Session Sheet - 3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): Applied ML for Manufacturing III

3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): Applied ML for Manufacturing III
Program Organizers: Remi Dingreville, Sandia National Laboratories; Ali Riza Durmaz, Fraunhofer Institute Iwm

Tuesday 9:10 AM
June 17, 2025
Room: Elite Ballroom 1 & 2
Location: Anaheim Marriott

Session Chair: Ashley Lenau, Sandia National Laboratories

9:10 AM
Application of ML to Ceramics Industry: Dominic Wadkin-Snaith¹; ¹Lucideon
    The traditional more established Ceramics industry is energy intensive with 30% of production costs coming from energy consumption alone. Although established, additionally, the industry can suffer from uncontrolled deformation of parts under firing. Whereas nascent techniques such as 3D printing of complex and intricate parts are in their development stage and need greater optimisation before coming to commercial market. We summarise recent advances made by the application of ML techniques to data collected via process sensors to drive towards an industry 4.0 approach enabling optimisation of industrial processes within traditional ceramics sector. As well as the latest in non-destructive computer vision techniques to monitor part distortion in real time for firing profiles of interest. We will also demonstrate how ML enhancement of design of experiment has led to new direction of development in the area of ceramic 3D printing.

9:30 AM
Understanding Manufacturing and Materials Design Spaces: Overcoming Expensive Data: Erick Braham¹; James Hardin²; ¹UES / Air Force Research Laboratory; ²Air Force Research Laboratory
    Materials discovery, manufacturing optimization, and many other goals in our community have benefited using data-driven methods like machine learning, design of experiments and AI. A large driver in the implementation of these tools is to understand complex design spaces from small amounts of data. Materials and manufacturing research lends itself inherently to expensive data with costly materials, researcher time, machine time, or other limited resources. In this work we explore ways to sample intelligently to provide the most meaningful design information in the most efficient manner. By examining two case studies we can demonstrate approaches that enable little waste when exploring opaque design spaces. Firstly we examine the use of an ensemble gaussian process approach on parameterizing direct ink write 3D printing. Secondly we examine using chemical and physical properties to guide selection of formulations of compositionally complex ceramics.

9:50 AM
Correlation of ULTEM 9085 Physical, Chemical, and Mechanical Properties: Kassandra Hernandez¹; Devin Roach¹; ¹Oregon State University
    ULTEM 9085 is the first Federal Aviation Administration (FAA) qualified 3D printed thermoplastic, well-known for its high strength-to-weight ratio, chemical resistance, and flame retardance. However, a large difference in mechanical properties exists between traditionally manufactured high-performance thermoplastics and 3D printed materials. This study investigated the effects of print parameters on material properties through density, surface profile, scanning electron microscopy (SEM), and void analyses across various coupon geometries and build orientations. Surface microscopy findings revealed that inter- and intra-layer bonding significantly influences density and tensile performance variations. Additionally, a machine learning (ML) model was developed to analyze tensile-tested ULTEM 9085 coupons, with 3D printing parameters as inputs and critical mechanical properties—such as strength, modulus, and yield strength—as outputs. The ML model uncovered previously unknown relationships between printing parameters and mechanical properties. Overall, this research provides essential insights for optimizing additive manufacturing processes for high-performance aerospace applications.

10:10 AM  Cancelled
Surrogate Modeling of Cluster Dynamics-Predicted Nucleation and Growth of Irradiation Defects Using Time-Series Neural Networks: Sanjoy Mazumder¹; Andrea Jokisaari¹; ¹Idaho National Laboratory
    A time-series based neural network model is presented, to predict the evolution of irradiation defects in structural materials, for application in modern fast reactors. Mean-field cluster dynamics (CD) has been extensively used to investigate the kinetics of nucleation and growth of extended defects, i.e., dislocation loops and voids in materials, under specific irradiation conditions. CD is computationally expensive to predict the population of TEM-observable large defects. The interaction of irradiation defects with microstructural features like grain-boundaries and network dislocations increases the model complexity. Also, the input parameter space expands to define the unirradiated microstructure. We have performed selection of input parameters based on sensitivity analysis of the CD predictions. A long short-term memory network (LSTM) has been trained with the CD-predicted time-series data, for the chosen parameters, to capture the temporal evolution of defects. High-throughput CD simulations were performed to generate the training and validation dataset for the LSTM model.

10:30 AM Break

10:50 AM
Hierarchical Bayesian Modeling for Enhanced Contamination Detection in Electron Beam Powder Bed Fusion Processes: Temilola Gbadamosi-Adeniyi¹; Tim Horn¹; ¹North Carolina State University
    Contamination in Electron Beam Powder Bed Fusion (EB-PBF) critically affects material integrity and component performance. Leveraging advancements in Total Electron Emission (TEE) data, which reveal material composition, contrast, and topography, this study introduces a hierarchical Bayesian model to detect and quantify contamination across all EB-PBF layers. The model proficiently distinguishes between the presence and absence of contamination at each spatial location and layer. Incorporating a Gaussian Process (GP) captures spatial correlations, ensuring the generation of coherent contamination maps throughout the build. Preliminary results demonstrate the model's effectiveness in accurately identifying contamination-free regions, reliably detecting intentionally introduced contamination, and quantifying contamination spread in subsequent layers. Posterior predictive checks validate the model’s robustness, showing strong alignment between simulated TEE signals and observed data. This hierarchical Bayesian framework offers a scalable and interpretable solution for enhancing process monitoring and quality control in additive manufacturing processes.

11:10 AM
Offline Programming for Wire-Arc DED Applications Using Machine Learning Algorithms: Kaue Riffel¹; Rakhi Bawa¹; Antonio Ramirez¹; Justin Chan¹; Rida Adhami¹; Daniil Gofman¹; ¹The Ohio State University
    With the expansion of Wire-Arc DED, optimization of the deposition process is crucial for efficiency. Optimal stepover distance and printing strategy is instrumental to producing a successful build, yet existing methods use low-flexibility models or rely on human experience, resulting in costly and inefficient trial-and-error methods. This project develops an artificial intelligence to identify the optimal stepover distance from input process variables like wire feed speed and travel speed. Using non-linear regression models on experimental data of overlapping bead depositions, hundreds of synthetic data points are simulated to train a neural network. The network then identifies the best stepover distance by optimizing multiple output variables, like height difference between beads and the area of the valley of the deposition. Additional parameters, like waveform and deposition area, are also incorporated, creating a versatile tool tailored to specific applications. A graphical user interface further enhances usability for efficient DED programming.