In 3D geo-electromagnetic modeling, an adequate discretisation of the modeling domain is crucial to obtain accurate forward responses and reliable inversion results while reducing the computational cost. This paper investigates the mesh design for subsurface models, including steel-cased wells, which is relevant for many exploration settings but still remains a numerically challenging task. Applying a goal-oriented mesh refinement technique and subsequent calculations with the high-order edge finite element method, simulations of 3D controlled-source electromagnetic models in the presence of metallic infrastructure are performed. Two test models are considered, each needing a distinct version of approximation methods to incorporate the conductive steel casings of the included wells. The influence of mesh quality, goal-oriented meshing, and high-order approximations on problem sizes, computational cost, and accuracy of electromagnetic responses is investigated. The main insights of our work are: (a) the applied numerical schemes can mitigate the computational burden of geo-electromagnetic modeling in the presence of steel artifacts; (b) investigating the processes driving the meshing of models with embedded metallic infrastructures can lead to adequate strategies to deal with the inversion of such electromagnetic data sets. Based on the modeling results and analyses conducted, general recommendations for modeling strategies are proposed when performing simulations for challenging steel infrastructure scenarios.

Many current embedded systems comprise heterogeneous computing components including quite powerful GPUs, which enables their application across diverse sectors. This study demonstrates the efficient execution of a medium-sized self-supervised audio spectrogram transformer (SSAST) model on a low-power system-on-chip (SoC). Through comprehensive evaluation, including real time inference scenarios, we show that GPUs outperform multi-core CPUs in inference processes. Optimization techniques such as adjusting batch size, model compilation with TensorRT, and reducing data precision significantly enhance inference time, energy consumption, and memory usage. In particular, negligible accuracy degradation is observed, with post-training quantization to 8-bit integers showing less than 1% loss. This research underscores the feasibility of deploying transformer neural networks on low-power embedded devices, ensuring efficiency in time, energy, and memory, while maintaining the accuracy of the results.