Accurate nano-photonic device spectra calculation using data-driven methods

This study employed a data-driven approach involving the creation and training of a deep neural network model to swiftly compute spectral data for nano-photonic devices. Initially, the transfer matrix method was utilized to compute transmission and reflection spectra for one million layered materials composed of SiO₂/Si₃N₄. These spectra were then used as the training data for the deep neural network. Remarkably, despite using a training set that represented just one billionth of all possible samples within the design space, the resulting model displayed exceptional accuracy. More than 99.71% of the predictions demonstrated a standard error below 1%. This method represents a significant advancement over traditional design approaches, as it drastically reduces the complexity for designers. Moreover, the deep neural network model is less than 1 megabyte in size, making it easy to integrate into micro-optoelectronic devices.