搜索结果：Neural computing & applications

A computer's clock rate ultimately determines the minimum time between sequential operations or instructions. Despite exponential advances in electronic computer performance due to Moore's Law and increasingly parallel system architectures, clock rates of conventional processors have remained stagnant at ~5 GHz for nearly two decades. This creates a significant challenge for applications requiring real-time processing or control of ultrafast information systems. Here, we break this barrier by proposing and experimentally demonstrating computing based on an all-optical recurrent neural network leveraging the ultrafast characteristics of linear and nonlinear optical operations while circumventing electronic bottlenecks. The all-optical computer realizes linear operations, nonlinear functions, and memory entirely in the optical domain with accuracy surpassing a purely linear model up to 80 GHz clock rates depending on the task. We experimentally demonstrate a prototypical task of noisy waveform classification as well as perform ultrafast in-situ analysis of the soliton states from integrated optical microresonators. We further illustrate the application of the architecture for generative artificial intelligence based on quantum fluctuations to generate images even in the absence of input optical signals. Our results highlight the potential of all-optical computing beyond what can be achieved with digital electronics by utilizing ultrafast linear, nonlinear, and memory functions and quantum fluctuations.

Surrogate models for the rapid inference of nonlinear boundary value problems in mechanics are helpful in a broad range of engineering applications. However, effective surrogate modeling of applications involving the contact of deformable bodies, especially in the context of varying geometries, is still an open issue. In particular, existing methods are confined to rigid body contact or, at best, contact between rigid and soft objects with well-defined contact planes. Furthermore, they employ contact or collision detection filters that serve as a rapid test but use only the necessary and not sufficient conditions for detection. In this work, we present a graph neural network architecture that utilizes continuous collision detection and, for the first time, incorporates sufficient conditions designed for contact between soft deformable bodies. We test its performance on two benchmarks, including a problem in soft tissue mechanics of predicting the closed state of a bioprosthetic aortic valve. We find a regularizing effect on adding additional contact terms to the loss function, leading to better generalization of the network. These benefits hold for simple contact at similar planes and element normal angles, and complex contact at differing planes and element normal angles. We also demonstrate that the framework can handle varying reference geometries. However, such benefits come with high computational costs during training, resulting in a trade-off that may not always be favorable. We quantify the training cost and the resulting inference speedups on various hardware architectures. Importantly, our graph neural network implementation results in up to a hundred- to thousand-fold speedup on GPU, and twenty- to two hundred-fold speedup on CPU for our benchmark problems at inference.

This model has significant applications in advanced thermal and bioengineering systems where coupled fluid flow, heat transport, and microbial dynamics are vital. It is applicable to the design of bioreactors and microbial fuel cells, where Oxytactic bacteria improve mixing and mass transfer efficiency. The model optimizes heat transport in nanofluid-based cooling systems, particularly under magnetic fields (MHD), by employing the Xue thermal conductivity formulation. The inclusion of LTNE (local thermal non-equilibrium) effects and activation energy makes it applicable to chemical reactors, catalytic processes, and biomedical drug delivery, all of which rely on reaction rates and temperature gradients. This study used a neural network trained using the Bayesian-Regularized approach to inspect the impact of LTNE on oxytactic microorganisms in bioconvection flow of a HNF (hybrid nanofluid) with activation energy. There are several uses for Oxytactic microbes, especially in environments and industries. In wastewater treatment, oxytactic bacteria are frequently utilized to increase oxygen distribution in bioreactors and accelerators for the breakdown of organic contaminants. These bacteria move quickly in response to gradients in oxygen concentration. In many biotechnological applications, Oxytactic microbes must be able to situate themselves within fluid flows in order to efficiently combine and transfer nutrients. The concentration profile increases as the activation energy parameter's values rise.

Cable-driven planar robots have attracted increasing attention due to their simple mechanical structure and flexible workspace, while achieving accurate trajectory tracking in discrete-time settings remains challenging because of nonlinear kinematics and real-time computational requirements. In this paper, a discrete-time neural dynamics (DTND)-based online quadratic programming (QP) framework is developed for discrete-time trajectory tracking of a cable-driven planar robot with a structure consistent with typical wall-drawing (V-plotter) systems. The inverse kinematics problem of the robot is first formulated as an online QP by transforming the desired end-effector trajectory into a discrete-time velocity tracking objective. Inspired by DTND for time-varying quadratic programming, a DTND-based iterative solver is employed to compute the cable-length rates at each sampling instant, enabling real-time trajectory tracking without explicit offline optimization. The proposed method operates entirely in discrete-time form and is suitable for digital implementation. Numerical simulations based on a cable-driven planar robot demonstrate that the proposed DTND-based online QP approach achieves stable error convergence and accurate trajectory tracking. Furthermore, physical experiments on a cable-driven planar robot are conducted to validate the practical feasibility of the proposed DTND-based framework and demonstrate its applicability in real-world engineering scenarios.

