This research provides a comprehensive examination of hybrid modeling approaches that integrate physics-based and data-driven methods across three transportation sectors: automotive, aviation, and maritime. Motivated by the need for improved generalization, interpretability, and robustness, hybrid models combine the expressiveness of machine learning with the reliability of physical laws. Existing hybridization techniques are categorized according to the fusion point, including input features, loss functions, model architecture, and output correction, and their respective advantages and challenges are assessed. Domain-specific trends are identified: automotive and aviation benefit from high-fidelity physical models and abundant labeled datasets, while maritime applications often contend with sparse, noisy data and low-fidelity models, thus placing greater emphasis on hybrid methods to ensure safety and reliability. We further analyze strategies such as physics-informed neural networks, residual learning, transfer learning with synthetic data, and multimodel architectures, and evaluate their suitability under different data availability and physical modeling constraints. This review highlights the critical role of domain knowledge, the impact of physical model fidelity, and the importance of adaptive integration strategies in achieving robust and trustworthy dynamic models in real-world transport systems.

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries plays a crucial role in the sustainable development and efficient operation of fields such as new energy vehicles, numerous electronic products, and energy storage power stations. This study presents a dual-exponential model integrated with an Extended Kalman Filter (EKF) for RUL prediction, enhanced by Bayesian optimization to automatically adjust the model parameters for improved accuracy. A dynamic weight loss function is employed to adapt to battery characteristics at various degradation stages, enabling optimal model parameterization. The model is evaluated using four datasets under different starting prediction cycles (300, 400, and 500 cycles). Experimental results demonstrate that the model effectively tracks capacity degradation and predicts RUL with high accuracy, fitting well with actual degradation curves and showing strong generalization ability across various datasets. Based on the conducted experiment, the proposed model demonstrates improved accuracy in tracking capacity degradation and predicting RUL.

For decades, supercritical flame simulations incorporating detailed chemistry and real-fluid transport have been limited to millions of cells, constraining the resolved spatial and temporal scales of the physical system. We optimize the supercritical flame simulation software DeepFlame—which incorporates deep neural networks while retaining the real-fluid mechanical and chemical accuracy—from three perspectives: parallel computing, computational efficiency, and I/O performance. Our highly optimized DeepFlame achieves supercritical liquid oxygen/methane (LOX/\(\ce {CH4}\)) turbulent combustion simulation of up to 618 and 154 billion cells with unprecedented time-to-solution, attaining 439/1186 and 187/316 PFlop/s (32.3%/21.8% and 37.4%/31.8% of the peak) in FP32/mixed-FP16 precision on Sunway (98,304 nodes) and Fugaku (73,728 nodes) supercomputers, respectively. This computational capability surpasses existing capacities by three orders of magnitude, enabling the first practical simulation of rocket engine combustion with >100 LOX/\(\ce {CH4}\) injectors. This breakthrough establishes high-fidelity supercritical flame modeling as a critical design tool for next-generation rocket propulsion and ultra-high energy density systems.

Digital twin (DT) technology, which creates virtual representations of physical systems to optimize their life-cycle, has drawn significant attention across various industries. The automotive and aviation industries have been pioneers in adopting DTs for enhanced efficiency, predictive maintenance, and real-time decision-making. However, the maritime industry, crucial to global trade and logistics, has lagged in DT implementation. This paper aims to bridge this gap by systematically surveying DT applications in the automotive and aviation industries and exploring how this knowledge can be transferred to the maritime industry. By analyzing existing literature, identifying key trends, and summarizing best practices, a comprehensive roadmap is provided for maritime industry adoption of DT technology. The surveyed papers are selected systematically following the PRISMA statement and categorized based on characteristics such as single vs. multiple systems, modeling methods (model-driven, data-driven, and hybrid), and life-cycle phases. We introduce DT models using a five-dimensional framework and analyze their characteristics in terms of research object, subsystem application, and modeling method. Additionally, DT applications from a product life-cycle perspective, covering design, manufacturing, operation, and maintenance phases are examined. Knowledge transfer from the automotive and aviation industries to the maritime industry is summarized. In the automotive industry, DTs enhance vehicle efficiency and safety, particularly for autonomous and electric vehicles. Aviation DT research focuses on predictive maintenance, pilot training, and real-time monitoring to improve operational efficiency and safety. The maritime industry faces data challenges and operational complexity but has significant potential for DTs to enhance ship performance, safety, and predictive maintenance.

