Neko-TOP: Exascale Topology Optimization Framework for Transient Flow Problems
Presenter
DescriptionComputational design increasingly integrates machine-driven optimization with human engineering expertise. Ensuring trust in automated design recommendations is critical, particularly for complex systems like transient flow. Topology Optimization offers a mathematical framework for optimal design, but transient flow problems demand vast computational resources: full flow simulations to evaluate objectives and adjoint simulations for sensitivities which require storing entire simulation histories.
We present Neko-TOP, a GPU-accelerated Topology Optimization framework targeting human-machine co-design at exascale. Built upon the validated Neko spectral element CFD library, our framework provides engineers with trustworthy computational guidance through reproducible, high-fidelity simulations. Our implementation demonstrates near-ideal scaling up to 256 GPUs and has been validated on LUMI, verifying result consistency across systems. By grounding computational optimization in validated physics, we enable confident human-machine collaboration on complex fluid design problems previously infeasible on conventional systems.
In this presentation, we discuss scalability achievements, key challenges in adjoint storage and reproducibility, and demonstrate how close integration of CFD with Topology Optimization fosters trust in computational design recommendations for transient flow problems.
We present Neko-TOP, a GPU-accelerated Topology Optimization framework targeting human-machine co-design at exascale. Built upon the validated Neko spectral element CFD library, our framework provides engineers with trustworthy computational guidance through reproducible, high-fidelity simulations. Our implementation demonstrates near-ideal scaling up to 256 GPUs and has been validated on LUMI, verifying result consistency across systems. By grounding computational optimization in validated physics, we enable confident human-machine collaboration on complex fluid design problems previously infeasible on conventional systems.
In this presentation, we discuss scalability achievements, key challenges in adjoint storage and reproducibility, and demonstrate how close integration of CFD with Topology Optimization fosters trust in computational design recommendations for transient flow problems.
TimeWednesday, July 19:30 – 10:00 CEST
LocationBldg. 6 – Room 002
SessionMS4A – Performance-Portable High-Order CFD for Digital Twins on Heterogeneous Exascale Architectures
Session Chair
Event Type
Minisymposium
