| Abstract Scope |
Metal additive manufacturing (AM) accelerates alloy development, but localized layer-wise heating introduces stochastic melt-pool dynamics, defect formation, and microstructure evolution. In-situ monitoring captures process variability yet remains sparse and surface biased, while high-fidelity simulations provide three-dimensional process-to-structure information but are computationally expensive and often deterministic. Here, we present a conditional latent diffusion model (LDM) that fuses multi-modal in-situ signals, including coaxial melt-pool images, laser power, and laser spot position, with high-fidelity thermal and microstructure simulations for stochasticity-aware online prediction in metal AM. The monitoring data encode real-time process variability, while simulations provide physical supervision for unobserved quantities such as sub-surface melt-pool geometry, solidification conditions, grain morphology, and crystallographic texture. Operating in a compressed latent space, the LDM enables millisecond-scale inference of three-dimensional thermal and microstructural evolution. This framework supports defect-prone region identification, track-level uncertainty quantification, and process-parameter optimization, advancing digital twins for reliable qualification and accelerated alloy design. |