Search

Mining key information from trajectory data can effectively help people in their life. In the case of hurricanes, trajectory prediction can avoid losses caused by disasters. Deep learning has widely used in many fields. However, trajectory is simply information such as longitude and latitude. It is difficult to extract trajectory features using deep learning. Therefore, a trajectory prediction model based on deep feature representation (DeepFR) is proposed by extracting various features. The model consists of trajectory data preprocessing layer, deep feature representation layer and trajectory prediction layer. The trajectory data preprocessing layer divides a trajectory into subtrajectories by using a sliding window. The deep feature extraction layer performs feature extraction for each subtrajectory. The spatiotemporal feature and 3D dynamic feature of each subtrajectory are extracted using bidirectional long short-term memory network (BiLSTM) and 3D convolutional neural network (3DCNN) respectively. The trajectory prediction layer first uses the self-attention to assign different weight to the feature of different subtrajectories. Then, latitude and longitude are predicted simultaneously through multitask learning. The real hurricane and typhoon data are used for simulation experiments. It is found that the DeepFR model has the best prediction effect and is more accurate and stable in long term prediction. [ABSTRACT FROM AUTHOR]

Multi-dimensional classification (MDC) assumes that each instance has multiple heterogeneous class spaces simultaneously, and each class variable describes the semantic information of instances from a specific dimension. Recent studies have proven that encoding heterogeneous class spaces into a special logical-label space and employing the label enhancement technique to learn latent real-number labels (i.e., label distributions) of instances is an effective strategy for MDC. However, the adopted label enhancement methods can result that data whose features are quite different to each other have similar label distributions. To tackle this problem, we propose a novel probability-based label enhancement approach for MDC. Specifically, manifold structures of the feature and label distribution spaces are transformed into two different probability distributions, and we expect them to be close. Subsequently, it makes label distributions of samples whose features have large differences be more differentiated. Moreover, the logical-label mapping and reconstruction terms are designed to preserve the intrinsic information from the logical-label space. Besides, an improved multi-output support vector regression is developed as the prediction model, where we introduce mean squared error to reduce the risk of model underfitting. Experimental results on ten benchmark datasets clearly validate the superiority of our method over state-of-the-art MDC baselines. [ABSTRACT FROM AUTHOR]

To improve the low-quality reconstruction of compressed ghost imaging (CGI) caused by non-negative measurement matrix, and avoid the requirements of large amount of training dataset and time-consuming training process, compressed ghost imaging based on deep image prior using singular value decomposition (SVDCGI-DIP) is proposed. The measurement matrix and corresponded measurements are first generated through numerical simulation and then simultaneously optimized with singular value decomposition without the change of optical setup of CGI. Secondly, deep image prior (DIP) method is applied to untrained generative network models to avoiding the long training process, and learned regularization is incorporated to further optimize network parameters to iteratively reconstruct the original image from only a few measurements of an image without training. Numerical experiments verify the superiority of our proposed SVDCGI-DIP method. [ABSTRACT FROM AUTHOR]

We have presented a multiscale computational scheme for the simulation of interactions between ultrafast intense laser pulses and condensed medium. This approach solves the coupled time-dependent Kohn-Sham (TDKS) and Maxwell equations in real time and provides a unified description of microscopic electron dynamics and macroscopic light propagation from first principles and applicable to wide ultrafast phenomena. Especially, as the TDKS equation was solved within the framework of all-electron full-potential and the augmented plane wave with local orbitals basis sets were adopted, this approach overcomes the drawbacks of real space grids and pseudopotential approximations and works well for some extreme conditions that the information near the nucleus is important. We demonstrated the approach by calculating the microscopic electron dynamics of crystalline silicon (Si) excited with an intense laser pulse and the macroscopic propagation of a laser pulse in a ∼3.5 μm free standing Si film with the feedback of microscopic electron dynamics coupled together. We also verified the availability of this approach under extreme condition with hundreds TPa (1 TPa = 1012 Pa) pressure in a metallic solid neon, and reasonable results were obtained. Our approach provides a new paradigm for multiscale simulations of ultrafast phenomena that induced by intense laser pulses in condensed medium. [ABSTRACT FROM AUTHOR]

• A two-dimensional weighted-orthogonal multiple-relaxation-time pseudopotential lattice Boltzmann model is developed. • A new geometrical formulation scheme based on the interaction force is proposed. • The present model can be applied to the simulations of the wetting phenomenon at the large density ratio of water-air system at room temperature with high numerical accuracy. • The present model exhibits excellent numerical stability in the simulations with low surface tensions. The wetting phenomenon is widespread in nature and industry, but simulating it with large density ratios remains challenging. In this study, a modified two-dimensional weighted-orthogonal multiple-relaxation-time pseudopotential lattice Boltzmann (WMRT-PLB) model is extended to simulate the wetting phenomenon at large density ratios. A simple formula describing the contact angle and interaction force is derived, and an alternative geometrical formulation scheme based on the interaction force is proposed to model the contact angle. First, theoretical analysis shows that the continuum form of the alternative geometrical formulation scheme is mathematically equivalent to that of the geometrical formulation scheme based on the pseudopotential. In contrast, the discrete forms of the two schemes are different in implementation. Then, four classical benchmark cases (static droplet on a wall, static droplet under gravity, droplet spreading on a wall and capillary intrusion) under a large density ratio (ρ l / ρ g = 857.7) are performed to verify and evaluate the performance of the WMRT-PLB model coupled with the alternative geometrical formulation scheme. The numerical results show that in the framework of the WMRT-PLB model, compared with the two conventional geometrical formulation schemes, the alternative geometrical formulation scheme is superior in numerical stability, especially under low surface tension, and achieves comparable numerical accuracy within a large range of surface tensions and contact angles. [ABSTRACT FROM AUTHOR]