8th World Congress on Integrated Computational Materials Engineering (ICME 2025): Artificial Intelligence and Machine Learning in ICME II
Program Organizers: Victoria Miller, University of Florida; Stephen DeWitt, Oak Ridge National Laboratory
Wednesday 2:10 PM
June 18, 2025
Room: Platinum Ballroom 5
Location: Anaheim Marriott
Session Chair: David Montiel, University of Michigan
2:10 PM Invited
Recent Advances in Structure-Property Correlations and Their Impact on ICME-Driven Accelerated Design of Materials and Products: Amarendra Kumar Singh1; Satyam Shukla1; 1Indian Institute of Technology Kanpur
The Integrated Computational Materials Engineering (ICME) approach has transformed the design, development, and deployment of materials and products by enabling the prediction and optimization of composition, microstructure, and processing paths to meet specific property and performance targets, which may include strength, ductility, and corrosion resistance, depending on the application. Structure-property relationships used in ICME workflows are primarily based on experimental data, which significantly limits the capability for inverse design. Recent advances in modeling and simulations, high-throughput experimentation, and machine learning-based tools provide a promising alternative approach. Data from ab initio calculations, crystal plasticity, and representative volume element (RVE)-based simulations supplement the experimental data, now enhanced through high-throughput experiments. Advanced machine learning models are employed to develop predictive tools that capture complex relationships among composition, microstructure, and properties. This enables a more robust platform for inverse design of materials and products. A few illustrative examples will be presented.
2:40 PM Cancelled
Exploring Novel Alloys With Superior Specific Hardness Using Data-Driven Approaches: Taeyeop Kim1; Wook Ha Ryu2; Geun Hee Yoo2; Donghyun Park1; Ji Young Kim2; Eun Soo Park2; Dongwoo Lee1; 1Sungkyunkwan University; 2Seoul National University
The discovery of advanced alloys using experimental data-driven approaches is hindered by the challenges of managing large compositional design spaces and the risk of overfitting machine learning (ML) models. This study combines ML predictions with thin-film base high-throughput experimental verification to accelerate the discovery of novel ternary alloy systems with exceptional specific hardness. By applying ensemble learning to a dataset from combinatorial experiments, we efficiently explored a composition space involving 28 metallic elements and discovered tens of new compositions exhibiting superior specific hardness compared with previously reported alloys. The property was consistently observed in 2 mm thick ribbon samples, demonstrating scalability. Explainable AI revealed that elemental dissimilarities significantly enhance solid-solution strengthening and phase formation, offering key insights into the underlying mechanisms. This iterative ML-driven process provides a reliable approach for discovering high-performance alloys and could serve as a useful framework for future materials development.
3:00 PM Break
3:30 PM
Deep Learning-Based Platinum Particle Analysis for Corrosion Insights in BWR Systems: Txai Sibley1; Ryan Jacobs2; Dane Morgan2; Kevin Field1; Elizabeth Holm1; 1University of Michigan; 2University of Wisconsin-Madison
This research investigates the use of deep learning-based image analysis models to detect platinum particles on Boiling Water Reactor (BWR) system components. These platinum particles, added through noble metal chemical addition (NMCA), are essential for preventing stress corrosion cracking by controlling water chemistry, thereby extending the lifespan of reactor parts. The study utilizes a segmentation model trained on a limited set of images to analyze platinum particles on reactor surfaces. By linking microstructural attributes with electrochemical potential (ECP), we aim to better understand corrosion behaviors and refine NMCA effectiveness evaluation. This work lays the groundwork for improving segmentation accuracy, reducing data collection and annotation costs, and deepening our comprehension of how platinum particles influence BWR reactor performance.
3:50 PM
Multi-Modal Machine Learning Framework for Property Prediction in Ni-Based Superalloys and Aluminum Alloys Using Integrated Characterization Data: Jiwon Park1; Chang-Seok Oh1; 1Korea Institute of Materials Science
Multi-modal machine learning approaches have emerged as powerful tools for integrating diverse materials characterization data to advance materials discovery and optimization. In this study, we present a novel framework that combines microstructural images, X-ray diffraction patterns, compositional data, and processing parameters to predict properties of Ni-based superalloy and aluminum alloys. Our model architecture employs parallel neural network branches to process each data modality independently before fusion: convolutional neural networks for microstructure image analysis, XRD data without peak and phase assessments, and fully connected layers for composition and process variables. The fused representation enables both property prediction and interpretable feature importance analysis across modalities. This work demonstrates the potential of multi-modal learning to leverage complementary materials characterization techniques for enhanced materials informatics and accelerated alloy development.
4:10 PM
Machine Learning Models for Predicting 3D Microstructure--Property Relationships From 2D Images: Guangyu Hu1; Sheila Whitman1; Marat Latypov1; 1University of Arizona
Engineering properties of structural alloys depend on both alloy composition and 3D microstructure. Modeling the engineering properties requires access to 3D microstructure information, which can be obtained either from direct (yet expensive) 3D experiments or reconstruction from 2D section(s). In this contribution, we present machine learning (ML) approaches to modeling effective properties of heterogeneous materials directly from 2D sections. To this end, we consider statistical learning models based on spatial correlations and microstructure representations obtained with vision transformers as well as deep learning with convolutional neural networks. We train all models on data obtained from micromechanical 3D simulations. Upon training, the presented models only need 2D sections as input, whose experimental acquisition is much more accessible compared to 3D characterization.
4:30 PM
Machine Learning for the Computational Design of Single-Crystal Superalloys: Matthieu Degeiter1; Edern Menou2; Armand Barbot1; Patricia Klotz1; Sami Ben Elhaj Salah2; Didier Locq1; Yohan Cosquer3; Mikael Perrut1; 1ONERA; 2SAFRAN TECH; 3DGA TA
High Pressure Turbine (HPT) blades in gas turbines are designed with single crystal nickel-base superalloys, as their microstructure provides the blades with exceptional mechanical properties at high temperature. In service, the thermomechanical loadings imposed on the blade induce microstructure evolutions which eventually degrade the blade macroscopic properties. Improving the blade service life requires to develop new superalloy grades with improved creep resistance at high temperature. The fundamental mechanisms underlying the macroscopic behavior of materials are often complex, take place over extended ranges of length and time scales, and are strongly nonlinear. When the equations driving these mechanisms are not known, data-driven methods are particularly effective in guiding metallurgists and engineers. In this context, our work focuses on the construction of gaussian process models based on experimental data to estimate the creep life of superalloys as a function of their chemical composition and testing conditions.