The invention relates to the technical field of
semiconductor device modeling, in particular to a neural
network data DC preprocessing method and modeling
system.The method comprises the steps that transfer characteristic curve
test data, output characteristic curve
test data and
diode forward characteristic curve
test data of a
semiconductor device are obtained to serve as
original data; performing cleaning processes on the three types of data respectively and then merging and de-duplicating the three types of data; performing data deformation, normalization and
standardization on the merged data in sequence; taking Vgs, Vds, Fingernorm, Widthnorm and Tempnorm as input features, taking Idsy as an output target, adopting a full-connection
feedforward neural network for training, and obtaining a model weight file; the weight file is injected into the
Verilog-A language, the flow of anti-
standardization, anti-normalization and anti-data deformation is carried out in the
Verilog-A language to restore Ids, the mapping relation of the formula Ids = f (Vgs, Vds, size and temperature) is obtained, SPICE
simulation is achieved, multi-size and temperature
adaptation can be considered, the precision of Vds = 0 during
simulation can be guaranteed, process corner support can be provided, and SPICE
simulation is guaranteed to be on the ground.