AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Session VII
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Saurabh Puri, VulcanForms Inc; Francesca Tavazza, National Institute of Standards and Technology; Dennis Dimiduk, BlueQuartz Software LLC; Darren Pagan, Pennsylvania State University; Kamal Choudhary, National Institute of Standards and Technology; Saaketh Desai, Sandia National Laboratories; Shreyas Honrao, Aionics; Ashley Spear, University of Utah; Houlong Zhuang, Arizona State University
Thursday 8:30 AM
March 23, 2023
Room: Cobalt 520
Location: Hilton
Session Chair: Shreyas Honrao, Aionics; Sean Donegan, Air Force Research Laboratory
8:30 AM
Laser Powder Bed Fusion Process Design Via Machine Learning Augmented Process Modeling: Michael Groeber1; Sandeep Srinivasan1; Brennan Swick1; 1The Ohio State University
Laser powder bed fusion (LPBF) additive manufacturing (AM) is a highly active research area in the materials and manufacturing community, driven by promises of reduced lead time, increased design flexibility, and potentially location-specific process control. However, a complex processing space counters these benefits and results in difficulties when attempting to develop process parameter sets across different component geometries and sub-geometries. We develop a procedure for coupling physics-based process modeling with machine learning and optimization methods to accelerate searching the AM processing space for suitable printing parameter sets. We demonstrate the approach first on simple geometries that vary in size to show the methodology and then to a more complicated geometry to show the benefit of locally-tailored process parameters on component processing history.
8:50 AM
A Physics-based Machine Learning Study of the Hot Cracking Phenomenon in the Processes of Additive Manufacturing: Guannan Tang1; Anthony Rollett1; 1Carnegie Mellon University
The occurrence of hot cracking in the additive manufacturing process involves a variety of factors from different aspects. This leaves the effort trying to quantify hot cracking phenomena often lacking generality. Thus, unifying models that take account of processing parameters, thermodynamics, and mechanical properties remain a big gap in modeling the hot cracking phenomenon. Our current study intends to evaluate variables in different aspects but related to the hot cracking phenomenon. The topmost relevant variables will be identified and correlated with the occurrence of hot cracking through machine learning algorithms. To this end, synchrotron-based high-speed techniques together with a melt pool simulation model will be used to generate the training data. The end goal is to build up a unifying model that can predict hot cracking susceptibility based on information at different scales and aspects.
9:10 AM
Gaussian Process Ensemble Active Learning for Autonomous Parameterization of Direct Ink Write 3D Printing: Erick Braham1; Marshall Johnson2; Surya Kalidindi2; James Hardin3; 1Air Force Research Lab and NRC; 2Georgia Institute of Technology; 3Air Force Research Lab
At the frontier of additive manufacturing research there is a challenge to make complex geometries while incorporating new materials and methods into the fabrication process. Gaining precise control over a material’s behavior when introduced to a new process is often a trial and error process led by a human’s intuition. Can machines develop an “intuition” and how much data does such a process require? To answer these questions, we are developing an automated rapid exploration protocol for enabling a machine to gain its own “intuition”. This work proposes a multiple-gaussian-process based automated parameterization demonstrated with the direct ink write printing of freestanding structures. This case study examines several factors crucial to the success of this automation including efficiency, design space considerations/constraints, and use of prior data. Ideally, automated training of machine intuition will enable more agile manufacturing, high-throughput screening of materials and geometries, and more efficient collaboration across systems.
9:30 AM
Autonomous Path Planning in Additive Processes Using Semi-supervised Machine Learning: Sean Donegan1; James Collins2; Edwin Schwalbach1; 1Air Force Research Laboratory; 2The Ohio State University
As additive manufacturing (AM) continues to expand in use cases across the industrial base, increasing focus is being placed on development of novel process optimization techniques that drive improvements in throughput and performance. An area of major untapped potential is local optimization of the toolpath used in the additive process: tailored, part-specific process planning that yields desired properties as a function of geometry. In major metals AM processes such as laser powder bed fusion, scan trajectories are usually designed under simple metrics that ensure complete coverage while minimizing processing time. We propose a framework for autonomous design of AM toolpaths using semi-supervised machine learning coupled to a physics-based process model. By abstracting elements of process design into multiple “agents,” we demonstrate a model system that can generate toolpaths for arbitrary geometries from minimal training. We showcase the framework via application to several synthetic test cases.
