OmniPMNet: Bridging discrete and gridded PM10 forecasts via omni-query neural processes

Batch of CS/ML arXiv announcements covering: (1) OmniPM-Net — a ConvCNP-based fusion model that lifts station GNN forecasts to a grid via terrain-aware Gaussian set convolution and Spatial Source Attention to produce consistent station and gridded PM10 forecasts with improved tail and dust-episode performance; (2) Semidirect Fourier Delta Attention (SFDA) — a phase-controlled generalization of delta/linear attention with block-rotational Fourier control and a constructive chunk-WY factorization that bounds rank growth and gives formal stability/complexity guarantees; (3) PhiCalNet — an FPP (fr