Session Sheet - 3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): Development of Novel ML Methodologies II

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

Monday 1:30 PM
June 16, 2025
Room: Platinum Ballroom 4
Location: Anaheim Marriott

Session Chair: Katherine Colbaugh, Leucite; Stephen Price, Worcester Polytechnic Institute

1:30 PM  Invited
Machine Learning-Driven Discovery of High-Hardness Multi-Principal Element Alloys With Physics Informed Priors: Eddie Gienger¹; Maitreyee Sharma Priyadarshini²; Jarett Ren³; Paulette Clancy³; ¹Johns Hopkins Unviersity - Applied Physics Laboratory; ²Virginia Tech; ³Johns Hopkins University
    Multi-principal element alloys (MPEAs) are known for their exceptional mechanical properties and thermal stability. However, traditional discovery methods are limited in their success given MPEAs complex compositions. Here, the previously developed Physical Analytics Pipeline (PAL 2.0), a Bayesian optimization framework employing active learning in a closed-loop setting is used to accelerate discovery. PAL 2.0 integrates Gaussian process modeling with experiments, guiding synthesis within an informed, optimized space. Through three experimental cycles, 20 novel MPEAs were synthesized. Samples in previously unexplored phase diagrams were characterized, achieving Vickers hardness values up to 1269, a 7% increase over previous benchmarks. A striking discovery is the appearance silicon and tantalum together, a combination not seen in the training dataset. The successful identification of new high-hardness alloys demonstrates PAL 2.0’s potential to optimize MPEAs efficiently. This approach presents a solution for the exploration of high-dimensional material systems, underscoring the framework's adaptability for advanced materials discovery.

2:00 PM
A Coupled Thermal-Mechanical Deep Material Network: Ashley Lenau¹; Dongil Shin²; Andreas Robertson¹; Ricardo Lebensohn³; Remi Dingreville¹; ¹Sandia National Laboratories; ²Pohang University of Science and Technology; ³Los Alamos National Lab
     Deep material networks (DMN) are tree-like machine learning networks that train on linear homogenization relationships for a given microstructure, and then act as a representative volume element to predict non-linear material responses. Most DMNs are utilized for “single physics” problems, such as a DMN predicting the thermal or mechanical response using its predicted homogenized thermal conductivity or stiffness, respectively. However, a DMN architecture or training strategy involving multi-physics homogenization tasks is still not yet well established. Combining thermal and mechanical tasks will ultimately result in a better description of the deformation behavior of the composite for a wider variety of boundary conditions. In this study, a DMN is simultaneously trained to homogenize the stiffness and thermal conductivity of a composite. After training, the network extrapolates stress, strain, heat flux, and temperature gradient for non-linear thermomechanical relationships. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525.

2:20 PM
A Thermodynamically Consistent Neural Ordinary Differential Equation for Constitutive Modeling of Polycrystalline Metals: Paul Christodoulou¹; Liam Mackin¹; David Najera-Flores¹; Rohan Patel¹; Bradley Davidson¹; Jarred Heigel¹; Elliot Haag¹; Reed Kopp¹; ¹ATA Engineering, Inc.
    It is difficult to formulate a robust constitutive model for path-dependent, nonlinear, heterogeneous materials, such as additively manufactured (AM) metals. The difficulty arising from material heterogeneity is compounded by competing material behaviors at various length scales. ATA recently developed a Neural Ordinary Differential Equation (NODE) constitutive model with multiplicative Deformation Gradient Decomposition (DGD) to enable automatic discovery of these path-dependent constitutive models from stress-strain data. DGD-NODE ensures a thermodynamically consistent set of internal state variables and state variable evolution laws to describe the elastic, plastic, and thermal components of deformation. Presented work includes training and verification of the DGD-NODE constitutive model using simulations of synthesized microstructures representative of an AM metal, and subsequent application of the constitutive model in AM process modeling simulations using the Multiphysics Object Oriented Simulation Environment (MOOSE) finite element solver.

2:40 PM
Physical-Informed Machine Learning for Silicone Formulation Development: Qingtao Cao¹; Lalitha Raghavan¹; Sheng Zhao¹; Ying Wang¹; ¹Saint-Gobain Research North America
    The application of machine learning (ML) in formulation development has been demonstrated through various Design of Experiment (DOE) cases. However, in the early stages of development, the limited number of available formulations restricts the effectiveness of purely data-driven models, particularly when developing new formulations that extend beyond the property ranges of existing ones. To overcome this challenge, Physics-Informed ML, integrating domain knowledge from physical models with the latent patterns captured by data-driven models, shows a more robust approach. In this case study, we present a successful implementation of Physics-Informed ML, leveraging a random forest model with packing value theory to develop silicone formulations characterized by two properties traded off with each other, based on a small set of formulation available within a narrow property range. The efficacy of this hybrid approach is further highlighted through a comparative analysis, demonstrating the superior performance of Physics-Informed ML over the purely data-driven method.

