Additive manufacturing (AM) processes such as laser powder bed fusion (LPBF) have received a great deal of attention due to their unique capability in producing parts with complex geometries. However, a major challenge limiting the application of LPBF today is the high variability in the quality of finished components. To address this challenge, HIDENN-AI, LLC and Northwestern University (NU) propose to establish a simulation-guided machine-learning (ML) based feedback control approach to additive manufacturing of metallic aerospace components based on Laser Powder Bed Fusion (LPBF). The project goal is accomplished by building a data-driven machine learning digital twin (ML-DT) framework that integrates real-time experimental data collection/process control with machine-learning-based process modeling and optimization. This integration is critical as it bridges the gap between high-fidelity physics-based models that are known to be computationally expensive and real-time process control and optimization. Seven technical tasks are proposed in this program for the Phase I base and optional period: Task 1: Calibration and validation of an existing physics-based computational model for LPBF process (Phase I Base for HIDENN-AI, LLC) Task 2: Development of machine learning based digital twin for online monitoring and process control (Phase I Base for HIDENN-AI, LLC) Task 3: Integration of feedback control model with the hardware platform of LPBF process (Phase I Base for NU) Task 4: Demonstration of the feasibility of the developed ML-DT for process optimization through LPBF experiment of simple AlSi10Mg Aluminum coupons (Phase I Base for NU) Task 5: Phase II demonstration plan for complex geometry with select materials and Phase II proposal preparation (Phase I Base for HIDENN-AI, LLC and NU) Task 6: Develop a framework of Convolution Hierarchical Deep Neural Network (C-HiDeNN) to significantly enhance the capabilities of the DT and demonstrate the capability for modeling part-scale LPBF process (Phase I option for HIDENN-AI, LLC) Task 7: Extension of ML-DT based control of additive manufacturing process to other process conditions (Phase I option for NU) The key deliverables of the Phase I project will be a software tool and its demonstration for real-time control of the laser power signal to control the Lack of Fusion (LOF) porosity and surface roughness in LPBF processing of AlSi10Mg material system. Future Phase II efforts will be directed towards the extension of the ML-DT based control system to manufacture complex geometry with varied AM materials and improving the predictive capability of the high-fidelity solver using the novel Convolution Hierarchical Deep Neural Network.
Benefit: The additive manufacturing (AM) sector is expected to grow tremendously within the next decade. The value of additively manufactured parts is reported to rise at a 15% compound annual growth rate (CAGR) from $12 billion in 2020 to $51 billion in 2030. The largest share of the growth is contributed by the end-user markets, which include the aerospace, automotive, biomedical, and maritime industries. In the commercial sector, the AM industry is facing the critical issue of process variability that leads to quality variability of finished components, both part-to-part and design-to-design. Modeling tools that can effectively provide real-time process control to reduce variability and optimization of the AM process to eliminate defects are in high demand. By integrating mechanistic-based data science methods with real-time process monitoring/control, the machine learning based digital twin developed by HIDENN-AI, LLC effectively addresses the critical limitations in the current AM applications. As such, the developed tool will benefit the manufacturers of the AM equipment by providing state-of-the-art software for processing control and optimization, thereby giving them a competitive advantage. In turn, it will also benefit the end-users in using the optimized process for the manufacturing of high-quality components/systems with reduced scrap rates.
Keywords: Graph Neural Network, Graph Neural Network, active learning, Convolution Hierarchical Deep Neural Network, Machine Learning, laser powder bed fusion, Digital twin, additive manufacturing, Feedback Control
