AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Poster Session
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
Tuesday 5:30 PM
March 21, 2023
Room: Exhibit Hall G
Location: SDCC
M-4: A Deep-learning Enabled Reliability Enhancement System for the Fused Deposition Modelling Process: Xiao Shang1; Xingchen Liu1; Jiahui Zhang1; Qiyan Mao2; Yu Zou1; 1University of Toronto; 2Crescent School
Fused deposition modelling (FDM) is an additive manufacturing (AM) technique that is prevalently applied in various industries such as aerospace, medical, and consumer goods. However, the low reliability hinders its advancements for mass production. In this work, we present an innovative reliability enhancement system (RES-FDM) that improves the systematic performance of FDM printers. RES-FDM uses a camera to capture real-time images of prints, which are inferred by a Deep Learning (DL) and a post-processing algorithm to terminate the printing process on detecting failures such as stringing and warping. The DL model is trained with a large dataset consisting of over 9000 experimentally collected images, achieving an accuracy of 95%. Compared with printers without RES-FDM, the printing failure rate is reduced from 16% to 4%. The RES-FDM proposed in this work shows significant improvements in the reliability of the FDM technology, paving the way for its applications for mass industry production.
Active Learning of Powder Milling Machine for Optimized Silicon Particle Size Control: Jong Ho Kim1; 1Research Institute of Industrial Science and Technology
Silicon powder has recently started to be used as a raw material for anode materials for lithium ion batteries. Silicon powder has a higher capacity than graphite, but has a brittle characteristic, so there is a limit to its usage. One of the prerequisites for applying the silicon powder is to control the particle size of the powder. In this study, a traditional milling device and machine learning were combined to control the particle size of silicon powder. In particular, in order to obtain a target particle size, an efficient optimization method was proposed by applying Bayesian inference and active learning methods. It was confirmed that the combination of traditional experimental equipment and machine learning is an efficient silicon powder particle size control method.
Adaptive Learning from Scarce and Multi-Fidelity Data: Amin Yousefpour1; Mehdi Shishehbor1; Zahra Zanjani Foumani1; Ramin Bostanabad1; 1University of California Irvine
Materials modeling and design are typically hampered by two major uncertainty sources: lack of data and model form discrepancies. In this talk, we will present a novel approach based on nonlinear manifold learning that addresses these (and more) uncertainty sources. Our approach is based on latent map Gaussian processes (LMGPs) and aims to leverage multiple data sources to quantify uncertainty sources and, more importantly, provide visually interpretable diagnostic measures that indicate the extent to which different data sources (e.g., experiments, simulations, etc.) agree with one another. We will demonstrate that our approach performs well in a wide range of applications where there may be missing data, calibration parameters, or source-dependent noise.
M-5: Advanced Analytics on 3D X-ray Tomography of Irradiated Silicon Carbide Claddings: Fei Xu1; Joshua Kane1; Peng Xu1; Jason Schulthess1; Sean Gonderman1; 1Idaho National Laboratory
Silicon Carbide (SiC) ceramic matrix composite (CMC) cladding is currently being pursued as one of the leading candidates for accident tolerant fuels. Material properties and defects were only measured on the material surfaces which cannot provide essential measurements to evaluate the material performance. Computed tomography (CT) is available at INL’s Materials and Fuels Complex, which can provide data-rich and non-destructive characterization. In this work, we presented a novel analytic method to unveil the defects in SiC CMC cladding and the discussion of the results from multiple samples. Moreover, the implementation of a new model to detect fiber toes structure on reconstructed 3D images is also introduced. Lastly, conclusive findings are included in this work. In the future, more transmission electron microscopy (TEM) experiments based on small-scale samples will be conducted to figure out the reasons causing cracks/defects, and a physic-informed model to predict crack propagation will be developed.
M-6: Design of Casting-friendly TiAl Alloy by Artificial Neuron Network: Yu-Jen Tseng1; Hong-Yuan Sun2; Yi-Hsuan Sun1; Cheng-Hsueh Chiang2; Hung-Wei Yen1; 1National Taiwan University; 2Metal Institute Research & Development Centre
TiAl alloys have been considered promising superalloys for aerospace and automobile industry. However, its fluidity limits its castability, alloy design and applications. Moreover, casting practice or related research is not attractive to young scientists. This study introduces artificial neuron network to disclose relationship between alloy composition and fluidity, which measured by spiral fluidity tests. Alloying features such as latent heat, peritectic solid fraction, superheat, and kinematic viscosity are obtained from calculated phase diagram and Scheil solidification. These features and processing variables act as inputs to build an ensemble model with single hidden layer. This model named feature model, which is coded with the relationship between alloy features and fluidity. Then, another model named composition model is trained with virtual data given by feature model, which provides relationship between composition and fluidity. This work demonstrates an approach to predict fluidity of TiAl and an integration of casting practice and artificial intelligence.
