3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): Development of Novel ML Methodologies III
Program Organizers: Remi Dingreville, Sandia National Laboratories; Ali Riza Durmaz, Fraunhofer Institute Iwm
Tuesday 9:00 AM
June 17, 2025
Room: Platinum Ballroom 4
Location: Anaheim Marriott
Session Chair: Sourav Saha, Virginia Tech
9:00 AM Invited
The Variational Deep Materials Network: Efficient Extrapolation With Uncertainty of Homogenized Material Responses: Andreas Robertson1; Dongil Shin2; Remi Dingreville1; 1Sandia National Laboratory; 2POSTECH
Surrogate models are a fundamental component in any machining learning framework for materials science. They provide the necessary computational efficiency for many downstream tasks, e.g., optimization in design. Importantly, useful surrogate models must be developed to account for both uncertainty and limited data. The Deep Material Network is a physics-informed machine learning framework that can stably extrapolate to predict non-linear homogenized material responses even though it is trained on only cheap elastic data. We present our extension: the Variational DMN. The VDMN naturally accounts for aleatoric microstructural uncertainty in its prediction. Importantly, this uncertainty prediction also extrapolates, allowing the VDMN to quantify uncertainty in both linear and nonlinear material responses without the need for nonlinear data. We present the algorithmic advances necessary for these changes and then present a series of examples exploring the strengths and limitations of the VDMN as a tool for accelerated uncertainty quantification in materials science.
9:30 AM
Enhanced Resolution and Image Contrast in 3D XRM Data Using Deep Learning Based Reconstruction Methods: Kaushik Yanamandra1; Hrishikesh Bale1; Rajarshi Banerjee2; 1Carl Zeiss Microscopy; 2University of North Texas
Deep learning-based 3D reconstruction methods can greatly improve the achievable resolution in image quality in 3D X-ray computed tomography compared to traditional reconstruction methods by utilizing trained models that can eliminate noise and increase contrast. Furthermore, combination with scintillator-based magnification objectives this technique can push the boundaries of achievable resolution. There have been significant developments since the early versions of deep learning based reconstruction. By adopting approaches like synthetic prior in the training step, a significant improvement in image sharpness and overall image quality along with improved throughput has been achieved. Results obtained from this new method are demonstrated on a model Ni-Ti-C based Metal Matrix Composites sample, wherein the complex microstructure and alignment of reinforcing titanium carbide (TiC) phases within the nickel matrix was revealed. This approach underscores the pivotal role of incorporating advanced deep learning-based reconstruction methods in pushing the boundaries of non-destructive 3D imaging characterizing advanced materials.
9:50 AM
MatCHMaker: Machine Learning for Microstructural Analysis in Materials Characterization Modeling: Christian Precker1; Vanessa Puertas1; Andrea Gregores Coto1; Satiago Muíños Landín1; Alexandre Ouzia2; Geoffrey Daniel3; 1AIMEN Technology Centre; 2Heidelberg Materials Global R&D; 3Université Paris-Saclay CEA, SGLS
The MatCHMaker project aims to reduce time, costs, and uncertainties in developing advanced materials, supporting the Green Deal's goal of industrial decarbonization and societal well-being. This initiative leverages materials modeling and characterization to create resilient new materials and discover novel uses for existing ones. By examining the relationships between processes, microstructure, and material properties, MatCHMaker helps predict and optimize material performance. Essential techniques, like Scanning Electron Microscopy (SEM), provide critical data on microstructures, which are analyzed using machine learning methods, including Convolutional Neural Networks and Autoencoders, to automate SEM image analysis. Focused on construction, energy, and mobility, the project integrates these ML tools to forecast materials' physical properties, ultimately developing a comprehensive materials characterization tool for industrial application. In this work, we present the progress done using ML methods to predict physical properties based on microstructure data.
10:10 AM
MLOgraphy++: A Context-Enhanced U-Net Approach for Robust Grain Boundary Segmentation in Metallographic Images: Inbal Cohen1; Julien Robitaille2; Francis Quintal Lauzon2; Ofer Beeri3; Shai Avidan1; Gal Oren4; 1Tel Aviv University; 2Clemex Technologies; 3IAEC; 4Stanford
Our work addresses the challenge of accurately identifying grain boundaries in metallographic images, where intricate texture boundaries complicate segmentation. Current state-of-the-art models, such as the Segment Anything Model (SAM), require prompts based on prior grain knowledge, limiting their usability in texture-only segmentation tasks. Manual annotation is also time-consuming and subjective. Existing methods often rely on small, annotated patches with post-processing steps, which can lead to overfitting and reduced generalizability. We introduce MLOgraphy++, a U-Net-based approach that leverages large context windows with partial labels, eliminating the need for post-processing. MLOgraphy++ effectively handles incomplete boundaries during inference by training with contextual variation. Using the Heyn intercept method as a more representative evaluation metric, we benchmark MLOgraphy++ against the state-of-the-art MLOgraphy on the Texture Boundary in Metallography (TBM) dataset, showing it achieves comparable results while enhancing generalizability and eliminating post-processing requirements.
