Transfer learning for dispersion-driven inverse design across photonic devices

Inverse design requires large, device-specific datasets. In this work, we show that a neural inverse model trained using dispersion curves on photonic crystal fibers (PCFs) can be transferred to Si₃N₄ strip waveguides using only a few dispersion points per device. Our approach freezes early PCF features, learns an input-scaler, and fine-tunes shared layers and heads with gradient normalization to balance width/height regression. On a small Si₃N₄ database, we recover geometry from just 5 wavelength-dispersion points, reaching R² ≈ 0.98 (width) and R² ≈ 0.91 (height) on test data. Querying the same target dispersion on both PCF and strip waveguide exposes the platform feasibility limits (material/geometry), not merely a model fit. The result is a sample-efficient, cross-platform inverse design recipe that repurposes a single trained model across multiple photonic platforms.