Architecture of the multimodal gating network. Separate encoders process the PS1, unWISE, and tabular inputs; an adaptive gating module assigns normalized modality weights; and the fused representation is passed to a multilayer perceptron (MLP) classifier to produce the candidate score used for first-stage IR-excess preselection. Here, CNN, ResSE, and GELU stand for convolutional neural network, residual squeeze-and-excitation block, and Gaussian error linear unit, respectively.