PINN-KK: A Physics-Informed Neural Network for Causality-Compliant Spectroscopic Phase Retrieval in Low-Loss Regimes
Abstract: Spectroscopic phase retrieval becomes difficult when weak absorption, finite measurement bandwidth, and noise limit the information available for reconstruction. This study presents PINN-KK, a physics-informed neural-network framework that combines supervised in-bandwidth phase learning with a low-loss, physics-motivated out-of-bandwidth phase constraint and adaptive GradNorm loss weighting. The method is evaluated using synthetic Lorentz-oscillator spectra through an ablation experiment, 28 bandwidth and SNR conditions, and three low-loss thresholds. Removing the physics-based loss produces an approximately four-order-of-magnitude deterioration in out-of bandwidth phase RMS at the reference condition. Across the evaluated conditions, PINN-KK maintains strong in-bandwidth accuracy while substantially reducing out-of-bandwidth phase error relative to Classical KK and the supervised baseline. Performance also improves as the imposed low-loss threshold becomes stricter. These synthetic-data results demonstrate the value of incorporating an explicit low-loss physical prior into neural-network phase retrieval and motivate subsequent validation using experimental spectra.
Keywords: Kramers-Kronig relations, low-loss optics, phase retrieval, physics-informed neural networks, spectroscopic inversion
How to Cite?: Olatinwo Adenike Sola, Salako Najeem Abiodun, "PINN-KK: A Physics-Informed Neural Network for Causality-Compliant Spectroscopic Phase Retrieval in Low-Loss Regimes", Volume 15 Issue 8, August 2026, International Journal of Science and Research (IJSR), Pages: 1305-1314, https://www.ijsr.net/getabstract.php?paperid=SR26810183613, DOI: https://dx.doi.org/10.21275/SR26810183613