The Weather Research and Forecasting (WRF) model is a well-established model used broadly in atmospheric research for decades. At Uppsala University, WRF has traditionally run on CPU-based HPC systems such as Tetralith, but with the move to Arrhenius and the GPU acceleration it makes possible for weather models, we aim to optimize our setup for the new hardware. We have ported the physics schemes used in our configuration; Thompson microphysics, RRTMG long/shortwave radiation, the YSU boundary layer, the revised MM5 surface layer, and the Noah land surface model, to compile and run on GPUs using OpenACC. The port has been validated using a local NVIDIA GPU (limited to 6 GB VRAM), but performance gain was minimal due to hardware limitations. The Grace Hopper (GH200) chip with a larger memory and bandwidth should allow the same code to deliver a substantial performance improvement. Prior attempts to port WRF's physics to GPU have not been merged into the mainline WRF distribution, and known community efforts were abandoned some years ago. A proprietary GPU-accelerated version, AceCast, exists commercially but is not available to the scientific community. Our port is therefore novel within openly available WRF. We also run a coupled configuration of WRF with the WaveWatch III (WW3) wave model and the NEMO ocean model via the OASIS3-MCT coupler. We aim to test this setup on Arrhenius by running WRF on the GPU while NEMO and WW3 run concurrently on the chip's Arm CPU cores. Deutscher Wetterdienst (DWD) has already shown to run such a configuration efficiently on Grace Hopper hardware. This evaluation is carried out within the FESPAN project, which aims to improve modelling of electromagnetic (EM) signal propagation in marine environments such as the Baltic Sea.