Registered Data

[02514] Developing Performance Portable, Scalable and AI enabled Fusion Energy Physics Framework

  • Session Date & Time :
    • 02514 (1/2) : 5C (Aug.25, 13:20-15:00)
    • 02514 (2/2) : 5D (Aug.25, 15:30-17:10)
  • Type : Proposal of Minisymposium
  • Abstract : The plasma physics and tokomak engineering are widely recognized as a challenging "exascale" multi-scale multi-physics application. Recent advances in exascale computing, artificial intelligence (AI), and machine learning have paved the way for unprecedented opportunities in "in-silico" fusion reactor interpretation and design. Deep learning and the availability of powerful, easy-to-use HPC/ML toolboxes have played a significant role in achieving such breakthroughs. In this minisymposia, we are aiming to present recent advances in general numerical methods with adaptive meshes, AI methods, surrogate modelling, and performance-portable programming techniques for current and future computing architectures.
  • Organizer(s) : Xiaohu Guo, Jony Castagna, Vignesh Gopakumar, Stanislas Pamela
  • Classification : 65M25, 68T07, 35Q83, 65M50, 68Q85
  • Speakers Info :
    • Xiaohu Guo (UKRI STFC Hartree Centre)
    • Karthikeyan Chockalingam (UKRI STFC Hartree Centre)
    • Jony Castagna (UKRI STFC Hartree Centre)
    • Alex Grant (STFC Hartree Centre)
    • Lorenzo Zanisi (UK Atomic Energy Authority)
    • William Hornsby (UK Atomic Energy Authority)
  • Talks in Minisymposium :
    • [04465] Develop next generation CFD tools using AI libraries
      • Author(s) :
        • Xiaohu Guo (UKRI STFC Hartree Centre)
      • Abstract : Due to new hardware technologies, increases in computing power and developments in AI software, the benefits of combining AI techniques with traditional numerical methods for solving governing equations of dynamical systems are becoming apparent. This talk introduces a revolutionary approach to the discretization and solution of PDEs. Our approach implements CFD models using neural network with the aim of simplifying the software development and building on the very substantial developments already made in AI software.
    • [04466] Cheap training sets of gyrokinetic surrogate models with active learning
      • Author(s) :
        • Xiaohu Guo (UKRI STFC Hartree Centre)
        • Lorenzo Zanisi (UKAEA)
      • Abstract : Surrogate models of gyrokinetic turbulence play critical role in accelerating integrated models thus leading to faster post-discharge analysis, operations optimisation and flight simulator applications. Training sets for surrogates are obtained by brute-force approaches which may cause unnecessary, expensive oversampling, which limits the dimensionality of the input space and applications to high fidelity codes. We develop two-stage Active Learning pipeline to efficiently sample the input parameter space of gyrokinetic models, benchmark applications in multichannel integrated models.
    • [04627] Implementation of a finite element PDE solver in AMReX
      • Author(s) :
        • Karthikeyan Chockalingam (UKRI STFC Hartree Centre)
        • Alex Grant (UKRI STFC Hartree Centre)
        • Xiaohu Guo (UKRI STFC Hartree Centre)
      • Abstract : The scope of the current work is to embed a finite element implementation in AMReX. One of the advantages of using block-structured meshes is that each element does not require a unique isoparametric mapping. Initial testing of the implementation on the Poisson problem shows considerable performance gain compared to an unstructured finite element framework. The framework allows for the ease of implementing different types of higher-order elements. It has also been extended for solving Navier-Strokes.
    • [04652] AI deconvolution operator for plasma turbulent simulations on complex geometries
      • Author(s) :
        • Jony Castagna (UKRI-STFC Hartree Centre)
        • Francesca Schiavello (UKRI-STFC Hartree Centre)
      • Abstract : We use Generative Adversarial Networks (GANs) to model the nonlinear terms in partial differential equations on coarse structured grids. The idea is to reconstruct the high-resolution fields exploring the latent space of the GANs after being properly trained on curvilinear coordinates. The nonlinear terms are then found and mapped back to the coarse structured mesh where the filtered equations are solved. Results for Navier Stokes and Hasagawa-Wakatani plasma equations are presented.
    • [04741] Performance and scaling of amrPX: a multiphase CFD framework
      • Author(s) :
        • Alex Grant (UKRI STFC Hartree Centre)
        • Xiaohu Guo (UKRI STFC Hartree Centre)
        • Karthikeyan Chockalingam (UKRI STFC Hartree Centre)
      • Abstract : amrPX is a modularised, multi-model adaptive mesh refinement framework built using AMReX. Each model has a common structure with Data, Solver, Physics and Problem containers, with shared utility modules such as numerics or materials, to promote code re-use and ease of development. A compressible multiphase Five-Equation model has been implemented and benchmarked. Performance results and scalability across thousands of CPUs and GPUs is discussed.
    • [05229] A highly parallel simulation of patient-specific hepatic flows
      • Author(s) :
        • Zeng Lin (Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences)
      • Abstract : Computational hemodynamics is being developed as an alternative approach for assisting clinical diagnosis and treatment planning for liver diseases. The technology is non-invasive, but the computational time could be high when the full geometry of the blood vessels is taken into account. In this work, we study a highly parallel method for the transient incompressible Navier-Stokes equations for the simulation of the blood flows in the full three-dimensional patient-specific hepatic artery, portal vein and hepatic vein. As applications, we also simulate the flow in a patient with hepatectomy and calculate the portal pressure gradient. One of the advantages of simulating blood flows in all hepatic vessels is that it provides a direct estimate of the portal pressure gradient, which is a gold standard value to assess the portal hypertension. Moreover, the robustness and scalability of the algorithm are also investigated. A 83% parallel efficiency is achieved for solving a problem with 7 million elements on a supercomputer with more than 1000 processor cores.