Hybrid Inference Optimization for AI-Enhanced Turbulent Boundary Layer Simulation on Heterogeneous Systems - Supplemental Material

This is supplemental material for the paper "Hybrid Inference Optimization for AI-Enhanced Turbulent Boundary Layer Simulation on Heterogeneous Systems" submitted to the MMCP 2026 workshop.

Repository Structure

• CPP-ML-Interface: Source code of the ML-Module partially ported to C++ from Fortran. It also includes the AIxeleratorService library for hybrid inference as a submodule.
• maia: Source code of the finite volume solver from the multi-physics simulation framework m-AIA coupled with the TBL-Transformer model.
• input: contains the required input files for the solver (e.g. grid, restart file, Transformer model)

Installation

Software Environment

The code was tested on the CLAIX-2023 cluster at RWTH Aachen University. On CLAIX-2023 you can setup the correct environment via

source setup_env_claix23.sh

which will load some modules corresponding to the following toolchain:

• GNU compilers 13.2.0
• FFTW 3.3.10 (MPI parallel version)
• PnetCDF 1.12.3
• HDF5 1.14.3
• Python 3.11.5
• CMake 3.29.3
• Intel oneAPI Math Kernel Library 2024.2.0
• CUDA 12.4.0
• cuDNN 8.9.7.29
• Score-P 8.4

On other systems you may need to create a similar software environment yourself. Please report any issue on other systems!

Dependencies

The m-AIA solver code depends on the CPP-ML-Interface library, so this should be built first by using the provided install script.

cd CPP-ML-Interface
./install-scorep.sh

This script should work on any Linux-based system and will automatically install the following dependencies into CPP-ML-Interface/extern/:

• a Python virtual environment
• Physics Deep Learning coupLer (PhyDLL)
• prebuilt libtorch 2.6.0
• AIxeleratorService
• HighFive

and finally the compiled CPP-ML-Interface library will be found in CPP-ML-Interface/BUILD-SCOREP. Further information can be found in the README.

m-AIA Solver

To build the coupled m-AIA solver use the provided install script.

./install-MAIA.sh

Further information regarding m-AIA's build process can be found in the README.

Reproducing Results

To reproduce the results shown in the paper, we provide an automated benchmark suite. Ensure that you have successfully built the solver (see Installation) before proceeding.

1. Setup Benchmark Suite

First, run the setup script with your SLURM account. This script performs the following actions:

• Downloads the required large input files (grid, restart files) into the input/ directory.
• Sets up a Python virtual environment (.venv) with all necessary plotting dependencies.
• Generates the complete directory structure for all experiments in data/.

# Replace <ACCOUNT> with your actual Slurm account (e.g., rwth1234)
./create_benchmark_suite.sh <ACCOUNT>

2. Run Experiments

Once the suite is set up, navigate to the data directory and execute the experiment runner. This will submit the jobs to the cluster.

cd data
./run_all_experiments.sh

3. Generate Plots

After the experiments have finished, you can generate the plots used in the paper by running:

./run_all_plotting.sh