Machine Learning Based Data Driven Prediction of Process-Induced Porosity in LPBF Using CT Scan Data and Thermal Modeling

Abstract Laser Metal Additive Manufacturing (MAM) offers a unique opportunity to produce complex parts with internal structures, enabling tailored mechanical and functional properties. Laser Powder Bed Fusion (LPBF) is widely used to fabricate intricate geometries, including porous structures that serve diverse applications. In dense components, porosity is typically minimized to enhance mechanical performance. However, in specific applications, controlled porosity can be beneficial such as increasing permeability for filtration, enhancing osteointegration in biomedical implants, or reducing strength to mitigate stress shielding effects. Porosity in LPBF parts can be introduced either through designed lattice pore structures or process-induced porosity. The latter, controlled at the melt pool level, is particularly challenging to predict due to complex interactions of melt pool dynamics, surface tension effects, and localized instabilities, leading to striation and pore agglomeration at melt pool nodes. Accurate prediction of the resulting porous structure under different processing conditions is crucial for optimizing the functional performance of LPBF components. This study presents a data-driven modeling approach for rapid and accurate prediction of process-induced porous structures in LPBF. A deep learning framework integrating Convolutional Neural Networks (CNN) and Multi-Layer Perceptron (MLP) is developed to predict structural features based on process parameters. The model is trained and validated using CT scan slice data extracted from multiple regions of fabricated samples, ensuring a comprehensive representation of pore structures. To streamline data extraction, a custom Python script is developed to automate CT scan processing. This enables efficient handling of large datasets while maintaining accuracy in capturing structural details. The extracted data is converted into structured 256 × 256 grayscale images. To further enhance predictive accuracy, a conduction-based thermal model using ANSYS is employed to supplement the dataset with simulated results. This physics-based model provides insights into heat distribution, offering a more comprehensive understanding of process-induced porosity. By incorporating both experimental CT data and simulation results, this hybrid approach creates a robust dataset for modeling the interactions governing porosity formation in LPBF. One key advantage is its ability to overcome the limitations of experimental data, which are often constrained by specific process conditions. Integrating simulation-derived data extends the model’s applicability across a wider process space and improves predictive performance. The results demonstrate a highly accurate and computationally efficient framework for mapping process-structure relationships in LPBF-fabricated 17-4 PH stainless steel components. This study supports the optimization of process conditions, enabling improved control over porosity and enhancing the functional performance of additively manufactured porous structures. The proposed AI-driven model significantly reduces computational costs and experimental effort, providing a scalable approach for process optimization in metal additive manufacturing.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.