Effective Bulk Properties and Structure-property Relationships in Additively Manufactured Metal with Micron- and Nanometer-scale Structural Complexity: Mir Al-Masud1; Ryan Griffith1; Naji Mashrafi1; Mujan Seif1; Matthew Beck1; 1University of Kentucky
Modeling additively manufactured materials is challenging due to their inherent complexity and structural randomness, which is typically present at multiple length scales. The recently developed Kentucky Random Structures Toolkit (KRaSTk) implements a high-throughput approach to generate and compute properties of model representative volume elements (mRVEs) based on physics-based geometric seed descriptions capturing relevant structural complexity. We have applied this technique to compute bulk effective properties and distributions of size-scale-dependent local properties for a novel 3D-printed metal alloy that exhibits structural complexity at both the 100 nm and 10 micron scale. Automated data-mining approaches have been used to extract structure-property relationships governing effective bulk elastic properties.
Glass Forming Ability of Silica Glasses with Machine Learning Based Prediction Technique: Jong Ho Kim1; 1Research Institute of Industrial Science and Technology
Silica glass is the most common and widely used material. In particular, when a specific element is added to silica glass, it forms a glass ceramic and has low thermal expansion characteristics, so it is used in various applications. For the properties of silica glass, amorphization is important and can be predicted through the glass forming ability. It is difficult to obtain sufficient data to measure the glass forming ability because it is experimentally based on various compositions and cooling conditions. Therefore, many researchers have conducted research to predict the forming ability through several physical properties. In this study, we tried to predict the glass forming ability using machine learning techniques and an open database. By applying several machine learning techniques, the accuracy was compared and the optimal model was selected. Based on this result, a composition that can be manufactured more easily and can maintain amorphous properties is suggested.
M-7: How Can I Use Machine Learning to Predict all the Process Parameters that will lead to a Specific Material Property in my Advanced Manufacturing Process?: Lizzy Coda1; Loc Truong1; Colby Wight1; WoongJo Choi1; Tegan Emerson1; Keerti Kappagantula1; Henry Kvinge1; 1Pacific Northwest National Lab
The relationship between process parameters and properties in advanced manufacturing is often many-to-one; there are many different process parameters that can yield the same property combination in a component. While standard deep learning architectures are well-adapted to predicting properties given process conditions (learning in one direction), the task of predicting process parameters that will yield a single material property combination is significantly more challenging. However, being able to efficiently and systematically solve this problem in the absence of large amounts of data would be a significant benefit to the materials and manufacturing community since this addresses experimental design and process optimization. In this presentation, we describe how a recent deep learning-based approach to this problem, Bundle Networks, can be adapted to better model this challenging problem.
How Do You Optimize Your Parameters? Realistically Complex Hyperparameter Optimization of 23 Parameters of a Black Box Function over a Realistically Low Budget of 100 Iterations: Sterling Baird1; Marianne Liu2; Taylor Sparks1; 1University of Utah; 2West High School
Expensive-to-train deep learning models can benefit from an optimization of the hyperparameters that determine the model architecture. We optimize 23 hyperparameters of Compositionally-Restricted Attention-Based Network (CrabNet), over 100 adaptive design iterations using two models within the Adaptive Experimentation (Ax) Platform. This includes a recently developed Bayesian optimization (BO) algorithm, sparse axis-aligned subspaces Bayesian optimization (SAASBO), which has shown exciting performance on high-dimensional optimization tasks. Using SAASBO to optimize hyperparameters, we demonstrate a new state-of-the-art on the experimental band gap regression task within the materials informatics benchmarking platform, Matbench (∼4.5 % decrease in mean absolute error (MAE) relative to incumbent). Characteristics of the adaptive design scheme as well as feature importances are described for each of the Ax models. SAASBO has great potential to both improve existing surrogate models, as shown in this work, and in future work, to efficiently discover new, high-performing materials in high-dimensional materials science search spaces. (https://github.com/sparks-baird/crabnet-hyperparameter)
M-9: Loss Curvature-informed Multi-property Prediction for Materials and Chemicals via Graph Neural Networks: Alex New1; Michael Pekala1; Nam Le1; Janna Domenico1; Christine Piatko1; Christopher Stiles1; 1Johns Hopkins Applied Physics Laboratory
Properties of interest for crystals and molecules, such as band gap, elasticity, and solubility, are related to each other: they are determined by the same underlying physics. However, when state-of-the-art graph neural networks predict multiple properties simultaneously via multi-task learning, they frequently underperform a suite of single property predictors. This suggests graph networks may not be fully leveraging these underlying property similarities. Here we investigate a potential explanation for this phenomenon – the curvature of each property’s loss surface significantly varies, leading to ineffective learning. This difference in curvature can be assessed by looking at spectral properties of the Hessians of each property’s loss function, which is done in a matrix-free manner via randomized numerical linear algebra. We evaluate our hypothesis on two benchmark datasets (Materials Project for crystals and QM8 for molecules) and consider how these findings can inform the training of novel multi-task learning models.