Biological visual systems perceive information by processing light signals through the regulation of synaptic weights in the nervous system. Consequently, optoelectronic synaptic devices that directly respond to light stimuli and mimic synaptic plasticity hold immense potential for constructing highly efficient neuromorphic computing systems. This paper reports an optoelectronic synaptic transistor utilizing amorphous indium gallium zinc oxide (IGZO) to serve as the active channel layer of the device. Owing to its wide bandgap, the transistor shows a strong positive photoresponse under ultraviolet (UV) light, leading to substantial photocurrent enhancement. Simultaneously, applying electrical pulses suppressed the current response of the device, achieving negative modulation of synaptic weights, which successfully simulates the dynamic balance mechanism between excitatory and inhibitory effects in biological synapses. Based on this phenomenon, this work defines optoelectronic co-modulation as an operation mode that achieves bidirectional dynamic regulation of channel conductance via ultraviolet light-induced carrier excitation and electrical pulse-induced charge trapping. Leveraging the photoelectric synergistic properties of the device, we successfully simulated key biological synaptic functions including postsynaptic current (PSC), paired-pulse depression (PPD), and the transition from short-term plasticity (STP) to long-term plasticity (LTP). Based on this, we achieved fundamental "AND" and "OR" logic gate functions. Furthermore, when the device was applied to handwritten digit recognition tasks, the neural network achieved an accuracy of 90%.

Anxiety disorders pose an increasing challenge to the mental health of individuals, particularly in regions with limited healthcare access. This study investigated the potential of integrating a brain-computer interface for processing electroencephalography (EEG) data with deep learning models to accurately classify anxious and non-anxious states. In the first phase, a convolutional neural network (CNN) was developed and validated on the public GAMEEMO dataset, achieving a classification accuracy of 95.72%. In the second phase, we conducted a separate experimental validation with seven participants (aged 18-60 years) using a within-subjects design. The protocol comprised a custom Stroop test to elicit acute cognitive stress and anxiety-related arousal, followed by a guided 4-7-8 breathing exercise to induce relaxation. EEG data from this experiment were used to classify anxious versus non-anxious states with the same CNN architecture after domain adaptation. On this self-collected dataset, the CNN achieved an accuracy of 86.58%. These results demonstrate proof-of-concept transferability while highlighting the performance gap between controlled benchmark data and real-world, small-sample recordings. The deep learning model can subsequently be coupled with neurofeedback techniques to manage anxiety levels. Overall, the findings support the potential of the developed automated system for detecting stress-induced anxious states, with possible future integration into neurofeedback-based management systems.

Network functions virtualization (NFV) is an emerging technology that enables flexible service deployment for supporting the Beyond 5G/6G network. NFV transforms physical network devices into virtual network functions (VNF) over Edge Computing capabilities, thereby facilitating the agility of network services and reducing management costs. To effectively monitor Internet of Things (IoT) network resources, service function chaining (SFC) is used for its virtualizations to ensure the multi-service requirements are sufficiently in capability, scalability, and flexibility for computation workloads alignments. However, to satisfy the resource availability requirements and efficiency under several conditions, SFC reconfiguration methods face the challenges in meeting significant latency requirement of delay-sensitive applications while reaching the importance of energy saving on orchestration timespan. In this paper, we propose task management-aware SFC and orchestrating schemes, namely GNN-PPO. In this framework, we utilize the Graph Neural Network (GNN), which relies on the message-passing neural network (MPNN), to capture all the abstraction of physical resource nodes and link capabilities over MEC node states. In particularly, GNN is divided construction into two phrases: (1) GNN represents nodes for all the Mobile edge computing (MEC) nodes, which have a global view on resources of computation and communicational capabilities that could serve as carriers; (2) VNFs are transferred into graph networks by using feature-extraction MPNN to manage each VIM that seeks an optimal and reliable analysis of traffic fluctuations. Lastly, Deep Reinforcement Learning (DRL) is used to embrace the network determination in policy strategy, which utilizes a Proximal Policy Gradient (PPO). On the other hand, we propose a novel network architecture based on PPO to perform the design for the optimization of resource utilization and facilitate energy consumption on MEC servers under diverse setting scenarios, which enables continuous policy enforcement for our system. With the experimental results, we compare our proposed solution with reference schemes in terms of rewards with learning rate and batch size, average request acceptance, SFC success, packet delivery, throughput, and resource utilization ratio that confirm the scheme's scalability and practical suitability for IoT network deployment.

