Abstract:
To alleviate the formidable training and dataset accumulation workload of the neural network for systems with multiple input/output parameters, i.e., high dimensionality and complexity, we present an extended neural network (ENN) and flexible ENN (FENN) to help modeling the scalable photonics devices and systems. ENN can save from 19.16% to 40% of the database collection cost comparing to the artificial NN (ANN) method when extending the modeling from 4- to 10-layer and 10- to 12-layer cases in the thin optical film studies. And the FENN can generate an appended general network for the iteratively added layers, therefore the decomposed functional sections of the system could be represented by the network, especially for structurally similar components in the photonic circuits.
