Fig. 1: Caveats of using DL and direct solvers in PDD.

i) DL models promise speed and efficient searching of design spaces, taking seconds to process hundreds of designs, but lack physical intuition, suffer from slow training procedures, and lack universality. Conversely, direct solvers are accurate, grounded in physics, and convergent, but the overall computation is slow and expensive, taking hundreds of hours to process a similar number of designs. Mode analysis image adapted with permission from ref. ⁷⁵, Copyright Ansys Lumerical, 2025. Data for the processing time plot retrieved from ref. ⁷⁶, Copyright 2020 AIP Publishing and ref. ¹⁷. Design Accuracy plot adapted from ref. ⁷⁷, Copyright 2018 IEEE. ii) DL performance is often governed by the problem formulation, data quality, and the network that interprets the data. Waveguide image adapted with permission from ref. ⁷⁸, Copyright 2014 Elsevier. Photonic crystal and bandgap images adapted with permission from ref. ⁷⁹, Copyright 2019 Elsevier. Dataset and network images adapted with permission from refs. ^16,80, Copyright 2023 Optica Publishing Group and ref. ⁸¹, Copyright 2020 American Chemical Society. iii) DL is very resource intensive, often exceeding that of classical methods. Google TPU image adapted with permission from ref. ⁸².