Diffusing Heat: Generative Scan Path Optimization for Laser Powder Bed Fusion

Island scanning is widely adopted in Laser Powder Bed Fusion (LPBF) to mitigate residual stress and distortion. The global visiting order of islands still governs the layer-wise thermal history. Existing geometric heuristics are efficient but largely ignore history-dependent heat accumulation on irregular topologies, while physics-driven global optimization requires repeated thermal evaluation and is typically computationally expensive for online planning. We present Diffusing Heat, a diffusion-model-based planner that generates heat-aware island sequences for complex layer geometries. We first couple a high throughput Gaussian-kernel thermal surrogate with simulated annealing to build an expert dataset under a long-horizon thermal-uniformity objective that penalizes both sustained non uniformity and peak overheating events. Conditioned on the set of active island centroids, a conditional diffusion model denoises a permutation-like successor representation in parallel and produces score matrices that are close to valid permutations. At inference time, sampled scores are decoded into feasible open scan paths and further improved by a lightweight objective consistent refinement stage under the same surrogate metric. Extensive simulated evaluations demonstrate that our planner consistently improves thermal uniformity over representative heuristic baselines, while approaching the solution quality of offline combinatorial optimization and accelerating the planning process by over 20 times. Furthermore, physical LPBF experiments show our planner reduces maximum distortion by up to 53% compared to standard geometric patterns.