A Sensitivity Analysis Framework for Gerchberg-Saxton Based Physics-Inspired Neural Networks in Computer-Generated Holography

IEEE Transactions on Computational Imaging, vol. 12, pp. 1636–1650

Abstract

Computer-generated holography (CGH) enables applications in holographic augmented reality (AR), 3D displays, systems neuroscience, and optical trapping. CGH involves solving an ill-posed inverse problem of phase retrieval from intensity measurements. Physics-inspired neural networks (PINNs), particularly Gerchberg-Saxton based PINNs (GS-PINNs), have been explored as learning-based alternatives to classical iterative methods like Gerchberg-Saxton (GS) algorithm. GS-PINNs are typically trained in an unsupervised manner and can enable fast inference once trained, making them attractive for real-time or near-real-time CGH applications. In practice, GS-PINNs are co-designed and trained for specific physical forward models (FMs) and associated hyperparameters (FMHs). As a result, their observed learning behavior and performance are sensitive to variations in these physical parameters, motivating a systematic analysis of FMH-dependent effects. In this work, we present a systematic sensitivity analysis framework based on Saltelli's extension of Sobol's method to quantify FMH impacts on GS-PINN performance under controlled training conditions. Our analysis shows that SLM pixel-resolution is the primary factor affecting neural network sensitivity, followed by pixel-pitch, propagation distance, and wavelength. We further observe that different forward models induce distinct sensitivity profiles, with free-space propagation exhibiting more favorable performance trends for GS-PINN than Fourier holography under the studied settings. By characterizing FM and FMH-driven sensitivity, this work provides practical guidance for forward model and hardware parameter selection in learning-based CGH under investigated simulation conditions. All results are interpreted as FMH-conditioned learning behavior rather than converged performance. Our research highlights the importance of sensitivity-aware analysis when interpreting neural network performance across physical configurations. Furthermore, it lays the groundwork for more principled evaluation and future development of physics-informed and interpretable CGH models.

Authors 6

1. Ankit Amrutkar Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

2. Bjorn Kampa Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

3. Volkmar Schulz Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

4. Johannes Stegmaier Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University

5. Markus Rothermel

   Otto-von-Guericke-Universität Magdeburg

   Affiliation as printed

   Otto-von-Guericke-University Magdeburg

6. Dorit Merhof Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University