Physics-driven tandem inverse design neural network for efficient optimization of UV–Vis meta-devices

• DOI: 10.1016/j.apsadv.2023.100503
• Publication Date: 2023-11-18T22:53:21Z
• Authors: Sadia Noureen; Iqrar Hussain Syed; Sumbel Ijaz; Alaa Awad Abdellatif; Humberto Cabrera; Muhammad Zubair; Yehia Massoud; Muhammad Qasim Me...
• Keywords: Nano-structures; Industrial electrochemistry; Model reduction; TA401-492; 006; Deep learning; 02 engineering and technology; 0210 nano-technology; Materials of engineering and construction. Mechanics of materials; Tandem inverse design neural network; TP250-261

This paper presents two tandemly stacked forward and inverse deep neural networks to model optimize cylindrical shaped transmissive meta-atoms for UV–Vis regime. Conventional modeling of these subwavelength calls repetitive solution Maxwell equations over each instance a mesh grid using some high-end commercial EM simulator. In contrast, Deep Learning (DL)-based approaches can significantly expedite this process owing their abilities intelligently learn approximate Maxwell's without explicitly solving them, hence anticipating the electromagnetic (EM) response given meta-atom within split-second. Despite advantages DL-based solutions metasurface design optimization, dataset collection training DL models still requires time-tedious conventional approaches. Here, we aim enhance learning performance while reducing requirements complexity proposed by enriching them with more underlying physical facts substantial knowledge particular problem. Therefore, in addition list geometrical parameters meta-atoms, network is also fed additional material physics related as part its input, including spectral information, type (dielectric or plasmon), maximum height (based on fabrication constraints), period rod. results enhancing model's utilizing minimum samples, predicting transmission amplitude phase negligible error tolerance, i.e. MSE: 2.1 × 10−3, dielectric plasmonic at various operating wavelengths (within targeted regimes). Optimization such meta-structures another challenging task that hit trial intuitive guesses, lengthy parametric sweeps time taking intelligent algorithms. an behind trained tandem assembly together predict best set dimensions most suitable achieve desired response. The same information all aforementioned parameters, mapping network. order assess impact incorporating supplementary problem specific comparative study regarding number hidden layers amount size carried out networks. analysis shows loss greatly decreased, learns efficiently scheme about under consideration. As result, show even smaller lesser layers, realize appropriate MSE.