Session Sheet - 3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): HPC4EI Program

3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025): HPC4EI Program
Program Organizers: Remi Dingreville, Sandia National Laboratories; Ali Riza Durmaz, Fraunhofer Institute Iwm

Wednesday 1:40 PM
June 18, 2025
Room: Platinum Ballroom 7&8
Location: Anaheim Marriott

Session Chair: Aaron Fisher, Lawrence Livermore National Laboratory

1:40 PM
The HPC4EI Program Connects Industry Partners With National Lab HPC and AI/ML Resources: Aaron Fisher¹; Nick Killingsworth¹; Victor Castillo¹; ¹Lawrence Livermore National Laboratory
    The High Performance for Energy Innovation (HPC4EI) program fosters public-private partnerships by awarding funding to collaborative R&D projects between national laboratories and industry partners. Focused on addressing critical manufacturing challenges, HPC4EI leverages the Department of Energy's (DOE) supercomputing resources and expertise. A key strategy employed in numerous HPC4EI projects involves high-performance computing (HPC) simulations to develop digital twins of manufacturing processes. These simulations generate massive datasets used to train advanced machine learning tools and reduced-order models capable of real-time predictions. This capability enables analysis and optimization of complex processes, with the potential to save hundreds of millions of dollars and millions of metric tons of CO2 emissions. This presentation will showcase successful examples of past HPC4EI manufacturing projects and outline how prospective industry partners can engage with the program.

2:00 PM
Faster Prediction With AI/ML for Manufacturing: Using Simulations and Production Data Develop Fast-Running Inference Models and Knowledge Tools for Manufacturing: Victor Castillo¹; Yeping Hu¹; Bo Lei¹; ¹Lawrence Livermore National Laboratory
    Manufacturing processes often involve numerous control parameters that influence final product quality. While computer simulations offer a significant advantage over physical experiments in navigating this high-dimensional control space, they can still be computationally expensive. This work leverages deep learning to develop fast-running surrogate models that accurately capture the dynamics of complex industrial processes. These surrogate models enable near-real-time predictions, facilitating efficient optimization of manufacturing parameters. Furthermore, a novel method is presented for integrating sparse manufacturing data with simulation outputs, enhancing the model's ability to predict actual production quality. The effectiveness of this approach is demonstrated through case studies involving both benchmark problems and real-world manufacturing systems.

2:20 PM
Development of a Massively Parallel Reduced-Order Model Based Design/Optimization Tool for Power Generation Using Natural Gas-H2 Blended Fuels: Shashikant Aithal¹; Nick Killingsworth²; Bob Schrecengost³; Aaron Fisher²; Victor Castillo²; ¹Argonne National Laboratory; ²Lawrence Livermore National Laboratory; ³Department of Energy
    Hydrogen use in power generating equipment such as gas turbines or internal combustion engines, traditionally fueled by natural gas, promises to reduce the generation of CO2. The fraction of hydrogen in the fuel mixture has a significant impact on the overall combustion characteristics and can pose unique operational challenges such as flashback in gas turbines and knocking in IC engines. Design and optimization of such power generation equipment fueled by NG-H2 blends present unique challenges on account of the large design space and conflicting constraints. We present the development of a massively parallel framework for generating data needed for AI-based models. Open-source code Cantera was used to generate over 16000 data points predicting ignition delay, flame-speed, burned gas temperatures and/or emissions over a range of conditions relevant to the operation of engines/gas turbines to generate a fast-running ROM. We will discuss current applications and possible future work.

2:40 PM
Physically Aligned Hierarchical Mesh-Based Network for Dynamic System Simulation: Yeping Hu¹; Bo Lei¹; Vic Castillo¹; ¹Lawrence Livermore National Laboratory
    Dynamic systems evolve through intricate interactions where local events impact global behavior, reflecting real-world interconnections. Modeling these systems requires capturing both local and long-range dynamics with accuracy and efficiency, a balance that current mesh-based Graph Neural Network (GNN) methods often struggle to achieve, especially with large datasets and complex meshes. Inspired by real-world dynamics, we introduce the Mesh-based Multi-Segment Graph Network (MMSGN), a framework designed to address these challenges through a hierarchical information exchange mechanism that aligns with physical properties. MMSGN integrates micro-level local interactions with macro-level global exchanges to accurately capture both local and global dynamics while remaining computationally efficient. Our model demonstrates superior accuracy, mesh quality, and scalability on multiple dynamic system datasets, outperforming several state-of-the-art methods and proving well-suited for large-scale, complex system simulations across diverse scenarios.

3:00 PM
Conditional Continuous Normalizing Flows for Damage Studies in Unidirectional Fiber-Reinforced Composites: Jihye Hur¹; Adam Generale²; Keith Ballard³; Vikas Varshney³; Craig Przybyla³; Surya Kalidindi¹; ¹Georgia Institute of Technology; ²Pratt & Whitney; ³Air Force Research Laboratory
    The microstructure-sensitive prediction of transverse failure strength in unidirectional polymer matrix composites (PMCs) remains elusive. This is mainly due to difficulties in quantifying the stochasticity in PMC microstructures, which propagates into variable damage property predictions. Traditionally, finite-element (FE) simulations have been used to extract microstructure-sensitive properties over representative volume elements of the material. However, the computational cost of running FE simulations can be incredibly high, especially when a large number of evaluations are needed to quantify stochasticity in the material response and characterize relevant properties. Statistical machine learning tools such as continuous normalizing flows are well suited for addressing this challenge, as they enable rapid microstructure to property predictions and perform efficient and exact inference of arbitrarily complex densities. This talk will discuss the advantages of using continuous normalizing flows in the context modeling PMC damage resilience, as well as demonstrate their efficacy in extracting microstructure-damage resilience property relationships.