Thermal–fluid simulation is becoming increasingly important in laser powder bed fusion (LPBF), as melt-pool hydrodynamics strongly influence track continuity, keyholing, porosity, and lack-of-fusion defects. However, high-fidelity CFD models that resolve free-surface flow and phase change remain computationally prohibitive for large-scale analysis and process optimization. In such models, pressure and velocity fields primarily serve as latent variables that mediate the evolution of melt-pool morphology, governing free-surface motion, keyhole formation, and solid–liquid interface dynamics, while the only external actuation is provided by the moving heat source. This structure suggests that the temperature field, particularly its spatiotemporal history, encodes rich information about the underlying multiphase dynamics. Motivated by this observation, we develop a neural surrogate that predicts melt-pool morphology evolution, represented by material volume fraction and solid–liquid phase fields, conditioned solely on temperature histories and scan parameters, thereby bypassing the computationally expensive fluid solving that requires extremely fine spatial and temporal grids. To further overcome the computational challenges of 3D spatiotemporal modeling highlighted in recent work, we introduce two key strategies: first, spatial dimensions are compressed through an autoencoder, enabling efficient autoregressive time integration in latent space; second, the surrogate operates on a temporal grid over 10 times coarser than typical CFD simulations. The model is trained on high-fidelity thermal–fluid simulations over localized melt-pool patches and generalizes across a spectrum of scanning speeds and laser powers within the standard processing window. This capability makes the surrogate differentiable and suitable for both gradient-based process optimization, providing a pathway toward physics-embedded, defect-free LPBF processes.