8th World Congress on Integrated Computational Materials Engineering (ICME 2025): Artificial Intelligence and Machine Learning in ICME I
Program Organizers: Victoria Miller, University of Florida; Stephen DeWitt, Oak Ridge National Laboratory

Tuesday 1:30 PM
June 17, 2025
Room: Platinum Ballroom 1
Location: Anaheim Marriott

Session Chair: Xiawa Wu, Penn State Behrend

1:30 PM  Invited
Alloy Design for Additive Manufacturing: Expanding the Scope of Alloy 3D Printing for Resource-Constrained and Location-Specific Applications: Wei Xiong1; 1University of Pittsburgh
    Additive manufacturing becomes critical in advanced materials processing, allowing us to produce complex shape components. This study presents a novel approach to alloy design for additive manufacturing in resource-constrained environments. Utilizing a mixture of commercially available stainless steel 316L and Inconel 718 powders, the research demonstrates the creation of a functionally graded alloy through directed energy deposition. Post-heat treatment processes are optimized using Calphad-based ICME modeling, resulting in an alloy with properties comparable or superior to pure Inconel 718 at a reduced cost. A high-throughput heat treatment method is introduced, significantly accelerating design and optimization. This approach, combined with machine learning, offers a powerful pathway for rapid alloy development in additive manufacturing, particularly in resource-limited scenarios. The findings suggest this method can accelerate alloy design, enabling cost-effective, customized solutions for challenging environments while maintaining or enhancing material properties.

2:00 PM  
Generative AI for Inverse Design of Inconel 718: Jarvis Loh1; Nigel Neo2; Zhidong Leong1; Wen Jun Wee2; Yang Hao Lau1; Xinyu Yang1; Mark Jhon1; Rajeev Ahluwalia1; Robert Laskowski1; Wei-Lin Tan2; 1Institute of High Performance Computing (IHPC), Agency for Science, Technology and Research (A*STAR); 2DSO National Laboratories
    A challenge in employing machine learning for the inverse design of materials is explainability – many frameworks use property as input and processing parameters as output, resulting in a process-property ‘black box’ AI, bypassing microstructural information. In this work we attempt to plug the process-structure-property gap by using a Continuous Conditional Generative Adversarial Network (CcGAN) architecture to inversely generate microstructures corresponding to specified yield strengths. A convolutional neural network is then employed to predict the requisite processing parameters to attain these generated microstructures. To demonstrate this architecture, we generated an extensive training set of microstructures using phase field simulations to model the precipitation of γ’ and γ” in Inconel 718, and a crystal plasticity model to calculate the corresponding yield strengths. Our methodology resulted in qualitatively well-aligned CcGAN-generated microstructures compared with those derived from physics-based calculations. Furthermore, the predictions for processing parameters achieve a root mean squared error within 7%.

2:20 PM  
Inverse Alloy Design: An Alloy Composition Generation Framework With Flexibilities: Mohammad Abu-Mualla1; Ellis Crabtree2; Fredrick Michael2; Yayue Pan1; Jida Huang1; 1University of Illinois Chicago; 2NASA Marshall Space Flight Center
    Inverse design, generating material compositions with targeted properties, offers a promising approach to material discovery. Machine learning has emerged as a key enabler for generative design. However, existing models face challenges such as non-uniqueness (multiple compositions yield the same property), the complexity of high-dimensional compositional spaces, and the generation of physically non-achievable alloys. This work proposes a latent diffusion generative model for the inverse problem and a Variational Autoencoder (VAE) to obtain the latent representation; we constrained the latent space to ensure the generation of physically valid alloys. Our framework successfully mapped both the composition-property and property-composition correlations, using the VAE for the forward and the latent diffusion model for the inverse design. The model is trained on the Granta Alloys dataset. Experimental results reveal the proposed framework could accurately predict both pathways and generate multiple feasible designs, providing flexibility for selecting compositions that meet target properties.

2:40 PM  
Surrogate-Model-Assisted Multi-Objective Calibration of Crystal Plasticity Finite Element Method (CPFEM) Models: Janzen Choi1; Mark Messner2; Zhiyang Wang3; Tao Wei3; Tianchen Hu2; Jay Kruzic1; Ondrej Muransky3; 1University of New South Wales; 2Argonne National Laboratory; 3Australian Nuclear Science and Technology Organisation
    Crystal plasticity finite element method (CPFEM) models are powerful tools for simulating the deformation behaviour of polycrystalline materials, capturing the influence of a material’s microstructure on its macroscopic properties. However, identifying a unique set of crystal plasticity parameters presents significant challenges due to the vast parameter space and the high computational cost associated with microstructure-informed finite element simulations. To address these challenges, a surrogate model is developed to approximate the CPFEM response, which is then integrated with a multi-objective genetic algorithm (MOGA) to determine the crystal plasticity parameters. This approach optimises the material parameters using multiple objective functions, such as for the stress-strain curve and grain rotation measurements obtained via in-situ electron backscatter diffraction (EBSD) during tensile loading. The calibration workflow is demonstrated using Alloy 617 at room temperature, showing that the calibrated CPFEM model can accurately capture the stress-strain behaviour, overall texture evolution, and reorientation trajectories of individual grains.

