Deep learning-based reconstruction of irregular laser ablation crater geometry from single plasma projection

During ultrafast laser ablation, a plasma plume forms as the laser interacts with the material, while the resulting crater geometry reflects how energy is deposited. We present a machine learning framework that reconstructs the 2D ablation pattern using only a single side-view image of the laser-induced plasma. A conditional generative adversarial network is trained to map the plasma projection directly to the corresponding crater morphology, incorporating an edge-aware loss function to improve the reconstruction of irregular contours. Despite relying on a single projection, the model successfully recovers the dominant spatial structure of the ablation pattern. Analysis of the learned representations indicates that spatial features within the plasma image contain information beyond the primary viewing axis, enabling the reconstruction of laterally irregular geometries. While fine surface textures remain challenging to resolve from a single view, the method accurately predicts the dominant crater spatial morphology. The approach complements spectral techniques by providing morphological information from indirect imaging, supports real-time in situ monitoring, and offers a practical route toward predictive diagnostics in laser-material processing. The framework could be extended to multi view fusion potentially when higher reconstruction fidelity is required.