<p>Accurate photometric redshift estimation is essential for large-scale cosmological surveys, enabling studies of galaxy evolution and the large-scale structure of the Universe, when spectroscopic measurements are impractical. Template fitting and machine learning are the two primary methods for estimating photometric redshifts, each with its own set of advantages and drawbacks. We present a hybrid framework that unifies the two through a Physics-Guided Neural Network (PGNN), integrating Spectral Energy Distribution (SED) templates with a multimodal deep learning architecture. The model combines Convolutional Neural Networks (CNNs) to extract morphological information from galaxy images, bidirectional LSTM (BiLSTM) layers to capture wavelength-ordered dependencies among photometric bands, and a cross-attention mechanism that adaptively fuses photometric and image-derived features for enhanced redshift inference. Bayesian layers further provide uncertainty estimates. Using the publicly available “GalaxiesML” dataset, which comprises <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\sim \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>∼</mo> </math></EquationSource> </InlineEquation>300,000 galaxies from the Hyper Suprime-Cam survey’s PDR2 release, our method achieves an RMS error of 0.0942, a catastrophic outlier rate of 1.28%, and a bias of 0.0002, surpassing existing approaches on the same dataset. By embedding physical priors, spectral correlations, and cross-modal interactions within a unified architecture, this work establishes a robust, interpretable, and scalable framework for photometric redshift estimation in next-generation surveys such as LSST and Euclid .</p>