9:50 AM
Data-enhanced Hybrid Machine Learning Model for Solid-state Friction Surfacing Process: Benjamin Klusemann1; Frederic Bock1; Zina Kallien1; Norbert Huber1; 1Helmholtz-Zentrum Hereon
Friction surfacing is a solid-state layer deposition process to additively manufacture structures of metallic materials. Since materials remain in solid state, deteriorations of microstructural characteristics and defect formations typical for fusion-based processes are circumvented. Thermal process histories can be determined with solid heat-transfer finite element models to identify and utilize relationships between process parameters and properties. Such models are computationally efficient and convenient; however, require final deposit geometries a priori, which makes experimental measurements prior to simulations almost inevitable. The aim of the proposed data-enhanced hybrid machine learning approach is to reduce the amount of those experiments to a minimum via exploiting links between process parameters, substrate and backing materials as well as maximum process temperatures and resulting deposit geometries. Initial experimental data is used to validate numerical models which are data-mined to feed a machine learning framework that can ultimately perform predictions for different process parameter and material combinations.
10:10 AM Break
10:30 AM
Developing a Physics-informed Machine Learning Model to Predict Melt-pool Shape in Additive Manufacturing: Mohammad Parsazadeh1; Sharma Shashank1; Sameehan Joshi1; Venkata mani Krishna Karri1; Narendra Dahotre1; 1University of North Texas
The laser powder bed fusion process is powder-based additive manufacturing, which is capable to produce 3D-printed parts from a 3D CAD model. Processing parameters are highly affecting the quality of the 3D-printed parts, which are linked to the shape of the melt pools. In this study, an extensive database is obtained by a single-track deposition of Ti-6Al-4V at various processing conditions. The micrographs of these single-track depositions are used to find the melt pool shapes. A set of dimensionless numbers, affecting the melt pool shapes is identified using scaling analysis. The melt pool shapes are then linked to these dimensionless numbers using a supervised machine learning model. The model is later used to predict the melt pool shape and the dominant heat transfer mode during melt pool formation.
10:50 AM
Interrelated Process-Geometry-Microstructure Relationships for Wire-feed Laser Additive Manufacturing: Sen Liu1; Craig Brice2; Xiaoli Zhang2; 1Stanford University; 2Colorado School of Mines
Wire-feed laser additive manufacturing has gained attention for years with its promises of high-level automation, reducing materials waste, overall costs, and large-scale volume production. However, printing material with multiple desired bead properties is a great challenge due to the high cost of experiments and the interdependent correlations among the properties. This study proposes a comprehensive multi-property integrated design framework based on machine learning from experiments data to enable quality control of part surface quality, geometry and microstructural characteristics. The printed bead process – microstructure – geometry relationships and process features' importance are investigated. The overall bead quality (smooth, ripple, and failed), geometry (e.g. bead height, width, fusion zone depth, fusion zone area), and microstructures are simultaneously optimized to obtain the optimal processing window. The validation results highlight that the process parameterizations by this framework can speed up the manufacturing of materials with the desired bead multi-performance.
11:10 AM
Anomaly Detection in Composite Manufacturing Using Zero-bias Deep Neural Network: Deepak Kumar1; Sirish Namilae1; Yongxin Liu1; Houbing Song1; 1Embry Riddle Aeronautical University
The manufacturing of high-performance composites is rapidly increasing to fulfill the needs of several industrial areas. Methods to detect and correct the processing defects during the composite manufacturing are needed to further expand their usage. Our novel experimental approach utilizes a custom autoclave with borosilicate glass viewports equipped with a 3D digital image correlation (DIC) system. The DIC is used to take timed pictures of the composite specimen during cure, which are analyzed using conventional and deep learning methods. A zero-bias Deep Neural Network (DNN) abnormality detection model is used to identify and categorize the anomalies that emerge during manufacturing. The model is trained and tested using the DIC images of composite structures generated during processing. The precise location and size of defects are determined using, a mask region based convolutional neural network (R-CNN) image segmentation technique. The model detects defects like wrinkles with a classification accuracy of 93.9 percent.
11:30 AM
Simulation of Mechanical Properties of TPMS-based Osteoporotic Bone by the Neural Network-Enhanced Finite Element Method: Yan-Zhen Chen1; Chu-Hao Wang1; Tsung-Yeh Hsieh1; Tsung-Hui Huang1; Cheng-Che Tung1; Po-Yu Chen1; 1National Tsing Hua University
Osteoporosis is one of the common symptoms for the elderly people around the world. Owing to the complex structures of cortical and trabecular bones, billions of elements have to be built inside the finite element analysis (FEA) model, which makes the simulation time-consuming in both modeling and solving processes. In this work, we aim to prune the miscellaneous simulation cost by integrating triply periodic minimal surfaces (TPMS) and neural network (NN) approaches. We generate a parameterized density neural network enhanced finite element method (NN-FEM), which involved representative volumetric element (RVE) construction, off-line machine learning, and deployment on a homogenized surrogate model. By assigning the TPMS parameters to this novel scheme, modeling and calculation costs can be reduced more than 100 times. The mechanical properties of normal and osteoporotic femur bones are simulated and compared. The scheme provides a new and efficient way to further mechanical properties of diseased mineralized tissues.