Physics-based optimization of Scan Strategies in Laser Powder Bed Fusion (LPBF) additive manufacturing process

Laser powder bed fusion (LPBF) is an additive manufacturing technique increasingly used to produce metallic parts, though it often results in residual stress, deformation, cracks, and other quality defects due to uneven temperature distributions. To address these issues, an intelligent method for determining laser scan sequences in LPBF, called SmartScan, was applied to simple 2D geometries previously. In this work, we extended SmartScan to arbitrary 3D geometries by expanding the thermal model and optimization approach to multiple layers, enabling processing of arbitrary shapes and infill patterns, balancing exploration and exploitation in the optimization process, and improving computational efficiency using model order reduction. Simulations and prints of 3D test artifacts using SmartScan showed reductions in temperature inhomogeneity (92%), residual stress (86%), maximum deformation (24%), and geometric inaccuracy (50%), without significantly affecting print speed. The study also demonstrated the practical integration of SmartScan as a plug-in to commercial slicing software, significantly enhancing printed part quality.

Enhanced Milling Stability Predictions Using Ensemble Transfer Learning with Experimental Data

In this work, we developed a novel approach for updating model-based stability chart predictions in milling using experimental data. This method leverages Deep Neural Networks (DNNs), pre-trained with simulated data generated by predicting machine dynamics through receptance coupling and evaluating stability via an analytical stability model. The DNNs are fine-tuned using a small experimental dataset of only a few dozen samples to align network predictions with experimentally observed stability states under various cutting conditions. This approach bypasses the need for measuring or estimating cutting force coefficients, tooltip dynamics, or extensive model parameter identification, making it highly suitable for industrial applications. Experimental validation shows that this method accurately predicts stability charts for different engagement conditions and tool clamping lengths. The use of an ensemble learning method, combining predictions from multiple networks, enhances prediction accuracy. Additionally, this new approach requires about five times fewer experimental samples than previously proposed model-free machine learning methods to achieve the same prediction accuracy on a test set.

Detection of workpiece inhomogeneity using advanced milling process monitoring approaches

The intricacies of manufacturing models result these models to be highly susceptible to process variatons. In this work, we proposed an approach to monitor such a variation - workpiece inhomogeneity, using machine tool's controller signals and machine learning algorithms. Steel specimens are end-quenched to introduce a controlled hardness gradient. This gradient is detected from the controller signals without the need for additional sensors. It is demonstrated that adjusting process parameters can prevent chatter caused by increasing hardness and enhance productivity by modifying parameters according to local material hardness. to.

Instructor:

• College of Engineering, Michigan State University, East Lansing, MI
  
  • ME 391 Mechanical Engineering Analysis: Spring 2024, Fall 2024

Graduate Teaching Assistant:

• Department of Mechanical and Process Engineering, ETH Zurich, Switzerland
  
  • Engineering Materials and Production I: Fall 2014, Fall 2015, Fall 2016
  • Manufacturing II: Spring 2014, Spring 2015, Spring 2016

• Mechanical Engineering Department, Koc University, Istanbul, TURKEY
  
  • Dynamics: Spring 2011, Spring 2012
  • Computer Integrated Manufacturing and Automation: Spring 2013