Machine Learning-assisted Investigation of Spatial-temporal Correlations from Process to Quality in Laser Directed Energy Deposition

Laser directed energy deposition (LDED) is gaining increasing industrial traction in sectors such as aerospace, automotive, and energy due to its ability to fabricate complex, high-performance components with near-net-shape geometry, high material utilization, and process efficiency. Despite these advantages, consistent quality assurance remains a major obstacle, often hindered by dynamic phenomena like melt-pool fluctuations and thermal accumulation effects. This study presents a machine learning-assisted framework to investigate and model the intricate spatial-temporal relationships between process dynamics and as-built part quality in LDED. The process behavior is captured through the evolution of the melt-pool area, automatically segmented from in-situ monitoring images using a semantic segmentation neural network. Surface quality is quantified by spatially resolved deviations derived from 3D point cloud measurements. By temporally and spatially aligning the process and quality data, dynamic heatmaps are constructed to visualize the evolution of quality features across build layers. Analysis of these maps reveals consistent patterns such as edge-induced anomalies and periodic surface variations linked to heat accumulation. Building on this understanding, a regression-based machine learning model is developed to learn and predict the process–quality relationship. Trained on thin-wall samples manufactured under varied parameters, the model successfully identifies localized quality regimes, offering valuable insight for process control and post-processing optimization. This work highlights the potential of machine learning to advance intelligent quality prediction and deepen understanding of process–structure–property interactions in additive manufacturing.