Estimation of hand kinematics from surface electromyography (sEMG) signals is crucial for the effective control of multi-degree-of-freedom (multi-DoF) robotic neural interfaces. However, the neural coding model has not been fully clarified, which restricts the ability to extract motor information from neural signals. Due to insufficient information mining, current methods to estimating biological kinematics have inherent limitations in accuracy, stability, and generalization ability. A scalogram image-based method for continuous estimation on multi-DoF finger joint angles was presented. Specifically, the continuous wavelet transform (CWT) was applied to sEMG signals to obtain scalogram images, and then to reduce the dimensionality of these images via digital image processing (DIP). Subsequently, the processed images were input into a convolutional neural network (CNN) to enable end-to-end autonomous learning. We selected all 40 subjects from the Ninapro DB2 dataset to verify the performance of the proposed method. In addition, we set up two independent control experiments, i.e., CNN method with sEMG image (sEMGimage-CNN) and support vector regression method with CWT (CWT-SVR). The experimental results demonstrated that the proposed method achieved an average correlation coefficient (CC) of 0.9793 ± 0.0089 and an average normalized root mean square error (nRMSE) of 0.0490 ± 0.0073, outperforming the other two methods in estimation accuracy. These results demonstrate that the proposed method provides an effective approach for continuous estimation of hand kinematics. This work opens up a new perspective for myoelectric control and potentially achieve more natural human-computer interaction (HCI) in practical application.

This review synthesizes the methodological evolution of artificial intelligence and evolutionary computation in hydrology and hydraulics, from early machine learning (artificial neural networks, support vector machines, fuzzy systems, metaheuristics) to recent advances, including deep learning (long short-term memory, Convolutional Neural Network (CNN), Convolutional Long Short-Term Memory (ConvLSTM), graph neural networks), symbolic regression, and physics-informed modeling. Evidence is organized by application domain: model calibration, streamflow forecasting, hydrological regionalization, reservoir operation, and surrogate models for hydraulic simulation. We emphasize that algorithm selection should reflect hydrologic understanding, data constraints, and operational needs - not model complexity. Beyond cataloging techniques, the shift from purely data-driven prediction toward hybrid, physics-aware, uncertainty-quantified frameworks is described here. Case studies from Mexico illustrate end-to-end workflows, including reservoir operation using dynamic programming and genetic algorithms, guide-curve optimization in the Cutzamala System, stochastic dual dynamic programming applications in the Miguel Alemán-Cerro de Oro system, and flood mapping with U-Net using Sentinel-1 imagery. A public repository provides reproducible scripts and a structured reference database. This review concludes with a research agenda addressing extreme events, uncertainty, interpretable surrogates, and standardized multibasin evaluation.

Lung cancer remains one of the leading causes of cancer-related mortality worldwide. Accurate prediction of relapse is notoriously difficult, posing substantial challenges to patient care and necessitating advanced tools to improve prognostic outcomes. MicroRNA (miRNA) expression profiles hold promise as biomarkers for predicting relapse, yet existing predictive models lack interpretability or sufficient predictive performance. Biologically informed neural networks have emerged as a modeling approach incorporating biological interpretability and predictive accuracy. Here, we introduce MiracleNet, to our knowledge the first visible neural network in which sparse connectivity is structured by the miRNA → target gene → pathway hierarchy for disease-free survival prediction from circulating miRNA in non-small-cell lung cancer and the first to expose interpretable importances jointly at all 3 biological layers, with nodes connected by prior knowledge about miRNA targets and related biological pathways. Our model, which also integrates clinical data, achieves a maximum concordance index of 0.76, demonstrates improved generalization over unconstrained neural networks of the same dimensionality (including both dense and sparse architectures lacking biological knowledge), and provides explicit biological interpretability. Our model also highlights several important biomarkers in the form of predictive miRNAs and connected biological pathways. We additionally evaluate MiracleNet under a nested repeated 80/20 protocol, augmented with patient sex and tumor stage as clinical covariates and combined across circulating free and extracellular-vesicle-associated miRNAs through early and intermediate fusion; these analyses are reported as separate sections.

