3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): Poster Session
Program Organizers: Remi Dingreville, Sandia National Laboratories; Ali Riza Durmaz, Fraunhofer Institute Iwm
Tuesday 4:40 PM
June 17, 2025
Room: Platinum Ballroom 6
Location: Anaheim Marriott
A Novel Approach for Input-Output Modeling and Optimization of EBW on HY282 Alloy Using Artificial Neural Networks and Metaheuristic Algorithms: Anupam Kundu1; Mrinmay Pal1; Dilip Pratihar1; Debalay Chakrabarti1; 1IIT Kharagpur
Electron Beam Welding (EBW) was employed to join HY282 alloy. However, modeling and optimization of EBW of HY282 alloy have not been reported yet. In this study, a novel and robust approach has been proposed for the input-output modeling and optimization of EBW of HY282 alloy, using Artificial Neural Networks and metaheuristic algorithms. The output parameters were the Ultimate Tensile Stress (UTS) and Microhardness (MH). The present study implemented four ANN models trained using Genetic Algorithm (GA), Bonobo Optimizer (BO), Gray Wolf Optimizer (GWO), and Particle Swarm Optimization (PSO), respectively. Among these models, ANN trained by GWO was found to provide the best predictions for the EBW weld attributes, with an average absolute deviation of 2.8%. Further optimization of the ANN model to maximize the UTS, PSO yielded the highest UTS, around 849.7 MPa. Finally, confirmatory test results closely matched the simulated ones.
An Optimized Neural Network for Detecting Complex Glass Defects in Automated Inspection: Shihyu Chen1; Yen-Hsiang Chen2; Fong-Ji Tsai1; 1National Yunlin University of Science and Technology; 2yuntech university
Glass defect detection is critical in manufacturing, yet traditional machine vision methods struggle with complex and subtle defect patterns. This study proposes a deep learning-based approach using a concatenated hybrid network to enhance defect classification accuracy. By integrating efficient feature extraction with deep feature reuse, the model effectively captures intricate defect variations.The proposed network is evaluated against state-of-the-art deep learning architectures using a curated glass defect dataset. Performance metrics—including accuracy, precision, recall, and confusion matrices—assess each model’s effectiveness. Experimental results demonstrate that the hybrid network outperforms conventional CNNs, achieving superior classification robustness and reliability. This research advances automated glass inspection by introducing a deep learning framework capable of handling complex defects, thereby enhancing manufacturing efficiency and quality control.
Computer Aided Design and Mesh Modeling Automation for FEA and Developing Surrogate Model: Dongwoon Han1; Hyo Kyu Kim1; Seongtak Kim1; 1Korea Institute of Industrial Technology
Designing an appropriate shape that meets the required performance of a product is crucial for ensuring its quality. This paper describes an automated method for design, finite element modeling, and finite element analysis (FEA) data collection, which are essential for developing surrogate models for optimal shape design. An algorithm was developed to automatically generate design models using a design program API based on the factors proposed through the design of experiments. Another algorithm was developed to perform automatic mesh modeling and apply analysis conditions tailored to the model. Subsequently, the process of conducting FEA, automatically collecting analysis results, and organizing them into a dataset for surrogate model training is detailed. As a result, this study significantly reduced the time required for the data collection process necessary for surrogate model development, and it is expected to be utilized for optimal design using surrogate models in the future.
Deep Learning-Based Analysis of XRD Patterns for SOEC Cell Performance Evaluation: Hunmin Park1; Sun-Dong Kim1; Yoonseok Choi1; 1Korea Institute of Energy Research
X-ray diffraction (XRD) is a crucial tool for characterizing crystallinity, but conventional methods struggle to analyze complex material compositions in solid oxide electrolysis cells (SOECs). Recently, artificial intelligence (AI), particularly convolutional neural networks (CNNs), has shown promise in enhancing XRD pattern interpretation.In this study, we applied deep learning techniques to analyze XRD data from SOEC cells after 1,000 hours of operation. CNN-based models effectively detected subtle structural changes and established correlations with performance degradation. Our results demonstrate that deep learning enables meaningful insights beyond conventional analysis, offering a powerful tool for SOEC material diagnostics. While AI applications in SOEC research remain limited, our study highlights its potential for improving material characterization and performance evaluation, paving the way for advanced analytical methodologies in the field.
