Data-driven Predictive Models for Manufacturing Glass Fiber Composites and 3D-printed Metals Using Neural Networks and X-ray Imaging

Date
May 2023
item.page.orcid
Journal Title
Journal ISSN
Volume Title
Publisher
item.page.doi
Abstract

Advanced manufacturing requires a close monitoring of process parameters, and real-time control for rapid response to fine-tune the process conditions to produce high-quality products. While multi-physics models provide high-fidelity simulation results, the computational time involved prohibits from those to be used directly for feed-back control. The physics-based trained data- driven models have the capabilities to replicate the multiphysics simulation results at the fast speed required for design optimization and process control. The data-driven models take inputs such as component geometry, process parameters, material properties, and output outcomes in the components including temperature profile, residual stress, microstructural evolution, and material property distribution. This investigation will focus on developing data-driven models for two specific manufacturing processes, namely vacuum-assisted resin infusion molding (VARIM) and wire arc additive manufacturing (WAAM). The data-driven predictive models are established using deep machine learning (ML). Several ML models are implemented, including deep convolutional neural network (CNN), for processing spatial information; recurrent neural network (RNN), and long short-term memory (LSTM) for processing temporal information. The manufacturing of large wind turbine blades requires well-controlled processing conditions to prevent defect formation such as thermal waves. The VARIM process is the most prevalent method implemented in the industry and is often studied and optimized using the physics-based finite element models that provide accurate computational capabilities but suffer from high computational costs in the meantime. Considering the limitations, an ML approach that employs a deep CNN and RNN/LSTM model is established to predict the spatial-temporal temperature distribution during the VARIM process. The ML model is trained with the “big data” that are generated from the physics-based high-fidelity simulations, validated by a lab-scale VARIM experiment conducted in the factory setting. Once fully trained, it can provide “real time” predictions of the blade manufacturing process. Powder-based additive manufacturing (AM) process, such as direct energy deposition (DED), is widely used in fabricating metallic functional gradient materials (FGM) parts, which have mechanical properties changing with locations in a part, since multiple metal powders are mixed and used in the DED process. Hybrid manufacturing, including the DED and machining processes, to fabricate stainless steel 316L/Inconel 718 FGM specimens are experimentally studied. The molten pool evolution during the printing is observed; influences of the machining process on the printed parts due to milling, including the surface roughness, and the hardness of the specimens are evaluated. Towards the goal of sustainability and eco-friendly process for manufacturing, the wire-feed-based AM process using WAAM provides porTable freeform fabrication capability, along with precision manufacturing at both small- and large-scales. However, internal defects such as porosities are often formed in additively manufactured metal components, the defects will nucleate, grow, and coalesce to form cracks under loads, leading to eventual catastrophic failure. To understand this failure process under loading, full-field porosity evolution in a WAAM aluminum alloy cylinder under tension is observed with in-situ X-ray micro-computed tomography (μCT). The analysis is performed with the assistance of a CNN algorithm that provides rapid analysis of over thousands of slice images at various strains. The results provide quantitative evaluations of the evolution of macropores inside the WAAM specimen under tension.

Description
Keywords
Engineering, Mechanical
item.page.sponsorship
item.page.rights
Citation