Natural language processing (NLP) has recently shown significant progress in rich-resource scenarios. However, it is much less effective for low-resource scenarios due to the model easily overfitting to limited training data and generalizing poorly on testing data. In recent years, consistency training has been widely adopted and shown great promise in deep learning, but still remains unexplored in low-resource settings. In this work, we propose DM-CT, a framework that incorporates both data-level and model-level consistency training as well as advanced data augmentation techniques for low-resource scenarios. Concretely, the input data is first augmented, and the output distributions of different sub-models generated by model variance are forced to be consistent (model-level consistency). Meanwhile, the predictions of the original input and the augmented one are also constrained to be consistent (data-level consistency). Experiments on different low-resource NLP tasks, including neural machine translation (4 IWSLT14 translation tasks, multilingual translation task, and WMT16 Romanian$\to$English translation), natural language understanding tasks (GLUE benchmark), and named entity recognition (Conll2003 and WikiGold), well demonstrate the superiority of DM-CT by obtaining significant and consistent performance improvements.

In this work, a survey of DT-related research in the maritime industry is presented. A five-dimensional DT model for the maritime industry is presented and explained. Moreover, research object and characteristics of DT in maritime industry are categorized and discussed. The research objects of DT in the maritime industry are classified into ships, marine structures, underwater vehicles and marine engines. The characteristics of the maritime industry DT models are discussed in terms of target system, research contribution and simulation method. In addition, DT in the context of product life-cycle perspective in the maritime industry is analyzed and discussed in terms of design, manufacturing, operation, and maintenance phases. Based on our analysis and discussion of the research, we found that the current DT research in the maritime sector is mainly focused on relatively small modules, or small systems, or even only on individual components. In addition, only a small fraction of the reviewed DT-related papers focuses on the whole life-cycle in the maritime industry. The reason may come from that the models and sub-models are not yet flexible and adaptive enough at different life-cycle stages. Therefore, we believe that the development of DT technology is still in the developmental stage in the maritime industry.

In the last decade researchers have increasingly considered eye tracking of the operators of cars and airplanes as a means to address human error and evaluate operational effectiveness. This article presents a systematic survey of recently published papers about this approach in service to the question as to whether eye tracking can be used to address operational safety in marine operations. The surveyed papers are selected systematically and were categorized according to several defined characteristics. Eye tracking depends on defining operators' areas of interest (AOIs) and measuring operators focus on them over time. We identified the method of defining AOIs as a key distinction between studies; the papers fell into four categories, depending on whether researchers relied on an expert, based it on the stimulus itself, or used an attention map or a clustering algorithm to define the AOIs they used. The article also summarizes and analyzes the design and procedure of the eye-tracking experiments in the papers. Based on the features of marine operation, instruction on AOI definition in different scenarios is extracted; guidelines on experimental design and procedure selection are provided. In the article's conclusion we apply the results to a case study of a heavy-lifting operation to demonstrate the effectiveness of eye-tracking in marine operations.

Improving the training programs for maritime operations is beneficial to enhance the maritime safety in practice. In this article, we propose a novel approach to the assessment of visual attention in a maritime operation so as to support an expert-in-the-loop training program. Experts' knowledge of maritime operation and experiences in the simulator are incorporated into the training program in three ways. First, through a questionnaire, information about task division, identification of critical operation, and definitions of areas of interest (AOIs) are incorporated as prior knowledge for modeling visual attention. Second, a weight scale factor that emphasizes the high importance of visual focus in critical operations is utilized to generate an operations-dependent attention map. Third, based on an expert's attention map and visual switch between AOIs, a similarity metric is designed as a comprehensive evaluation between saliency and visual transition. A case study of heavy lifting operation is performed by two groups of trainees who have received different briefings about “critical operation.” The assessment result shows that the group with more detailed briefing obtains a 6% similarity score higher than the other group, which is consistent with the debriefing result of a superior performance of that group. The proposed approach is thus verified effective in assessing visual attention for the expert-in-the-loop training program.