Deep Learning for Industrial-Scale Modeling of the Basic Oxygen Furnace Process: Maryam Khaksar Ghalati1; Zhou Daniel Hao1; Jianbo Zhang1; Hongbiao Dong1; 1University of Leicester
The basic oxygen furnace (BOF) steelmaking process is a cornerstone of modern steel production, where accurate modeling is critical for optimizing operations, improving process control, and enhancing energy efficiency. This study investigates the application of advanced deep learning models, including deep transformers, to model the BOF process. Representing the first extensive deployment of such architectures on BOF operational data, we evaluated multiple state-of-the-art models using a comprehensive dataset comprising over 10,000 samples from a large-scale industrial setting. Our research introduces novel approaches tailored to the complexities of BOF data, leveraging insights from exploratory data analysis to enhance predictive performance. The proposed models demonstrated improvements in accuracy compared to traditional methods, highlighting the transformative potential of deep learning in optimizing industrial processes.
Development of Milling State Assessment System Using Machined Surface Images and Engineer's Sensory Evaluations as Supervised Data: Ryo Tanaka1; Tatsuya Furuki1; 1Chubu University
Milling manufacturing for products such as mold or medical products requires high-quality machined surfaces, engineers have achieved stable and efficient production by visually evaluating the appearance of machined surfaces and tools. However, such tacit intellectual skills are difficult to pass on. In this study, we attempted to develop a digital triplet-typed machining condition evaluation system based on a machine learning algorithm with excellent human interpretability, such as decision trees, using images of machined surfaces and the engineer's sensory evaluation as training data. Furthermore, milling produces scaly cutting mark patterns, although if variations of cutting mark shapes occur, the product's appearance like the glossiness deteriorates. Therefore, the developed system recognizes the shape of cutting marks and matches the pattern to evaluate the stability of the machined surface. This system enables unskilled engineers to know the results of evaluating superior engineers simply by taking pictures of the cutting surface.
Cancelled
Ontology-Based Digital Representations of Materials Testing in the MaterialDigital Initiative: Hossein Beygi Nasrabadi1; Birgit Skrotzki2; Harald Sack1; 1FIZ Karlsruhe — Leibniz Institute for Information Infrastructure; 2Bundesanstalt für Materialforschung und -prüfung (BAM)
The MaterialDigital (PMD) platform has been funded by the German Federal Ministry of Education and Research (BMBF) by 2019. The platform aims to digitalize materials and processes including the provision of infrastructures to represent complete material lifecycles, considering the FAIR principles (discoverable, accessible, interoperable, reusable). In recent years, PMD achieved lots of progresses for developing the data ecosystem for digital materials research. The fundament of the data ecosystem is ontology-based digital representations of materials characteristics. In the current research, we present the methodology and toolchains for the development of domain-level ontologies for materials testing that address the requirements of materials testing standards. The collection of the required terminology from the testing standard, the semantic representation of the process graphs, the conversion of the ontology files, their integration with the upper-level ontologies, and the data mapping processes were presented for the several mechanical testing use cases.
Optimal Design Method Using High-Fidelity Surrogate Modeling Based on Finite Element analysis Data: Seongtak Kim1; Dongwoon Han1; Hyokyu Kim1; 1Korea Institute of Industrial Technology
This study proposes an optimal design method for photovoltaic (PV) module frames by developing a deep learning model based on finite element analysis (FEA) data. For decades, PV modules have maintained standardized structures, resulting in similar frame designs. However, the increasing size of PV cells, the emergence of building-integrated photovoltaic (BIPV) modules with diverse shapes, and the shift away from aluminum frames to reduce carbon emissions necessitate customized frame designs. To address this need, we apply a deep learning-based high-fidelity surrogate model with over 99% accuracy for frame optimization. This approach enables the design of lighter frames while enhancing performance compared to conventional designs.
Optimization of Machining Conditions for Improved Machining Quality Based on a Digital Twin of Machine Tools: Beomsik Sim1; Jae-Eun Kim1; Wonkyun Lee1; 1Chungnam National University
Optimizing machining conditions is essential for enhancing machining quality. This study proposes an approach to optimize machining conditions using a digital twin of a machine tool. The digital twin integrates key components, including the controller, feed drive system, and physical models based on cutting theory to estimate machining quality. It enables virtual experiments that incorporate the machine’s dynamic behavior and control characteristics, offering a more accurate representation of the machining process. Based on these virtual experiment results, optimization techniques adjust the feed rate and spindle speed. This approach provides precise optimization by calculating machining quality as influenced by control characteristics and dynamic behavior, which traditional methods, such as design of experiments or basic models, cannot account for. The method is validated by comparing machining quality and machining time between non-optimized and optimized conditions, along with assessing the accuracy of the digital twin’s predictions.Acknowledgment: This work was supported in part by the Technology Development Program for Smart Controller in Manufacturing Equipment (No. 20012834, Development of Smart CNC Control System Technology for Manufacturing Equipment) funded by the Ministry of Trade, Industry & Energy (MOTIE, Korea) and in part by a Korea Institute for Advancement of Technology (KIAT) grant funded by the Korean Government (MOTIE) (No. P00020616, The Competency Development Program for Industry Specialist).