3:00 PM Break

3:30 PM
Out-of-Distribution Surface Anomaly Detection Using Masked Autoencoder Vision Transformers: Pierre Belamri¹; Henry Proudhon¹; David Ryckelynck¹; Damien Texier²; ¹Mines Paris - PSL University; ²Institut Clément Ader
     Automated surface-anomaly detection using machine learning has become a promising area of research with high impact on visual inspection. However, traditional supervised models rely on large, labelled datasets, making them difficult to apply in this context. Large Vision Transformers, leveraging masked image modelling, offer a solution by creating modality-agnostic latent spaces that can enhance multimodal materials characterization. We propose a self-supervised learning approach for out-of-distribution anomaly detection using EBSD orientation maps, a key modality in material science as it provides direct information on grains orientations. A Vision Transformer Masked Autoencoder is trained to reconstruct quaternions maps, learning latent representations that capture grain boundaries and orientation patterns.Our results show clear differences in Mahalanobis distances distributions for twin-free and twin-containing samples, indicating the model’s ability to detect anomalies linked to sigma-3 grain boundaries and/or twin geometries. This demonstrates the potential of multimodal transformers for defect detection in polycrystalline materials.

3:50 PM
Rapid Bubble and Cavity Tracking Utilizing Machine Learning: Riley Wheeler¹; Christopher Field²; Khalid Hattar¹; ¹University of Tennessee; ²Theia Scientific LLC
    Bubble to cavity evolution in materials could greatly influence the mechanical stability of materials utilized in extreme environments, such as nuclear reactors. Nanoscale bubble counting is typically done by hand using transmission electron microscope (TEM) micrographs and annotation software in a non-scalable, post-acquisition workflow. Recently, machine learning (ML) models have been utilized to count and measure bubbles in a fraction of the time compared to manual annotation. A rapid neural network option is the You Only Look Once (YOLO) model. Factors that influence the model include: data volume and training time through epochs. This technique has been applied to many different materials, such as Ni, PdNi, and LiAlO3 with some degree of success. The selection of model development parameters will be explored to determine the optimal setting for time efficient ML model-based bubble counting.

4:10 PM
Prediction of Creep Life Using K-Means Clustering and Gaussian Process Regression: Sami Ben Elhaj Salah¹; Edern Menou¹; Matthieu Degeiter²; Armand Barbot²; ¹Safran; ²ONERA
     In this study, a model involving Gaussian Process Regression (GPR) is introduced to predict creep life based on chemical compositions of nickel-based superalloys, utilizing a large and heterogeneous dataset. The high variability in the data introduces significant challenges for model accuracy; thus, a clustering approach using K-means is used to split the entire dataset into more similar sets.A separate GPR model is trained to capture the specific data patterns within each cluster. These individual kernels are then combined into a global kernel, representing the cumulative predictive power across clusters. This method enables a more adaptive modeling approach, improving predictive accuracy by leveraging the local similarities within each cluster. The proposed framework provides a robust tool for predicting creep life time in order to guide novel alloy design.

4:30 PM
Multidimensional Analysis for Correlation of Mechano-Physico-Chemical Attributes With Bio-Functionality in Eight TiNbZrSnTa (TNZST) Alloys: Carmen Torres-Sanchez¹; Paul Conway¹; ¹Loughborough University
    Data-enabled approaches with experiments allow exploration of the inter-play between chemical, physical and mechanical attributes of alloys with their impact on biological behaviours. Characterisation of bulk and surface properties has been performed on eight exemplar TiNbZrSnTa of different allotropes. Their β or β+α′ (α″) crystal structure deﬁned their mechanical properties, with dual β+α″ more suitable for load-bearing applications. As these are intended for implantation, surface assessment was a primary focus. A statistical correlation was undertaken between oxide layer composition, layer thickness, surface free energy and contact angles, and behaviour of cells progressing through adhesion, attachment, proliferation, differentiation and to maturation. Pairwise Pearson coefﬁcient correlations, visualised in a heat-map matrix, allows elucidation of inferred agnostic correlations and supported the study of primary physical and chemical attributes most affecting cell behaviour; enabling both identiﬁcation of primary contributors to osteoblastogenesis and, which experimental tests reveal stronger predictors for osteogenesis in these TNZST alloys.