Neural-Network-Driven Fast Design of Magnetophotonic Refractometric Sensors in Optical Telecommunication-Band

A comprehensive dataset was generated to train the neural network model for predicting the TMOKE spectral response of magnetophotonic nanogratings. The dataset was created using an automated pipeline that couples full-wave finite-element simulations (performed in COMSOL Multiphysics®) with scripted parameter sampling and post-processing via LiveLink™ for MATLAB®. The input space comprises seven parameters per sample: four geometric design variables—grating period (Λ), ridge width (w_g), grating thickness (t_g), and silicon layer thickness (t_Si)—along with three fixed analyte refractive index values (n_a = 1.33, 1.36, and 1.39). The geometric parameters were pseudo-randomly sampled within the following ranges: Λ ∈ [645, 660] nm, w_g ∈ [320, 360] nm, t_g ∈ [120, 150] nm, and t_Si ∈ [190, 210] nm. For each configuration, the working wavelength was swept from 1500 nm to 1600 nm (covering the optical C-band) at a fixed incidence angle of 15°. For each input sample, the simulation pipeline computes the transmittance for both magnetization states (m_z = +1 and m_z = -1) and extracts the TMOKE spectrum using the definition TMOKE = 2(T(+1) - T(-1))/(T(+1) + T(-1)). The output data for each sample includes: (i) the full TMOKE spectral curve over the wavelength range, (ii) the maximum TMOKE amplitude (|TMOKE_max,j|) and its corresponding resonance wavelength (λ_max,j) for each of the three analyte RI values, and (iii) the refractive index sensitivity S, calculated from the linear shift of λ_max with respect to n_a. A total of 47,262 input-output pairs were generated through this automated pipeline. The dataset was structured and serialized using the joblib library to enable fast disk access and efficient training. All input and output variables were normalized to zero mean and unit variance before training, and the complete dataset was randomly partitioned into training and testing subsets using an 80/20 split to ensure robust generalization assessment. This dataset serves as the foundation for training the neural network model, enabling the learning of the nonlinear mapping between nanograting geometry, analyte refractive index, wavelength, and the corresponding TMOKE spectral response. Once trained, the model acts as a fast surrogate for full-wave simulations, allowing rapid prediction of analyte-dependent spectra and sensitivities without the need for repeated computationally expensive electromagnetic simulations.