Neurodevice - neural network device modelling interface for VLSI design

A novel, fast and accurate neural network tool is proposed for efficient technology independent implementation of the interface between device modelling and circuit simulation. Modified backpropagation, conjugate gradient and Levenberg-Marquardt optimization algorithms are applied in network learning. Simulations show fast convergence and an excellent fit of recalled characteristics to the measured device data. The utilized algorithms are robust and capable of presenting the entire device characteristics unaltered even with largely reduced amount of the learning material. The good monotonicity of the neural network generated device data facilitates the usage of the method in circuit simulation purposes. Possible further applications of implementing circuit level macromodels with this technique are discussed