Optical waveguides can be used to trap atoms in well-defined locations close to the waveguide surface and enable strong photon-atom interactions. Calculating the accurate parameters of shape, depth, and width (frequency) of the trapping optical potential are hampered by limited computational resources and capabilities of numerical tools.

Our atomic trapping system is composed of a toroidal nano-antenna placed on top of a waveguide, illuminated by blue- and red-detuned fields from the atom D-line. The focused evanescent field from the nano-antenna generates the trap. However, in order to efficiently tailor the potential trap, a nested sweep over multiple parameters needs to be performed. As more parameters need to be tuned, more simulations need to be performed. Given that a single FDTD simulation in Lumerical with a fine mesh takes 3-5 hours for one set of parameters, a fine-tuned parametric sweep over so many variables is unrealistic.

In this work, we utilise deep learning algorithms in order to efficiently predict the parameters to design an atomic trap. The results presented here are generated using a multi-input and multi-output deep learning model for a regression problem implemented using Keras. Our initial guess of the geometry of the nanoantenna is a toroid. The parameters to be predicted are 1) the geometric properties of the toroid, and 2) the input powers of the blue- and red-detuned fields.