ERI: vision-simulation-driven in-situ repair of recoater streaks in laser powder bed fusion

The National Science Foundation Division of Civil, Mechanical, and Manufacturing Innovation awarded the University of North Florida $200,000 on July 1, 2026, under the Engineering Research Initiation program to develop a vision-simulation-driven in-situ quality-control system for laser powder bed fusion manufacturing.

The project addresses defect-driven scrap and rework in metal additive manufacturing by creating a within-layer detect-predict-decide-act loop that couples real-time perception, modeling, and control. A lightweight vision model will segment and quantify recoater streaks on each layer in no more than one tenth of a second with calibrated confidence scores. A layer-aware digital twin implemented as a fast hybrid surrogate will map streak features, scan parameters, and energy/cooling descriptors to a porosity-risk map, scoring candidate repairs by predicted risk reduction and time cost. This in-process capability observes each layer and applies corrective actions only where needed, reducing defect propagation to later layers that degrades reliability and increases cost, energy use, and material waste.

The recipient is the University of North Florida's Office of Research and Sponsored Programs in Jacksonville, Florida. Performance extends from July 1, 2026, through June 30, 2028. The project will integrate results into course modules and laboratory exercises for student training in data-centric manufacturing and conduct outreach with regional manufacturers and community colleges to expand participation in manufacturing education and accelerate adoption of modern quality practices.

Abstract

This engineering research initiation (ERI) award will strengthen metal additive manufacturing by reducing defect-driven scrap and rework in laser powder bed fusion, a process used to produce geometrically complex metal parts for aerospace, biomedical, and energy systems. A common reason of build failure is the formation of streaks when the powder spreading recoater damages or disturbs. These streaks can trigger porosity and incomplete melting that propagate to later layers, degrading reliability and increasing cost, energy use, and material waste. This project will create a practical, in-process quality-control capability that observes each layer and applies the smallest safe corrective action only where IT is needed. By turning layer images into risk-aware interventions, this work will advance the national interest by promoting the progress of metal additive manufacturing, supporting a more resilient industrial base through higher yield and less waste. Results will be integrated into course modules and laboratory exercises that train students in data-centric manufacturing. Outreach with regional manufacturers and community colleges will expand participation in manufacturing education and training and accelerate adoption of modern quality practices. The technical goal of this project is to develop a within-layer detect-predict-decide-act loop that couples perception, modeling, and control under explicit safety and time limits. A lightweight vision model will segment and quantify recoater streaks on each layer in no more than one tenth of a second and will output geometry features with calibrated confidence. A layer-aware digital twin, implemented as a fast hybrid surrogate maps streak features, scan plan, and energy/cooling descriptors to a calibrated porosity-risk map and exposes a what-if interface that scores candidate repairs by predicted risk reduction and time cost. A minimal-intervention controller will then select among a compact action set, such as micro-remelt, selective RE-scan, modest power/speed edits, or partial recoat, while enforcing machine and thermal guardrails and defaulting to no action when expected gains are small. Performance will be validated using synchronized layer images, standard machine logs, and micro computed tomography measurements, targeting at least a forty percent reduction in streak-region porosity with no more than a three percent increase in build time. Public releases of datasets, trained models, digital twin packages, baseline decision policies, and deployable repair tools will support replication and technology diffusion across institutions and small manufacturers. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are not planned for this award.