3:00 PM Break

3:30 PM  
Integrated Microstructure and Mechanism-Guided Multimodal Machine Learning for Advanced Steel Design: Wei Xu1; 1Northeastern University
    Machine learning is revolutionizing the material design community, especially for high-performing steels. However, emerging methods such as automated laboratories suffer from poor interpretability and requirements for huge amount of data. The current study combines physical metallurgy knowledge and microstructural data to accurately predict properties of steels and improve steel design. Guided by thermodynamic data, deep learning models achieve precise predictions for frictional work and the martensite transformation start temperature. For more complex scenarios such creep and fatigue, transfer learning is employed where source models are established to understand the relationship amongst composition, processing, and mechanical properties. These learned mechanisms are then used to predict fatigue and creep properties. Recognizing that most problems in steel design rely heavily on microstructural information, rapid quantification methods for microstructural images are developed, and multimodal information pertaining to the images is extracted to develop site-specific AI strategies, finally achieving a comprehensive integration of prediction capabilities.

3:50 PM  
Linking Multiple Length Scales Using Material Data Driven Design (MAD3): David Montes De Oca Zapiain1; Hojun Lim1; 1Sandia National Laboratories
     Metal alloys used in stamping and forming processes exhibit polycrystalline structures at the lower length scale that cause the metal to display plastic anisotropy. Accurate predictions of the metal’s plastic anisotropy are crucial in manufacturing given the effect it has on the macro scale. Material Data Driven Design (MAD3) is an innovative software that leverages the power of machine learning to link the micro and macro scales and thus modernize the forming and stamping processes of sheet metals by predicting the parameters that characterize the load-dependent behavior of a metal alloy 1000 times faster than existing solutions. This software is conveniently packaged in a simple and easy-to-use graphical user interface that is deployed using cloud computing. In this talk, we present the structure and functionality of MAD3 and how this technology can be obtained by external users. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525. SAND2024-14838A

4:10 PM  
Micromechanics Surrogate Model for Fatigue Life Prediction of Composites: Brandon Hearley1; Steven Arnold1; 1NASA Glenn Research Center
    Fatigue modeling of composites, particularly using a multiscale approach, can be very costly, due the number of iterations that must occur when evaluating the damage state of each constituent within each ply, thus making it difficult for engineering in early design stages to evaluate a large number of potential candidate material configurations. In this work, a machine learning surrogate model for ply level micromechanics fatigue damage of composites is presented, enabling S-N curve prediction of laminates for any arbitrary number of plies. Training data is created using they physics-based Micromechanics Analysis Code with Generalized Method of Cells (MAC/GMC) tool for a single ply under multiaxial loading, predicting the damage increment and corresponding cycles to damage for each ply. The developed machine learning model will allow engineers to quickly get a reasonably accurate estimate of the fatigue life a composite with any arbitrary number of plies subject to any multiaxial load.

4:30 PM  
Parameter-Efficient Quantized Mixture-of-Experts Meets Vision-Language Instruction Tuning for Semiconductor Electron Micrograph Analysis: Sagar Srinivas Sakhinana1; Venkataramana Runkana1; 1Tata Research Development and Design Center
     Semiconductors, crucial to modern electronics, are generally under-researched in foundational models. It highlights the need for research to enhance the semiconductor device technology portfolio and aid in high-end device fabrication. In this paper, we introduce sLAVA, a small-scale vision-language assistant tailored for semiconductor manufacturing, with a focus on electron microscopy image analysis. It addresses challenges of data scarcity and acquiring high-quality, expert-annotated data. We employ a teacher-student paradigm, using a foundational vision-language models like OpenAI GPT-4o, Google Gemini as a teacher to create instruction-following multimodal data for customizing the student model, sLAVA, for electron microscopic image analysis tasks on consumer hardware with limited budgets. Our approach allows enterprises to further fine-tune the proposed framework with their proprietary data securely within their own infrastructure, protecting intellectual property. Rigorous experiments validate that our framework surpasses traditional methods, handles data shifts, and enables high-throughputscreening.