Uploaded November 2026 | Updated November 2026, 3 hours ago
Asynchronous Field-Particle Coupling for Multiphase Cloud Simulation using Heterogeneous HPC by Henrik Rusche
PreTalx: pretalx.com/juliacon-2026/talk/VFZ9X7
Schematic execution timeline for one GPU driven by a host CPU and a slave CPU. Asynchronous overlap of Eulerian and Lagrangian calculations is clearly visible.: pretalx.com/media/juliacon-2026/submissions/VFZ9X7/resources/Figure_2-1-5_ZO8BpL0.png
Snapshot of droplets in the 81m³ cloud chamber. Two side walls and the top are removed for illustration purposes and only a selection of droplets is shown. The size of the droplets is scaled up proportional to their diameter, but with an additional scaling factor to make them visible.: pretalx.com/media/juliacon-2026/submissions/VFZ9X7/resources/Figure_4_FjM0hrN.png
flyer-VFZ9X7: pretalx.com/media/juliacon-2026/submissions/VFZ9X7/resources/VFZ9X7_fB2FUEp.png
public GIT repository: github.com/Wikki-GmbH/SCALE-TRACK
Efficient simulation of multiphase flows remains a major challenge, particularly for cloud microphysical processes in which interactions between turbulent airflow and suspended droplets must be resolved in detail. We present a novel asynchronous two-way coupled Euler–Lagrange simulation framework that exploits heterogeneous computing architectures to achieve unprecedented scalability.
The proposed method executes Eulerian field calculations on CPUs using the OpenFOAM software package, coupled asynchronously to Lagrangian particle tracking on GPUs implemented in Julia, minimizing computational idling times and synchronization barriers. Data transfers are initiated immediately upon data availability, with Eulerian source terms predicted from previous time steps and subsequently corrected to ensure conservation of mass and momentum. Particles are organized into cache-friendly chunks with maintained bounding boxes, enabling dynamic load balancing across GPUs and optimized CPU-GPU data transfers.
Comprehensive testing on a local workstation and on a EuroHPC JU supercomputer revealed dramatic improvements: the algorithm achieves excellent scalability up to 256 billion droplets. The overall time-to-solution improved by a factor of 4.5, while energy efficiency improved 3.4 times compared to established methods. Weak and strong scaling tests demonstrated very good efficiency and speedup using up to 2500 cores paired with 256 GPUs.
The software is available in a public GIT repository at github.com/Wikki-GmbH/SCALE-TRACK
Asynchronous Field-Particle Coupling for Multiphase Cloud Simulation using Heterogeneous HPC by Henrik Rusche
PreTalx: pretalx.com/juliacon-2026/talk/VFZ9X7
Schematic execution timeline for one GPU driven by a host CPU and a slave CPU. Asynchronous overlap of Eulerian and Lagrangian calculations is clearly visible.: pretalx.com/media/juliacon-2026/submissions/VFZ9X7/resources/Figure_2-1-5_ZO8BpL0.png
Snapshot of droplets in the 81m³ cloud chamber. Two side walls and the top are removed for illustration purposes and only a selection of droplets is shown. The size of the droplets is scaled up proportional to their diameter, but with an additional scaling factor to make them visible.: pretalx.com/media/juliacon-2026/submissions/VFZ9X7/resources/Figure_4_FjM0hrN.png
flyer-VFZ9X7: pretalx.com/media/juliacon-2026/submissions/VFZ9X7/resources/VFZ9X7_fB2FUEp.png
public GIT repository: github.com/Wikki-GmbH/SCALE-TRACK
Efficient simulation of multiphase flows remains a major challenge, particularly for cloud microphysical processes in which interactions between turbulent airflow and suspended droplets must be resolved in detail. We present a novel asynchronous two-way coupled Euler–Lagrange simulation framework that exploits heterogeneous computing architectures to achieve unprecedented scalability.
The proposed method executes Eulerian field calculations on CPUs using the OpenFOAM software package, coupled asynchronously to Lagrangian particle tracking on GPUs implemented in Julia, minimizing computational idling times and synchronization barriers. Data transfers are initiated immediately upon data availability, with Eulerian source terms predicted from previous time steps and subsequently corrected to ensure conservation of mass and momentum. Particles are organized into cache-friendly chunks with maintained bounding boxes, enabling dynamic load balancing across GPUs and optimized CPU-GPU data transfers.
Comprehensive testing on a local workstation and on a EuroHPC JU supercomputer revealed dramatic improvements: the algorithm achieves excellent scalability up to 256 billion droplets. The overall time-to-solution improved by a factor of 4.5, while energy efficiency improved 3.4 times compared to established methods. Weak and strong scaling tests demonstrated very good efficiency and speedup using up to 2500 cores paired with 256 GPUs.
The software is available in a public GIT repository at github.com/Wikki-GmbH/SCALE-TRACK











- Provides a growing collection of benchmark combinatorial decision problems.
- Includes all necessary components to build and train a DFL policy for each problem.
- Offers a general interface for defining custom problems.
- Facilitates reproducibility and experimentation by allowing the same algorithm to be applied across multiple benchmarks with minimal changes.
- [`DecisionFocusedLearningAlgorithms.jl`](https://github.com/JuliaDecisionFocusedLearning/DecisionFocusedLearningAlgorithms.jl)
- Implements generic versions of state-of-the-art DFL training algorithms.
- Provides compatibility with problems defined through the benchmarks interface.
- While some classical DFL approaches can be implemented directly using the lower-level tools provided by InferOpt.jl, recent state-of-the-art methods have become increasingly complex. This package introduces higher-level implementations to make these advanced algorithms accessible without requiring users to master the technical details.
The goal of the talk is to demonstrate how users can leverage these packages to define a problem, select a training strategy, and train a policy for their own problem in a few lines of code, without necessarily needing deep expertise in DFL. JuliaDecisionFocusedLearning: A Practical Introduction to Decision-Focused Learning in Julia | Baty](https://i.ytimg.com/vi/g2XFyvBmsz0/mqdefault.jpg)