J. Eur. Opt. Society-Rapid Publ. 22, 20( 2026) 191
Figure
6. The input and output values associated with post-training group velocity predictions obtained using deep learning models, together with the resulting prediction – actual, residual – prediction, frequency – residual, and residual ratio comparison plots.
Figure 7. The input and output values associated with post-training bandwidth predictions obtained using deep learning models, together with the resulting prediction – actual, residual – prediction, frequency – residual, and residual ratio comparison plots.
approaches, the adequacy of these models was evaluated, and inverse design was employed to predict input configurations capable of yielding the desired output parameters. However, as noted in the previous section, certain aspects remain open to further improvement, one of which is the enhancement of the R 2 performance.
For this reason, the investigation of negative refractive index generation was extended in the final part of the study. In this stage, rather than considering a single incidence angle, wave-vector – normalized-frequency solutions corresponding to waves incident at multiple angles( 0 – 45 °) were computed using 10 000 randomly generated parameter sets to construct a new dataset. The geometry was kept identical to the binary geometry presented in the previous section, while deep learning models were employed for data processing and analysis.
Using this approach, the R 2 values for both output parameters were increased beyond 0.98 when evaluated on the training dataset. The improvements achieved for both output types are illustrated in Figures 6 and 7.