Process Control System and Cloud Simulation Platform for Continuous Casting Based on Digital Twin Technology: Cheng Ji1; Bochun Liang1; 1Northeastern University
In this work, a digital twin system is architected which predicts in real time the 3D temperature field, full-section reduction deformation and micro-structure evolution of the continuous casting. In temperature prediction, thermophysical parameters can be loaded in real time based on the specific location of the billet in 3D space, which ensured the prediction accuracy of solidification process. For deformation prediction, based on three-dimensional thermal/mechanical coupled FEM data-driven and mechanism-driven, a model for online prediction of deformation characteristics during the continuous casting process is established, with the calculation process employing nodal stress accumulation to determine pressure force. In micro-structure evolution, the precipitation of Mns, Ti(C, N) and ect were predicted by coupling microegregation with saturated solid solution product . In the practice, the digital twin system can control of process parameters rapidly, and it was appied in slab and bloom continuous casting process control more than 15 casters.
Quantum Machine Learning: Revolutionizing Fatigue Fracture Mode Prediction of Laser Powder Processed Inconel 625: Ravindranadh Bobbili1; Chitti Babu N1; 1DRDO
The present study employs a quantum machine learning approach to predict the fatigue test failure of laser powder processed Inconel 625 alloy. In this research, four machine learning models were applied for fatigue fracture modeling. The machine learning algorithms (logistic regression, random forest, gradient boosting and QLattice) are trained and tested for fracture prediction. The input features are porosity (%), alternated stress, cycles, defect size (Area 1/2), respectively. The QLattice, gradient boosting and random forest models have the best performance as far as ROC-AUC curves are concerned. In addition, the shapley additive explanations (SHAP) is introduced to improve the interpretability of model. This cutting-edge approach offers unparalleled efficiency, accuracy, and cost-effectiveness compared to traditional experimental testing methods. Its ability to harness the power of quantum computing to predict fatigue fracture of laser powder processed Inconel offers immense potential for improvingmaterials design and manufacturing processes.
Real-Time Data-Driven Optimization System for Hot-Rolled Coil Cooling Placement: Chihun Lee1; 1Korea Institute of Materials Science
Hot-rolled coils (HRCs) play a crucial role in diverse industries such as automotive, construction, and machinery. However, the cooling process of HRCs within storage yards often leads to nonuniform cooling due to complex thermal interactions between adjacent coils and variable environmental conditions, directly affecting the mechanical properties and overall steel quality. In this study, simplified heat transfer models based on the finite element method (FEM) were employed to generate realistic cooling scenario data. To address the computational limitations inherent in FEM, a novel management system integrating two artificial neural networks (ANNs) with deep and wide architectures, optimized through hyperparameter tuning, was developed. This system predicts temperature variations at multiple locations on coil surfaces in real-time, enabling strategic placements to minimize temperature disparities and enhance cooling uniformity. This real-time computational approach eliminates the need for additional cooling equipment while ensuring high-quality products.
Unsupervised Learning for Low Dimensional Corrosion Quantification of Aluminum Films: Sarah Firestone1; Nathan Brown1; David Montes de Oca Zapiain1; Aditya Venkatraman1; 1Sandia National Laboratories
Environmental corrosion at small length-scales can be assessed by the manual evaluation of topographical images of component surfaces using an Oxford-AFM. However, this process is time-consuming and subject to bias introduced by the researcher. In this work we address these challenges by developing an efficient and data-driven analysis of the images. Our proposed solution involves using unsupervised learning to identify trends of dynamic surface corrosion activity based on 2D topographical images of the samples subjected to harsh environments. Specifically, we leveraged computational techniques such as Generalized Extreme Value Distribution, binarization, spatial correlations, Principal Component Analysis, and K-Means clustering. Our protocols demonstrate excellent efficacy in identifying the evolution of corrosive features, even for previously unseen images. As a result, we have successfully established a continuum of corrosive states within the latent space, which allows for rapid, preemptive identification and facilitates localized mitigation. SNL is managed and operated by NTESS under DOE-NNSA-contract-DE-NA0003525.SANDNo.SAND2024-14510A