Brain-computer interface (BCI) technology is undergoing rapid translation from laboratory research to clinical applications, heralding a fundamental restructuring of human-machine relationships with profound societal implications. Existing bibliometric analyses of this domain have exclusively relied on scholarly article databases, critically overlooking patent data that is essential for capturing the full spectrum of technological innovation in this highly translational field. This study aims to investigate the overall scientific and technological trajectories, key research drivers (journals, institutions, countries, and funding agencies), as well as research topics and trends in the BCI field through a comprehensive analysis of scientific and technical literature (articles and patents). BCI-related articles (11,346) and granted patents (1,551) published from 2015 to 2025 were retrieved from the Web of Science (WOS) and the incoPat database, respectively. VOSviewer, CiteSpace, Microsoft Excel and InCites were used to summarize bibliometric features. Additionally, the Disruptive Index was calculated to characterize the developmental trajectory of scientific and technological advancements in BCI field. The number of BCI articles and patents has been continuously increasing over the past decade. JOURNAL OF NEURAL ENGINEERING published the largest number of BCI articles. The Chinese Academy of Sciences ranked first in article output, while the University of California System achieved the greatest citation impact; Tianjin University from China led in patent filings. At the country level, China dominated article output and patent filings, whereas the United States attained the highest article citation impact and the most extensive international patent portfolios. The National Natural Science Foundation of China (NSFC) was the most prolific funding agency for BCI articles. Disruptive Index analysis revealed that BCI scientific research (article-based) maintained sustained growth in disruptiveness, whereas technological development (patent-based) demonstrated a pattern of fluctuation rather than consistent growth during the observation period. Five major research topics were identified, with Neural Interfaces and Motor Control attracting the greatest attention. In parallel, the leading technology categories were computer input/output interface devices (IPC: G06F3) and diagnostic measurement and human identification (IPC: A61B5). Citation analysis revealed an average knowledge transfer lag of 8.8 years, alongside limited bidirectional article-patent linkages. This study reveals a rapidly expanding yet strategically differentiated global BCI landscape dominated by China and the United States, with the former leading in output volume and the latter achieving highest citation impact and international patent portfolios. BCI scientific research maintains active disruptive potential, whereas technological development lacks a commensurate upward trajectory; this asynchrony, compounded by prolonged knowledge transfer lag and weak article-patent linkages, points to translational challenges that require strengthened mechanisms for converting scientific discovery into technical innovations.

Graph-based machine learning approaches, including Knowledge Graph Embedding (KGE) methods and Graph Neural Networks (GNNs), have emerged as powerful tools for modeling complex biomedical data. However, a systematic and clinically grounded comparison of these approaches across heterogeneous healthcare graphs, accounting for both predictive performance and real-world deployment constraints, is still lacking. We introduce AMIUGraph, a comprehensive benchmarking framework for healthcare link prediction that integrates real-world clinical data with external biomedical knowledge bases. AMIUGraph evaluates eight state-of-the-art models, of which four are knowledge graph embedding (KGE) methods DistMult, CP, ComplEx, and ConvE and four are graph neural network (GNN) architectures GCN, GraphSAGE, GAT, and GIN. The models are evaluated across three heterogeneous bipartite graphs representing Patients-Diseases, Diseases-Drugs, and Drugs-Targets interactions. Models are assessed under both transductive and inductive learning settings using accuracy, AUC, precision, recall, F1-score, and training time as evaluation metrics. Experimental results show that model performance is strongly influenced by graph structure and sparsity. GNNs consistently achieve superior predictive performance on sparse interaction graphs, particularly for Diseases-Drugs and Drugs-Targets prediction tasks. In contrast, KGE models demonstrate competitive accuracy with substantially lower computational costs in inductive clinical scenarios involving unseen patients. These trends are especially relevant in clinically realistic settings characterized by multimorbidity, such as gastrointestinal and liver diseases, where frequent patient updates and complex therapeutic interactions are common. AMIUGraph provides a clinically grounded and utility-driven benchmarking framework that jointly evaluates KGE and GNN models across multiple healthcare graph types and learning settings. The findings offer practical guidance for selecting graph-based models in medical decision-support systems, including applications in gastrointestinal healthcare, while promoting transparency and reproducibility through the public release of all datasets, protocols, and code.

