Real-time semantic segmentation is a core perception capability for underwater robots and autonomous underwater vehicles (AUVs), yet it remains challenging because underwater imagery often exhibits low contrast, blurred boundaries, and strong appearance degradation under strict onboard computation budgets. This paper proposes MSNet, a multi-supervised two-pathway network that decouples feature learning into a semantic branch for context modeling and a detail branch for preserving high-resolution spatial information. MSNet introduces three complementary supervisory signals: (i) low-frequency semantic supervision derived from smoothed labels to encourage body semantics, (ii) high-frequency detail supervision derived from edge-enhanced labels to improve boundary localization, and (iii) category representation supervision implemented by a Category Representation Enhancement Module (CREM) to strengthen class discrimination at the deepest stage. To prevent auxiliary supervision from amplifying cross-resolution misalignment during fusion, we embed a Bilateral Flow-based Alignment Module (BFAM) into multi-stage feature fusion. Experiments on the SUIM benchmark show that MSNet achieves 79.83% mIoU and 86.57% F-score at 55 FPS with 6.2 M parameters on an RTX 3060 GPU, outperforming mainstream encoder–decoder and two-pathway algorithms. Compared with SFNet and BiSeNet V3, MSNet improves mIoU by 1.52% and 1.89%, and runs 9 FPS faster than SFNet. Ablation studies verify the effectiveness and complementarity of the proposed supervision and alignment strategies, indicating MSNet offers a practical accuracy–speed trade-off for marine engineering applications.

The analysis of cavitation pressure for high-speed underwater moving bodies is critical for hydrodynamic performance and stability. However, traditional methods relying on physical experiments or high-fidelity CFD simulations are computationally expensive and time-consuming, often resulting in small datasets that challenge data-driven approaches. This study aimed to evaluate efficient feature extraction techniques to overcome the limitations of small-sample scenarios. This paper systematically evaluated three feature extraction methods—Principal Component Analysis (PCA), Fast Independent Component Analysis (Fast ICA), and a customized 1-dimensional Convolutional Auto-encoder (Conv1D AE)—for processing cavitation pressure data. The evaluation was conducted through two experiments: the first assessed the ability of these methods to reconstruct the pressure evolution across the body surface in an unsupervised manner, while the second investigated their performance in predicting the peak pressure using a small set of labeled samples.The findings demonstrated a clear trade-off: Fast ICA exhibited the best performance in reconstructing the overall pressure evolution, followed closely by PCA, while the Conv1D AE showed limitations in capturing sharp pressure gradients. Conversely, for the critical task of peak pressure prediction from limited labeled data, the Conv1D AE model achieved a significant 10% increase in accuracy compared to the baseline model without feature extraction, with PCA providing a 3% improvement. Fast ICA, however, was less effective for this specific prediction task.These results underscore the effectiveness of tailored feature extraction in automating cavitation analysis. By reducing reliance on manual intervention and accelerating the extraction of key features like peak pressure, these methods offer a practical pathway to enhance the design and analysis cycle of underwater moving bodies. The study establishes a foundation for applying machine learning to small-sample fluid mechanics problems, with future work focused on optimizing network architectures for improved precision.

The application of Transformer-based large models has achieved significant success in recent years. However, the exponential growth in the parameters of large models introduces formidable memory challenges for edge deployment. Prior works to address this challenge mainly focus on optimizing the model structure and adopting memory swapping methods. However, the former reduces the inference accuracy, and the latter raises the inference latency. This paper ¹ introduces PipeLoad , a novel adaptive memory-efficient pipeline execution mechanism. It reduces memory usage by incorporating dynamic memory management and minimizes inference latency by employing parallel model loading. Based on the PipeLoad mechanism, we present Hermes \({}^{+}\) , a framework optimized for large model inference on edge devices. In Hermes \({}^{+}\) , we adopt a searching algorithm to balance inference latency and memory usage. We evaluate Hermes \({}^{+}\) with Transformer-based models of different sizes on a CPU server and edge devices. Our experiments illustrate that Hermes \({}^{+}\) achieves up to \(4.49\times\) speedup in latency and \(81.1\%\) lower memory footprint than the state-of-the-art pipeline mechanism for BERT and ViT models, \(24.3\times\) speedup in latency and \(67.2\%\) lower memory footprint for GPT-2 and GPT-J models with 128 tokens.

Rapid surrogate models are attractive for iterative computational fluid dynamics (CFD) design loops, though defining their operating envelope remains crucial. This study proposes RenaNet, a Reynolds-aware convolutional gated recurrent unit (ConvGRU) surrogate, for predicting two-dimensional laminar and transitional flows past cylinder and square obstacles. Using two initial flow snapshots and a Reynolds-number map, the model predicts spatiotemporal flow states up to 2000 time steps into the future, with Lattice Boltzmann Method (LBM) simulations serving as ground truth. Trained on Reynolds numbers of $1\le \mathrm{Re}\le 500$ (cylinder) and $1\le \mathrm{Re}\le 250$ (square), RenaNet achieves a minimum validation mean squared error (MSE) of $1.47\times {10}^{-5}$. A Reynolds-number ablation shows that removing the conditioning channel increases the validation MSE to $1.17\times {10}^{-3}$, while a ConvLSTM baseline gives $9.94\times {10}^{-4}$ with 24% more parameters. RenaNet also uses a direct long-horizon prediction interface for distant target frames. Auxiliary physics diagnostics confirm that predictions trained via MSE maintain acceptable continuity residuals across fitting, interpolation, and extrapolation cases. The average inference time for a 1000-step prediction horizon is approximately 1.25 s, delivering a 500-fold speedup over the reference LBM solver. Interpolation errors range from ${10}^{-4}$ to ${10}^{-2}$ depending on Reynolds number and geometry, while extrapolation beyond the training regime increases errors to the order of ${10}^{-2}$. These results establish RenaNet as a robust, parameter-efficient surrogate for laminar and transitional flows, with a clearly characterized operational boundary that informs future extensions into turbulent regimes.