2024 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2024): Special Session: Big Data Analytics II
Program Organizers: Joseph Beaman, University of Texas at Austin
Tuesday 8:00 AM
August 13, 2024
Room: 616 AB
Location: Hilton Austin
Session Chair: Srikanthan Ramesh, Oklahoma State University
8:00 AM
AMUI: A Data-driven Additive Manufacturing User Interface for Process Optimization: Peter Pak1; Chakradhar Guntuboina1; Odinakachukwu Ogoke1; Achuth Chandrasekhar1; Olabode Ajenifujah1; Amir Barati Farimani1; 1Carnegie Mellon University
Rapid prototyping in many additive manufacturing (AM) applications requires immediate, point-of-need access to the optimal process configurations for a specific build. Therefore, we introduce the Additive Manufacturing User Interface (AMUI), an intuitive end-to-end user interface for automatically optimizing process parameters for 3D printing with minimal requirements for user domain knowledge. It will enable individuals without prior technical expertise to upload 3D computer-aided-design models and determine the necessary processing conditions for their builds. Using surrogate models and analytical solutions, we automatically predict the optimal printing parameters for processes such as Laser Powder Bed Fusion (LPBF) and effectively mitigate potential sources of defects. Specifically, we leverage surrogate models of the temperature fields produced during AM processes in combination with physics-based simulations to create custom process maps for a specific build configuration. We also discuss extensions to other areas of the AM workflow enabled by the modular design of the interface.
8:20 AM
Exploring Machine Learning Classification of Porosity from Infrared Signatures During Laser-powder Bed Fusion: Matthew Roach1; Leah Jacobs1; Grant Wilmoth1; Brett Brady1; Caleb Campbell1; Bradley Jared1; Anahita Khojandi1; 1University of Tennessee, Knoxville
Internal defects are commonly formed during laser powder bed fusion (L-PBF) additive manufacturing of components. One common defect is porosity which is normally found after the manufacturing process using methods such as computed tomography (CT) or destructive sampling. Prior art in this field used costly cameras to explore specific defects in defined locations within a sample. This research uses real-world computed tomography data and a lower resolution infrared (IR) camera. Machine learning is used to relate the layer-wise IR images to layer-wise CT porosity to shed light on the in-situ conditions in PBF-LB additive manufacturing that result in non-defective parts.
8:40 AM
Predicting Print Morphology in Aerosol Jet Printing Using Deep Learning-based cGANs: Shihab Shakur1; Akash Deep1; Srikanthan Ramesh1; 1Oklahoma State University
Aerosol jet printing (AJP), a micro-scale additive manufacturing technique, is predominantly utilized for producing flexible electronic devices. In AJP, the morphology of the printed features contain rich information about the process phenomena. Early prediction of these morphological signatures can provide opportunities for process monitoring and mitigating suboptimal process conditions, thereby minimizing defect occurrence. Consequently, a scalable and efficient data-driven model is essential to effectively address this issue. This research presents the development of a deep learning-based conditional generative adversarial network (cGAN) model to predict and emulate the morphological signatures in AJP. The model generates images of print morphologies conditioned on the process parameters. It allows for early prediction of future morphological signatures for an in-process part, based on process parameters. The efficacy of the model is demonstrated in a case study on AJP of conductive circuits using a poly(3,4-ehtylenedioxythiophene) (PEDOT) and polystyrene sulfonate (PSS) ink (PEDOT:PSS) on flexible substrates.
9:00 AM
Generative Artificial Intelligence (GenAI) Prompt Engineering for Additive Manufacturing (AM): Nowrin Surovi1; Paul Witherell1; 1National Institute of Standards and Technology (NIST)
Additive manufacturing (AM) faces several challenges in achieving efficient and defect-free printing. Although traditional machine learning (ML) has proven effective in mitigating these challenges, it requires specialized models for solving specific problems with limited scopes. Generative artificial intelligence (GenAI) holds promise as a versatile tool capable of addressing multiple issues simultaneously, leveraging its expansive training data and robust problem-solving capabilities. However, getting the desired output from GenAI relies heavily on crafting effective prompts, as incorrect formulation of prompts can lead to unexpected responses. Prompt engineering is crucial for GenAI models to produce desired outputs efficiently. In our study, we explore how different prompt techniques affect the responses of GenAI tools in addressing AM problems. We examine five popular prompt engineering methods: Zero-shot, Few-shot, Chain-of-shot, React, and Directional Stimulus Prompting. We also use well-known GPT-4 models to evaluate these responses across various AM metrics.