Decoding short-window electroencephalography (EEG) signals is critical for low-latency brain-computer interfaces (BCIs), yet current models struggle to extract robust features under high cross-subject variability and low signal-to-noise ratios. To address this, we propose a spatiotemporal decoding framework integrating dynamic weighted permutation entropy (DWPE) with a hybrid neural network. We introduce DWPE to quantify nonlinear dynamic complexity while retaining amplitude information. These features are subsequently processed by a cascaded convolutional neural network (CNN) and bidirectional long short-term memory (BiLSTM) architecture with spatial attention, enabling the simultaneous extraction of topological patterns and temporal dependencies. The framework was evaluated on three public motor imagery datasets (hBCI, BCI Competition IV-2a, and IV-2b) using a fixed 3 s window. Empirical results demonstrate that our approach achieves an average accuracy of 84.35% and an AUC of 0.8821 on the hBCI dataset, significantly outperforming current representative recent baselines (p < 0.01). Ablation studies confirm that integrating DWPE yields a 3.89% accuracy improvement over the spatial-temporal backbone alone. With a single-sample inference time of 20.94 ms and an estimated total decision latency of approximately 3.02 s under the 3 s window setting, the proposed method provides a favorable balance between decoding accuracy and computational efficiency for short-window and near-online BCI applications.

Conventional fire monitoring systems frequently exhibit high false alarm rates, delayed response times, and a lack of closed-loop control capabilities, which severely constrain their deployment in complex real-world environments. To address these issues, this paper proposes an embedded fire detection, tracking, and extinguishing system based on multimodal information fusion and a lightweight neural model. The system follows a "Perception-Decision-Execution-Feedback" closed-loop paradigm and is implemented on a heterogeneous cooperative computing architecture comprising OpenMV4 H7 Plus and STM32F103C8T6 microcontrollers. The perception layer implements a decision-level RGB-infrared fusion mechanism that incorporates a pruned, INT8-quantized lightweight FOMO model, enabling real-time fire detection with an inference latency of 210 ms and a model size of merely 1.8 MB under resource-constrained embedded conditions. The decision layer employs a Bayesian inference-based multimodal fusion framework that effectively suppresses spurious fire interference. The vision-only false detection rate is 15.3%. After infrared fusion verification, the system-level false alarm rate is reduced to 2.0% on the interference test set. In the execution layer, a sixth-degree polynomial jet trajectory model was established and combined with an improved PID-PI dual-loop controller to enable dynamic optimization of spray angle and flow rate in real time. Experimental results demonstrate that the proposed system achieves an average fire recognition accuracy of 95.6% with a false alarm rate as low as 1.4%. Furthermore, it realizes an extinguishing accuracy better than &#xb1;5 cm within an effective operating range of 10-60 cm and completes the entire perception-to-extinguishing cycle within 8.5 s under illumination conditions ranging from 50 to 100,000 lux. These results demonstrate the excellent real-time capability, robustness, and energy efficiency of the proposed system, providing a practical and scalable solution for autonomous embedded fire-fighting applications in household, industrial, and warehouse environments.

Brain-computer interfaces (BCIs) offer the potential to restore function and augment human capabilities. However, non-invasive electroencephalography (EEG)-based BCIs still face challenges in learning efficiency and control precision, particularly for na&#xef;ve users performing complex tasks. Here, we present a sensory-guided joint learning framework that integrates human motor learning with adaptive machine learning to improve BCI training and performance. In 31 BCI-na&#xef;ve participants, the framework enabled rapid skill acquisition, achieving average online discrete accuracies of 86.0% for one-dimensional (1D) and 77.5% for two-dimensional (2D) motor imagery tasks, along with continuous control accuracies of 77.5% (1D) and 66.9% (2D). Mechanistically, tactile guidance reduced user exploration and accelerated neural adaptation, while sample reweighting aligned decoder updates with human learning trajectories. By coupling reinforcement-driven neural plasticity with adaptive algorithmic optimization, this framework advances BCI training from passive calibration to active human-machine joint learning, enabling practical and scalable neural interfaces for communication and rehabilitation.

