2023 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2023): Data Driven Modeling and Process Control
Program Organizers: Joseph Beaman, University of Texas at Austin
Tuesday 1:40 PM
August 15, 2023
Room: 415 AB
Location: Hilton Austin
Session Chair: Chinedum Okwudire, University of Michigan
1:40 PM
A Data-driven Surrogate Model for Time-dependent Scanwise Thermal Simulations of Laser Powder Bed Fusion Parts: Berkay Bostan1; Shawn Hinnebusch1; David Anderson1; Albert To1; 1University of Pittsburgh
The intricate thermal conditions that arise throughout the laser powder bed fusion (LPBF) process depend on the local geometry and have significant impact on the part quality in terms of defect formation, microstructure, and residual distortion. However, due to the large length scale difference between the laser beam (µm) and part (cm), scanwise simulations are intractable for part-scale models due to the immense computational expense. In this study, a surrogate model which represents time-dependent, part-scale scanwise thermal simulations is proposed. The developed architecture consists of a sequence of deep neural networks (DNNs), and long short-term memory (LSTM) units. The input vector encompasses information related to geometric, thermal conditions, scanning strategy, and simulation parameters. A 100x speedup is achieved by the proposed model, thus enabling defect formation and microstructure simulations of centimeters scale parts.
2:00 PM
A Machine Learning Approach to Part Scale Microstructure Predictions in LPBF: Mason Jones1; Brian Weston2; Theron Rodgers3; Daniel Moser3; Jean-Pierre Delplanque1; 1University of California Davis; 2Lawrence Livermore National Laboratory; 3Sandia National Laboratories
Laser powder bed fusion additive manufacturing uses microscale processes to create macroscale parts with complex microstructures. This makes part-scale modeling and simulation of microstructures challenging and expensive. A newly developed machine learning based surrogate microstructure model for this process has been shown to significantly accelerate predictions of microstructure statistics while maintaining accuracy. This work uses ensemble modeling in conjunction with this surrogate model to identify and explore viable routes for scaling this surrogate model to part scale predictions for the purposes of process parameter optimization. This will include analysis and refinement of the surrogate model as well as statistical analyses to identify shortcuts to part scale modeling.
2:20 PM
A Physics-guided Data-driven Model for Enhanced Temperature Prediction and Control of LPBF Additive Manufacturing: Cheng-Hao Chou1; Chinedum Okwudire1; 1University of Michigan
A hybrid (i.e., physics-guided data-driven) model is proposed for temperature prediction and control of laser powder bed fusion (LPBF) additive manufacturing. Parts produced by the LPBF process are subject to deformation or other defects due to the thermal behavior during the manufacturing process, which can be resolved by model-based control techniques. However, existing temperature prediction models are either inaccurate or computationally costly, and hence are not suitable for closed-loop control. To overcome these deficiencies, the authors propose a linear hybrid model, which cascades a physics-based finite difference method (FDM) model with a linear data-driven model that aims to correct the temperature prediction of the FDM model by learning the unmodeled dynamics. By simulation, the proposed hybrid model is shown to achieve significant improvement in the temperature prediction accuracy compared to the baseline model (i.e., the physics-based FDM model).
2:40 PM
A Scientific Artificial Intelligence (Sci-AI)-based Concurrent Multiscale Simulation Framework for Accurate Temperature Prediction of Large-scale Metal Additive Manufacturing: Lin Cheng1; 1Worcester Polytechnic institute
The broad applications of metal AM are hindered by the variability of mechanical performances due to time- and space-dependent thermal history. Although simulation models of multiple scales have been developed, either of them is able to provide accurate thermal history of large-scale production. This work aims to develop a concurrent multiscale framework capable of capturing the detailed melt pool dynamics in part-scale analysis. A Sci-AI model is to incorporate the in-situ data into the melt pool dynamics simulation for more accurate and efficient analysis. The computational cost of the Sci-AI model is at the microseconds level, allowing real-time coupling with part-scale analysis. This makes it possible to accomplish more accurate thermal history with detailed thermal fluid information for the part-scale fabrication and lays the foundation for concurrent multiscale modeling of metal AM process. Several numerical examples will be conducted to illustrate the performance of the proposed framework.
3:00 PM
System Identification of Fused Filament Fabrication Additive Manufacturing Extrusion and Spreading Dynamics: Osama Habbal1; Ali Kassab1; George Ayoub1; Pravansu Mohanty1; Christopher Pannier1; 1University of Michigan Dearborn
In fused filament fabrication additive manufacturing, polymer extrusion and spreading dynamics affect build quality in both surface finish and mechanical properties. However, the nonlinear extrusion and spreading dynamics are ignored by slicers. To advance the aim of slicing using accurate nonlinear dynamic models, this work presents a system and procedure for automated measurement of dynamic bead extrusion. The system uses a belt printer, iFactory3D One Pro, with nozzle tilted 45 degrees from the build belt, and a snapshot 3D scanner. Single layer prints in polylactic acid (PLA) are scanned and then automatically ejected. The G-code for the L-shaped print holds the gantry speed fixed while the extrusion flow rate is varied as a square wave signal in space. The experiment design matrix varies three variables: extrusion temperature, gantry speed, and extrusion flow rate. Time constants are fitted to bead width signal extracted from the scan data to obtain nonlinear models.