9:20 AM
Generative AI for AM Materials Optimization and Design: Patxi Fernandez-Zelai1; Jason Mayeur1; Jiahao Cheng1; Guannan Zhang1; Neil Zhang1; Amirkoushyar Ziabari1; Saket Thapliyal1; Rangasayee Kannan1; Peeyush Nandwana1; 1Oak Ridge National Laboratory
Materials inverse design problems are essential towards advancing various technologies from fuel cells to fusion materials. Purely experimental discovery is extremely laborious; physics-based computational routes are generally limited to solving forward problems. Machine learning based generative models are well suited for data fusion and, critically, enable inverse solutions. Very recently denoising diffusion probabilistic models have exploded in popularity with a number of successful materials specific applications. Here we explore the viability these models for various tasks ranging from structural grain-scale optimization to binder jet AM composition design. These case studies demonstrate that these models are extremely flexible optimization tools capable of various constrained optimization tasks. The probabilistic nature of these models also makes them well suited for quantifying uncertainty. We envision that future materials design frameworks will make extensive use of these models as ``search'' tools bolstering the utility of experimental and computational approaches.
9:40 AM Break
10:00 AM
Ontology-based Retrieval Augmented Generation (RAG) for GenAI-Supported Additive Manufacturing: Yeun Park1; Paul Witherell1; Nowrin Akter Surovi1; Hyunbo Cho2; 1National Institute of Standards and Technology; 2Pohang University of Science and Technology
Conventional data analytics often fail to capture the intricate context of Additive Manufacturing (AM) processes, leading to pointed solutions and suboptimal analytics outcomes. The performance of Generative AI (GenAI) models, such as Large Language Models (LLMs), largely depends on their ability to integrate and contextualize the vast data they are trained on. However, contextualizing is often directly driven by the data consumed, and not necessarily grounded in the fundamental truths. To address this issue, an ontology-based retrieval augmented generation (RAG) approach is proposed to enhance GenAI's capability to generate pertinent prompts and answers. The GenAI recognizes and applies relevant context by leveraging structured ontology, resulting in accurate and insightful interpretations. A use case showcases how the proposed ontology-based RAG framework operates to provide context-aware AM data analytics that promote analytical transparency through fundamental truths when executing AM data analytics.
10:20 AM
Applying Generative Deep Learning Models for Cost-Effective Monitoring and Simulation of LPBF Processes: Odinakachukwu Ogoke1; Quanliang Liu1; Sumesh Kalambettu Suresh1; Jesse Adamczyk2; Dan Bolintineanu2; Olabode Ajenifujah1; Alexander Myers1; Guadalupe Quirarte1; Anthony Garland2; Jack Beuth1; Jonathan Malen1; Michael Heiden2; Amir Barati Farimani1; 1Carnegie Mellon University; 2Sandia National Laboratories
The stochastic formation of defects during Laser Powder Bed Fusion (LPBF) introduces variation in the mechanical and fatigue properties of printed parts. Generative deep learning models can be used to achieve insights into the distribution of defects produced during these processes, reducing both the cost of simulation-based forecasting and the cost of experimental monitoring of defect formation. We demonstrate this through two case studies. Initially, we develop deep learning models to stochastically upscale low-fidelity multiphysics simulations of the melting process to their high-fidelity counterparts, identifying keyhole-forming behavior while bypassing the runtime required for high-fidelity simulations. Subsequently, we apply generative models to link low-cost, low-resolution, layer-wise optical images of the build plate to detailed high-resolution images of the build plate, enabling cost-efficient layer-wise process monitoring. We evaluate the performance of these models by analyzing the statistical properties of the generated samples in addition to the preservation of key LPBF-based metrics.