Session: Government Agency Student Posters

Paper Number: 173787

Porosity Evaluation and Reduction in Laser Powder Bed Fusion via Deep Learning on In-Process Layer Surface Measurement Data From Fringe Projection Profilometry 

Laser Powder Bed Fusion (LPBF) is a widely used additive manufacturing process capable of producing complex metal components with fine resolution. However, porosity remains a critical concern, compromising the structural integrity and mechanical performance of printed parts. Pore formation is driven by complex laser-material interactions, including repeated melting-solidification cycles and random spattering, which create surface features such as ridges, lumps, and craters across in-process layers. These surface irregularities influence subsequent powder spreading, fusion, and melt pool fluctuations, leading to defects like pores and weak inter-layer bonding. Existing research primarily relies on layer-wise optical imaging with basic machine learning approaches that analyze image intensity values, often neglecting the impact of quantitative surface characteristics (e.g., roughness, height of lumps, depth of cavities) on porosity formation.

To address this gap, we have developed a custom LPBF-specific Fringe Projection Profilometry (FPP) system, enhanced with super-resolution machine learning, to capture high-resolution layer-wise surface topography in real time. By correlating in-situ FPP measurements with ex-situ X-ray computed tomography (XCT)-characterized porosity data, deep learning models are trained to predict porosity defects based on surface topography inputs. These models identify potential defect-prone regions, providing feedback on defect location and size, which can inform process optimization strategies to mitigate porosity. Building on the predictive insights delivered by the process–property model, a series of laser-rescanning optimization trials are performed, aiming at enhancing surface integrity and suppressing defect formation in real time. First, anomalous features from the correlation model are fed into a rescan-planning algorithm that (i) classifies each defect by type and severity, (ii) evaluates a library of candidate repair strategies—including scan strategy and processing parameters—and (iii) solves a multi-objective optimization problem to select the laser path, power, and scanning speed that jointly minimize the characterized defects. The algorithm is developed based on a series of empirical laser rescanning results which act as the plant model. Results demonstrate the closed-loop potential of our framework: real-time multimodal sensing feeds a predictive neural network, whose outputs in turn inform adaptive process control, closing the gap between monitoring, diagnosis, and remediation.

By fusing high-frequency, multimodal sensing with physics-informed machine-learning analytics and optimization-driven laser control, our framework propels laser-powder-bed fusion toward the intelligent manufacturing paradigm. The closed-loop architecture delivers immediate gains in dimensional accuracy, surface finish, and defect suppression while continuously curating a rich, provenance-tagged data stream that captures the ultrafast process dynamics in the build. This digital thread not only underpins rigorous, model-based qualification protocols but also feeds life-cycle analytics that predict maintenance needs and component fatigue. As a result, the platform accelerates the certification of new alloys and complex geometries, lowers iteration costs, and furnishes the traceability demanded by mission-critical sectors where reliability and regulatory compliance are necessary.

Presenting Author: Haolin Zhang University of Pittsburgh

Presenting Author Biography: Haolin Zhang is a graduate research assistant in Mechanical Engineering at the University of Pittsburgh, where he develops real-time monitoring methods, physics-informed process models, and optimization strategies for additive manufacturing. His research spans laser powder bed fusion, binder jetting, and vat photopolymerization, with a common goal of linking in-situ sensor data to predictive control algorithms that boost build quality, reliability, and throughput.

Authors:

Haolin Zhang University of Pittsburgh
Md Jahangir Alam University of Pittsburgh
Alexander Caputo Georgia Institute of Technology
Chaitanya Vallabh University of Pittsburgh
Richard Neu Georgia Institute of Technology
Xiayun Zhao University of Pittsburgh