M-10: Modeling the Phase Transition of 2-D Magnetic Materials under the Effects of External Parameters Uncertainty: Md Mahmudul Hasan1; Zekeriya Ender Eger1; Arulmurugan Senthilnathan1; Pinar Acar1; 1Virginia Tech
Magnetic materials are essential to the advancement of industrial and scientific development. They are often utilized in electronic devices, analog and digital data storage, and medical devices, However, when the ferromagnetic to paramagnetic phase transition occurs at the critical temperature, the magnetic property of these materials is substantially reduced. The Ising model explains the magnetic phase transition. A phase transition zone rather than a critical point results from the uncertainties related to the ambient temperature and magnetic field. However, previous studies neglected the long-range spin interactions at Ising modeling considering only the nearest neighbor spin interactions. Moreover, the uncertainty of the external parameters and their effects on the magnetic phase transition has never been studied. This study investigates the effects of external parameters uncertainty on ferromagnetic-paramagnetic phase transition of 2-D magnetic materials. An analytical formulation of uncertainty quantification along with computational approaches will be applied to improve computational time requirements.
M-11: Optimized Print Parameter Prediction by Machine Learning: Kevin Graydon1; Yongho Sohn1; 1University of Central Florida
While laser powder bed fusion offers the opportunity for novel engineering designs, determination of material printability through optimization studies remains expensive in terms of time and cost. Machine learning algorithms such as neural networks have been employed to minimize these two parameters and to predict the optimal print parameters for a given material. Utilizing our wide, in-house, print parameter dataset of ferrous and non-ferrous alloys and ICME tools such as Thermo-Calc, models are trained on the thermophysical properties, experimental print parameters, and their resulting relative densities. Modelling with thermophysical properties allows for use not only with current engineering alloys but also extends capabilities to future novel and made-for-AM materials. In this way, print parameters granting fully dense parts can be predicted thereby reducing development time and cost.
M-12: Scaling Microstructure-dependent Mechanical Properties to Bulk Material Properties Using 3D Convolutional Neural Networks: Laura Vietz1; Carter Cocke1; Ashley Spear1; 1University of Utah
Experimentally characterizing bulk mechanical behavior for certain materials, including nuclear materials used in harsh operating environments, can be challenging and cost prohibitive. In such cases, characterization efforts might be limited to small-scale testing, which can exhibit size effects for sample sizes below that of a representative volume element (RVE), or the smallest volume of material above which a property of interest converges to that of bulk material. Volumes smaller than the RVE, called statistical volume elements (SVEs), exhibit scattered responses. This research aims to link microstructure-dependent SVE-derived properties to bulk material properties by coupling high-throughput numerical simulation with machine learning. SVE microstructures are simulated using an elasto-viscoplastic fast Fourier transform code. Three-dimensional images of the SVE microstructures and their corresponding mechanical responses are used to train a convolutional neural network to predict bulk mechanical response. This research could enable cost-effective methods for characterizing bulk materials by testing micro/nanoscale specimens.
M-13: Synthetic Data-assisted Unsupervised Domain Adaptation for Hierarchical Microstructure Reconstruction: Ali Durmaz1; Muhammad Karim1; Oleg Shchyglo2; Akhil Thomas1; Chris Eberl3; 1Fraunhofer IWM; 2Ruhr-Universität Bochum; 3University of Freiburg
In materials science, a variety of full-field simulation tools are available which provide solid estimates of crystal growth or damaging processes. While these rule-based approaches currently do not model relations comprehensively, they can provide synthetic and low-cost data to train statistical models along with real-world data. The presented work utilizes synthetic data from phase-field simulations of martensite formation along with corresponding experimental micrographs to improve the analysis of the latter. Specifically, the image segmentation of prior austenite grain boundaries is tackled, which has implications for a range of different materials properties such as crack growth in fatigue.