Deep Learning-Driven Fast and Efficient Inverse Design Model for Meta-Absorber and Meta-Emitters

Solar thermophotovoltaic (STPV) systems suffers from major difficulties in the realm of design and optimization of meta-surface based solar absorbers and selective emitters. Traditional techniques of designing them frequently include lengthy parametric sweeps and repetitive electromagnetic simulations, which can be time-consuming and computationally exhaustive. This research tackles the constraints of traditional design methodologies by proposing a deep learning-based transposed convolutional neural network (TCNN) for the inverse design of absorber and emitter meta-atoms. The proposed TCNN is especially developed to anticipate ideal metaatom shapes that fulfil required optical spectrum targets, which considerably speeds up the design process. Given the sequential and the complex nature of optical spectra, three critical preprocessing steps are followed which includes: Gaussian smoothing, discrete cosine transformation, and low pass filtering. These techniques condense the input data dimensions, reducing computational demands while preserving essential features necessary for accurate prediction. The resulting low-dimensional vector is then processed via proposed TCNN, to forecast structural images of meta-atoms. These predicted metaatom images undergo further refinement through Conditional Random Fields (CRF) and Connected Component Labeling (CCL) techniques. These post-processing steps enhance the detection and identification of individual nano-resonator structures within the generated images, ensuring high fidelity in the final output. By leveraging deep learning, our approach significantly accelerates the meta-atom design process, yielding accurate structural predictions in a matter of seconds, as opposed to the hours or days required by traditional methods.