Accurate color sensing is essential for applications ranging from materials analysis to autonomous vision, yet compact systems that can classify color directly at the device level without bulky optics or computationally intensive post-processing remain limited. Here, a machine learning-free, self-powered retinomorphic pyro-photodetector is demonstrated for direct electrical wavelength encoding from 365 to 940 nm. The multiterminal Ag/ZnO/n-Si/Ag device uses electrostatically balanced built-in potentials to convert incident wavelength into distinct photo and pyroelectric current fingerprints, enabling direct in-sensor spectral discrimination without spectral reconstruction and external neural processing. The device achieves wavelength decoding with less than 3 nm accuracy and a pyroelectric current rise-time of &#x223c;46 &#xb5;s. At the system level, direct in-sensor encoding enables end-to-end wavelength classification with 300 ms latency while reducing downstream energy and computational complexity. The encoded electrical readout enables accurate color classification and is further applied to plant health and pigment-state sensing, as well as quantification of 0% to 40% water adulteration in milk, curd, coconut water, and salt and sugar solutions, with classification accuracy above 92%. This work establishes a portable retinomorphic platform for real-time, energy-efficient, and high-precision spectral sensing, providing a general route toward direct in-sensor color recognition and wavelength classification.

Image retrieval systems rely on robust and discriminative feature descriptors for effective Content-Based Image Retrieval (CBIR). Traditional handcrafted descriptors, such as SIFT and SURF, although computationally efficient, exhibit limitations in handling large-scale datasets, domain variability, and semantic representation. To address these challenges, this paper proposes an AI-driven feature descriptor framework based on Convolutional Neural Networks (CNNs) and transfer learning. The proposed approach leverages pre-trained architectures, including ResNet and VGG, and integrates dimensionality reduction using Principal Component Analysis (PCA) along with attention-based pooling to generate compact and semantically rich feature representations. The framework is evaluated on benchmark image retrieval datasets, including Oxford5k, Paris6k, and Holidays, under standardized experimental settings. Quantitative results demonstrate that the proposed method achieves a mean Average Precision (mAP) of 85.9 %, outperforming traditional handcrafted descriptors and baseline CNN features. Furthermore, High-dimensional CNN features (e.g., 2048-d) are compressed to 128 dimensions using PCA, resulting in significant improvements in storage efficiency and retrieval speed. Overall, the proposed method provides an effective balance between retrieval accuracy, computational efficiency, and adaptability across diverse image domains, making it suitable for large-scale real-world CBIR applications.

Brain-computer interfaces (BCIs) have achieved transformative success in restoring movement and communication. However, extending these approaches to decoding or recovery of cognitive function, such as attention or memory, poses fundamentally new challenges. Cognitive BCIs will need to contend with distributed and dynamic neural processes that differ sharply from the more localized, stable representations underlying motor and language control, imposing new technical and conceptual demands. Conversely, neuromodulation, long used in neurological and psychiatric therapies, offers a complementary methodological path and initial translational applications through causal modulation of cognitive circuits. Integrating these approaches into adaptive, closed-loop systems could allow cognitive BCIs to restore mental function and bridge systems neuroscience and next-generation neurotherapeutics capable of monitoring and shaping human cognition in real time.

Semiconductor materials are widely used in electronic, optoelectronic, and energy applications. While DFT-based phonon calculations provide highly accurate assessments for dynamical stability of structures, their prohibitive computational cost poses a significant bottleneck for large-scale materials screening. Herein, we develop DynStabNet, an E(3)-equivariant graph neural network (E3GNN) framework that learns dynamical stability from phonon-informed data, enabling rapid prediction without the need for explicit phonon calculations at the inference stage. Rather than serving as a replacement for first-principles phonon calculations, DynStabNet is designed as a fast surrogate model to accelerate the prescreening of candidate structures prior to subsequent MLIP- or DFT-based phonon validation. Specifically, a crystal structure generation model is utilized to produce diverse candidates, and a pretrained machine learning potential is employed to rapidly compute their phonon spectra to construct the training data set. The E(3)-equivariant architecture enables DynStabNet to capture the complex relationships between crystal structures and their dynamical stability. Consequently, unstable configurations can be rapidly eliminated early in the materials design pipeline. The DynStabNet model achieves an accuracy of 97% while reducing the evaluation time per structure from several hours to approximately 1 ms. By dramatically accelerating the overall screening process, this approach provides an efficient framework for large-scale materials screening.