This paper presents a hybrid neural self-exciting point-process model tailored to capture the dynamics of extreme financial shocks. The research background stems from the limitations of classical parametric models in handling nonstationary baselines and the lack of interpretability in purely neural approaches. The primary research objective is to develop a model that couples a learned, time-varying baseline intensity produced by a recurrent encoder with a parsimonious, parameterized self-excitation kernel. The study utilized daily data from three distinct market classes—developed equity (S&P 500), emerging markets (Hang Seng Index), and cryptocurrency (BTC-USD)—over the period 2015–2023. Methodologically, a recurrent neural network captures exogenous drivers, while a parametric kernel models endogenous clustering. Results indicate that this hybrid architecture achieves superior early-warning performance compared to classical Hawkes and GARCH , baselines, particularly in markets with frequent regime shifts. Furthermore, the model provides stable, interpretable memory estimates, such as half-life summaries, which are critical for risk monitoring. The conclusion suggests that hybrid models offer a pragmatic balance between flexibility and interpretability for financial risk analysis.

The stock price sequence has a low signal-to-noise ratio, strong nonlinearity and statistical characteristics drift over time. Once it is necessary to continuously predict multiple moments in the future, the error will continue to accumulate along the step length, making the difficulty of prediction significantly increased. In response to this problem, this paper embeds the attention mechanism into the long-term and short-term memory network to construct an LSTM-Attention hybrid model for multi-step prediction: first, LSTM encodes the timing dependence of historical sequences, and then assigns learningable weights to the hidden state of each time step through the attention layer, so as to highlight the key points that are more distinguishable for future trends. Finally, the full connection layer gives the price of several future moments at once. The experiment selects the daily frequency data of the Shanghai-Shenzhen 300 Index from 2015 to 2024 and compares it horizontally with models such as ARIMA, LSTM, GRU, CNN-LSTM and Transformer. The results show that the proposed model leads the overall accuracy of RMSE, MAE, MAPE and direction, and the root mean square error decreases by about 20% compared with the single LSTM; and with the predicted step length increases, the attention mechanism can significantly inhibit the accumulation of error. The average absolute percentage error of 10-step prediction is about 31% lower than that of LSTM, indicating that the hybrid structure has a good effect in multi-step prediction tasks.

Full-waveform inversion (FWI) adopts the dynamic and kinematic information contained in pre-stack seismic wavefields and fits actual observed seismic data using the wave equation. It can realize high-resolution quantitative imaging of physical property parameters of underground media, with a theoretical imaging accuracy reaching half of the seismic wavelength. Since Tarantola and Lailly established the core theoretical framework in the 1980s, FWI has gradually evolved from two-dimensional acoustic approximate algorithms to a three-dimensional elastic and viscoelastic multi-parameter inversion system, and its application scope has expanded from simple geological structures to complex stratigraphic regions. In recent years, the iteration and upgrading of high-performance computing hardware, the application of deep learning algorithms in geosciences, and the popularization of multi-component and wide-azimuth seismic acquisition equipment have jointly driven FWI into a brand-new development stage. This paper systematically sorts out the theoretical system and classical research methods of FWI, and summarizes the development context of core technologies including multi-scale inversion, objective function construction, initial model establishment and computational efficiency optimization. Meanwhile, it concludes the latest achievements of research hotspots such as elastic and viscoelastic multi-parameter inversion, deep learning integrated FWI, uncertainty quantification, 4D time-lapse monitoring and cross-scale imaging. Finally, combined with the industrial development trend, this paper prospects future research directions including deep integration of physics and artificial intelligence, large-scale high-efficiency inversion, multi-physical field coupled imaging, and standardization of result interpretability. It aims to provide a systematic literature review and research reference for researchers engaged in seismic inversion and oil and gas exploration.

