under dynamic system state fluctuation, such as workshop work-in-process levels, current methods cannot stay accurate prediction due to the lack of interpretability to explain the changing rules of the forecasting model with system states. Therefore, an interpretable deep learning method(IDLM) for wafer manufacturing cycle time forecasting is proposed to clarify the organization rules of forecasting neural networks under different system states. First, a brain-inspired interpretable structural model of the wafer manufacturing cycle time forecasting neural network is constructed to provide a structural basis for the analysis of the network in the organization form of "neurons-neural circuits-neural network". Second, a key neuron recognition method of cycle time forecasting network is proposed to filter important neurons from the network with information entropy weighted rules constraint. Finally, a key neural circuit search algorithm is designed to quickly search for the optimal combination of similar neurons to obtain the key forecasting circuits. The experimental results show that IPM can extract the key neural circuits of the forecasting network while maintaining the accuracy, which provides a key structural basis for the network self-assembly under dynamic environments.

Key words: deep neural network, interpretable learning, network parsing, cycle time forecasting, semiconductor manufacturing