10:30 AM Break
10:50 AM
Transfer Learning for Ultrasonic Crystallography: Accelerating Orientation Mapping in Materials With Neural Networks: Rikesh Patel1; Wenqi Li1; Richard Smith1; Matt Clark1; 1University of Nottingham
In determining the crystallographic orientation using ultrasound measurements, brute force search algorithms are used to match measured surface acoustic wave phase velocities to crystallographic orientation. This process can take hours as it is computationally intensive. We introduce a method to transfer train neural networks using calculated surface acoustic wave (SAW) phase velocities to rapidly determine crystallographic orientations. For demonstration, a model has been trained using nickel SAW phase velocities, which achieved 93.6% in validation accuracy, and applied it to classify the planes on Inconel 617 and CMX4 polycrystalline alloys. Measurements were made using the laser ultrasound technique Spatially Resolved Acoustic Spectroscopy (SRAS) and full orientation maps were achieved on 1.4Mpixels in 16 seconds, compared with the brute force searching method that requires roughly 10 hours. This method aims to be a transformative tool for materials discovery by providing near real-time microstructural insights.
11:10 AM
Synthetic Microstructure Generation and Prediction Using Stable Diffusion: Hoang Cuong Phan1; Chihun Lee1; Hoheok Kim1; Sehyeok Oh1; Ho Won Lee1; 1Korea Institute of Materials Science
Diffusion models like Stable Diffusion XL (SDXL) excel in generating high-quality images but face challenges in adapting to specific domains with limited data, such as microstructure generation in materials science. This study presents a novel SDXL-based framework, optimized with Low-Rank Adaptation (LoRA) and DreamBooth techniques, to both generate and predict microstructure images across seen and unseen experimental parameters with minimal computational demand. By selectively fine-tuning the UNet and text encoders, with targeted modifications and optimal hyperparameters, our method accurately captures intricate microstructure characteristics. It enables controlled image generation across varied process parameters, such as temperature, annealing time, and cooling methods, thereby reducing the need for additional experimentation. Rigorous evaluations demonstrate that our approach outperforms benchmarks in both image quality and fidelity to real microstructures. This scalable strategy addresses data scarcity and costly experimentation, enabling extensive, high-quality dataset generation with predictive capabilities applicable to broader scientific domains.
11:30 AM
Texture Evolution Surrogate for Magnesium Materials: Kyle Farmer1; Elizabeth Holm1; 1University of Michigan
Understanding and predicting texture evolution in materials under deformation is essential for designing materials with targeted mechanical properties. Crystal plasticity simulations offer a robust method for predicting texture changes by modeling slip system activation and strain hardening behavior at the grain scale. However, the computational demands of crystal plasticity limit its application, especially for large-scale or high-throughput analyses. In this work, we present a deep learning-based surrogate model trained on finite element method (FEM)-based crystal plasticity simulations where magnesium - an underexplored material in texture surrogate development - polycrystals are deformed following an arbitrary strain path. Our model maintains high accuracy with a significant computational speedup over FEM simulations and demonstrates excellent scalability with respect to the system size and simulation settings.
11:50 AM
Topological Analysis of Processing to Microstructure Mappings: Zachary Varley1; Megna Shah2; Jeff Simmons2; Veera Sundararaghavan3; 1Carnegie Mellon University; 2Air Force Research Laboratory; 3University of Michigan
The relationship between processing parameters and resultant microstructures remains a fundamental challenge in materials science. We explore a neural network approach to map between processing parameter space and microstructure morphology space, with particular emphasis on quantifying topological faithfulness in latent representations. We evaluate several complementary approaches for measuring topology preservation, combining techniques from computational topology, manifold learning, and dimensionality estimation. These metrics assess both local neighborhood structure and global topological features of the paired point clouds representing processing conditions, input data, and learned latent representations. From a materials development perspective, maintaining accurate topological relationships in the learned representations could enable more reliable navigation of processing parameter space and facilitate materials optimization. We compare the effectiveness and computational efficiency of these metrics in the context of autoencoder architectures, providing insights for developing more robust computational tools for materials design and processing-structure relationship analysis.