The Epoch of Reionization was a major change in cosmic history because it altered the ionization state of the gas between galaxies. Before this transition, much of the universe contained neutral hydrogen. That hydrogen did not block every light frequency equally, but for radiation connected to hydrogen transitions, it was a significant obstacle. As the first stars and galaxies formed, their ultraviolet radiation gradually ionized surrounding hydrogen. Over time, the universe moved from a mostly neutral state toward a more ionized and transparent one. The broad sequence is known, but the detailed history is still uncertain: when reionization began, how quickly it advanced, how uneven it was, and which galaxies supplied most of the ionizing radiation remained unanswered. This paper examines Lyman alpha emitters as one way to study that history. Lyman alpha emission is useful because neutral hydrogen in the intergalactic medium affects whether Lyman alpha photons reach observers; changes in Lyman alpha visibility, rest-frame equivalent width, escape fraction, velocity offset, and emitter fraction can help infer how the neutral hydrogen fraction changed over time. Recent James Webb Space Telescope (JWST) Near-Infrared Spectrograph observations from the JWST Advanced Deep Extragalactic Survey (JADES) provide a stronger observational base, while simulations such as THESAN help interpret the final signal after photons pass through galaxies, circumgalactic gas, intergalactic gas, dust, scattering, and cosmic expansion. Lyman alpha emitters are powerful probes of reionization, but not direct measuring sticks. They are most useful when spectroscopy, completeness corrections, model comparisons, and simulations are used together.

Accurate trajectory tracking serves as a core technology for quadrotor unmanned aerial vehicles to perform inspection, mapping, monitoring and low-altitude autonomous flight missions. Affected by underactuated, nonlinear and strongly coupled dynamics, the horizontal motion of quadrotors is realized by adjusting attitudes to change the direction of thrust, while attitude responses are jointly restricted by rotational coupling, inertia parameters and external disturbances. This study adopts three typical trajectory tracking methods, namely proportional-integral-derivative control, sliding mode control and integral backstepping sliding mode control, and constructs a performance evaluation framework from the perspectives of trajectory response, position error, attitude variation and comprehensive error indicators. The results indicate that the three-dimensional root mean square error and maximum three-dimensional error of PID control under periodic trajectories are 0.5836 m and 0.8560 m respectively, which proves that the basic feedback structure presents noticeable deviations in curve tracking. Sliding mode control reduces the maximum three-dimensional error to 0.0509 m, showing excellent capability in limiting error bounds. For integral backstepping sliding mode control, its three-dimensional root mean square error and terminal three-dimensional error are 0.0419 m and 0.0050 m separately. It demonstrates that integral compensation and backstepping sliding mode structure help enhance overall precision and terminal convergence performance. The findings clarify the differences among the three control structures in error suppression, attitude regulation and terminal accuracy, and can provide references for the selection of trajectory tracking controllers for quadrotor unmanned aerial vehicles.

This paper applies three algorithms in deep reinforcement learning (DQN, Double DQN, Dueling Double DQN) to stock prediction and trading decisions. Using the historical trading data of Kweichow Moutai (600519.SS) from 2018 to 2023 as the experimental sample, a systematic stock trading reinforcement learning environment is constructed. Subsequently, the prediction performance and trading returns of the three algorithms are compared. Ultimately, the cumulative returns of the first two algorithms are 15.2% and 23.5% respectively, while that of Dueling Double DQN is 32%. In terms of maximum drawdown, Dueling Double DQN is only 8%, significantly lower than 18% and 12% of DQN and Double DQN respectively. Regarding the win rate, Dueling Double DQN reaches 65%, higher than 52% and 58% of the other two algorithms. In terms of average total reward, Dueling Double DQN reaches 1645.72, far exceeding 848.56 and 1196.89 of DQN and Double DQN respectively. The results of the experiment show that due to the adoption of the value-advantage separation network structure, Dueling Double DQN outperforms Double DQN and basic DQN in terms of cumulative return, maximum drawdown, and win rate. Therefore, the accuracy of stock prediction and the stability of trading decisions have been greatly improved.

