Physics-guided deep neural networks for Laguerre-Gaussian modal decomposition via PCA-ranked Zernike descriptors

Direct pixel-space learning for Laguerre-Gaussian (LG) modal decomposition suffers from high input dimensionality and severe computational redundancy. Here, we present a physics-guided framework that maps intensity patterns into a compact Zernike descriptor space coupled to a deep neural network (DNN). Using principal component analysis (PCA), we systematically interrogate the effective dimensionality of the sampled Zernike feature space and establish a variance-weighted ranking for Zernike features. Experimental validation using a digital micromirror device (DMD)-based setup demonstrates that dominant descriptors capture >99% of dataset variance. Truncated feature subsets achieve saturated reconstruction fidelity (∼ 98%) and near-unity spatial correlation under realistic optical perturbations. This framework eliminates full-resolution image processing, paving the way for potential real-time optical mode sorting applications.