In recent years, deep reinforcement learning has seen notable advances in areas such as game playing and autonomous driving. How the underlying reasoning and decision-making processes actually develop and improve remains an open question that deserves closer attention. This paper starts from the most basic decision-making dilemma and traces the optimization path of deep reinforcement learning in reasoning and decision-making along the logical thread of sequential choice under complexity. Through a comparative look at classic strategies in the multi-armed bandit problem, the theoretical structure of Markov decision processes, and major deep reinforcement learning algorithms from the past five years, the paper shows that the heart of deep reinforcement learning optimization lies in a shift from model-based deduction toward interaction-based experience building. The results suggest that current algorithms continue to grapple with a trade-off between sample efficiency and generalization, while structural causal reasoning is drawing growing interest as a new research direction. The aim of this work is to offer a reference point for thinking about the nature of decision-making and for designing more effective learning algorithms.

In recent years, with advances in astrophysical observation technology, direct imaging has become a powerful method for studying exoplanets, especially in the near-infrared band (11002526 nm). However, under the interference of strong stellar background light and instrumental residuals, planetary signals are extremely faint, typically requiring a contrast level of 10−5–10−8. Traditional image-processing methods such as Principal Component Analysis combined with Angular Differential Imaging (PCA-ADI) often fail to robustly de- tect planetary signals under low signal-to-noise ratio conditions. To address this, the PCA method is combined with three machine learning methods, including k-nearest neighbors (KNN), convolutional neural networks (CNN), and a discriminator based on generative ad- versarial networks (GAN). Using real observational data and precisely injected simulated-planet datasets from the VLT/SPHERE and Gemini/GPI exoplanet imaging challenge, a high-accuracy recognition model is constructed. Through an improved data-balancing strat- egy and enhanced preprocessing, remarkable performance is achieved in detection tasks: KNN reaches 97.83% accuracy and 98.81% F1 score, GAN achieves 100% precision and 96.43% recall, and CNN achieves 88.04% accuracy and 91.01% precision. This study successfully identifies exoplanet signals in eight datasets, verifies the effectiveness of machine-learning methods in high-contrast imaging tasks, and provides a new data-processing framework for future exoplanet detection.

Silicon photonic device has compatibility property with CMOS technique; however, it is restricted by the non-direct bandgap of silicon, which brings about bad light emission efficiency and a disability to detect light with wavelengths that surpass 1.1 μm. Germanium makes the detection scope become wider toward the short-wave infrared area. Even so, its indirect bandgap still brings a barrier to the development of high-efficiency luminescence devices. Si-Ge-Sn alloys have become the foreground as a hopeful solution because the adding of tin can bring out a direct bandgap changing. The band gap of these alloy materials can be regulated in the interval from 0 to 0.8 eV, which covers the short-wave to mid-wave infrared spectral range. This review makes a summary of current research progresses about Si-Ge-Sn alloys which are used for infrared optoelectronic devices. It concentrates tightly on four key domains: theoretical energy band construction, material epitaxial growth, device utilizations, and important difficulties. Worthy achievements that attract attention include direct-bandgap Ge-Sn light-emitting lasers, middle-infrared light detecting devices, and the first-time on-chip integration work. The still existing questions, for example tin separation, heat stability, ohmic contact points, and restrictions of the CMOS heat budget, are discussed deeply, together with the future research directions. Defeating these difficulties will let out the whole capability of Si-Ge-Sn compound materials for low-cost, high-effectiveness infrared optoelectronic devices which are totally integrated on the silicon base platform.

Power semiconductor devices are essential to modern power and electronic systems as they govern the switching, conversion, and delivery of electric power. Silicon has long been the dominant material in this field. However, the rapid growth of electric vehicles, renewable energy systems, fast chargers, and compact power supplies has created stronger demand for higher voltage capability, lower loss, faster switching, and better thermal stability. This paper compares three important materials used in power devices: silicon (Si), silicon carbide (SiC), and gallium nitride (GaN). The discussion focuses on material properties that directly affect device behavior, including bandgap, breakdown field, thermal conductivity, and carrier transport. The analysis shows that silicon still offers strong advantages in cost, manufacturing maturity, and broad industrial use. SiC is more suitable for high-voltage and high-temperature operation, while GaN is especially attractive for high-frequency and fast-switching applications. The main conclusion is that SiC and GaN are expanding rapidly because they overcome performance limits inherent to silicon. However, material selection in power electronics still depends strongly on the specific requirements of each application.