Wafer image denoising is essential in semiconductor manufacturing, impacting yield and defect detection accuracy. Unlike general images, wafer images exhibit unique micrometer-scale structures, low-contrast defects, and fine edges, making denoising particularly challenging. However, traditional methods perform poorly with complex noise, and supervised learning methods rely on clean labeled data that is difficult to acquire, limiting robustness in practical applications. In this study, we propose an adaptive dual-domain fusion and self-supervised network (ADF-SN). First, noise suppression and edge enhancement are performed in the transform and spatial domains to obtain complementary feature representations. Then, an adaptive dual-domain fusion mechanism achieves a dynamic balance between denoising and structure preservation. Finally, we present a dual attention self-supervised network (DASSN), which exploits redundant information from multiple noisy observations for high-quality restoration without requiring clean reference images. Experimental results on our Wafer4 dataset show that ADF-SN achieves an average PSNR of 31.2722 dB and an a structural similarity index (SSIM) of 0.8461. This method has been deployed in wafer inspection equipment, completing the closed-loop from algorithm to application and demonstrating effectiveness, robustness, and practicality in semiconductor manufacturing.

We propose a generalized reinforcement learning (RL) approach for personalized autofocus control in wafer micro-imaging, aiming to address the issue of inconsistent focal distances across different wafer regions. Our method integrates region selection with focus control by creating a deep network that estimates focal distances based on the current image frame. Through multiple rounds of image capture and evaluation in the RL framework, the network is fine-tuned to develop personalized models that predict optimal focal distances for interest regions based on engineer feedback. The Gaussian policy gradient algorithm is used to update the model's policy network during the fine-tuning process. To validate our approach, we constructed a dataset of wafer images captured at varying focal distances for training and prediction. Experimental results show that our network not only resolves the limited generalization of focus adjustment algorithms across regions but also achieves an average improvement of approximately 4.0% in focusing quality. This method eliminates the need for manual focus adjustment and region selection in wafer inspection, offering new insights for improving wafer micro-imaging quality.

Inverse design of materials (IDMs) has increasingly becoming an integral part of the manufacturing process. Despite significant advance in algorithmics for this task, it has still not yet achieved widespread acceptance, mainly due to a gap between the capabilities of current state-of-the-art algorithmics and the requirements of industry. For IDM to be truly useful for industrial use cases, it needs to be much more flexible and amenable to modifications stemming from specific material specifications and constraints. In this work, a symmetry and interface-guided diffusion model, a robust and intuitive diffusion generative framework, was introduced for target design of two-dimensional semiconductor structures. Our framework addresses the lack of robust and highly-controllable generative models in this important subfield of materials engineering, and provides a basis for a much more enlarged role of generative models in 2D semiconductor structures.

Ensuring that a robot employing demonstration learning models can simultaneously achieve accurate trajectory tracking of demonstrated paths and effective avoidance of moving obstacles in dynamic environments remains a critical research challenge. This paper proposes a real-time trajectory planning framework based on an enhanced artificial potential field (APF) approach to address this dual-objective problem. Specifically, the proposed method deploys a sequence of virtual target points along the demonstrated trajectory to guarantee both path-following precision and goal convergence for robotic systems. A dynamic obstacle repulsion model is developed by integrating velocity-coupled and acceleration-associated force components, enabling proactive obstacle motion anticipation and adaptive trajectory reconfiguration. Furthermore, a probabilistic obstacle motion prediction framework is established through motion pattern analysis to actively optimize the robot's motion strategy and reduce tracking errors. Simulation-based experimental results demonstrate that, under complex obstacle motion scenarios, the proposed method achieves a 55.8% reduction in trajectory tracking error compared with recently proposed improved APF methods and a 41.5% decrease relative to Dynamic Movement Primitives (DMP) baselines. These quantitative improvements validate the framework's superior robustness and safety performance in unstructured environments, with all evaluations systematically conducted in simulated settings.

With the widespread use of lithium-ion batteries in various application fields, accurate prediction of battery state of health (SOH) has become an important research topic to ensure battery performance and safety. To improve the accuracy of SOH prediction, this paper proposes a novel approach that combines multidimensional feature extraction and a transformer–LSTM fusion model. This method extracts time domain, frequency domain, and time dimension features from voltage, energy, and temperature curves. It evaluates feature importance, removes redundancy, and focuses on key features most relevant to SOH. Then, using the self-attention mechanism of transformer and the long-term dependency capture ability of LSTM, an efficient fusion model is constructed to further improve the accuracy and stability of SOH prediction. The proposed method is validated based on the cycling data from 124 commercial lithium iron phosphate/graphite batteries under fast-charging conditions. Compared with existing methods, the proposed approach effectively extracts key features closely related to SOH and builds models based on these features. It achieves a prediction accuracy exceeding 50% and demonstrates superior generalization performance relative to current methods.

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries is of great significance for preventing safety hazards such as thermal runaway caused by battery aging, optimizing battery maintenance strategies, and improving battery management efficiency. To solve the problem of early prediction of lithium battery RUL, this paper proposes a prediction method based on the GPR-SSA fusion model. This method uses the first 100 charge and discharge cycle data of lithium batteries to extract time domain and frequency domain features, and removes redundant features through feature importance evaluation, overcomes the limitation that the initial capacity degradation of the battery is not significant, and significantly improves the early prediction performance. At the same time, a Gaussian process regression model is constructed, and the model parameters are optimized in combination with an intelligent search strategy to maximize the prediction accuracy of the model. To verify the effectiveness of the method, this paper carried out an early prediction experiment of RUL based on the cycle data of 124 commercial lithium iron phosphate/graphite batteries under fast charging conditions. The experimental results show that the proposed method is superior to other comparative methods in terms of early prediction accuracy, and shows reliable prediction ability and good generalization performance on the test data set.

The circle fitting algorithm for wafer pre-aligners plays a pivotal role in wafer transmission technology. Traditional wafer pre-aligner adopts the method of one-time fitting after collecting complete data, resulting in low accuracy, high computational burden, and poor robustness. In this study, we propose an innovative online segmented fitting method based on a linear fitting model, which uses a recursive weighted least-squares (RWLS) algorithm to enhance the accuracy of fitting segmented data. To address computational complexities and improve algorithm robustness, we introduce the progressive sample consensus (PROSAC) algorithm, employing a uniform segmented strategy for rapid online selection of high-quality sampled data. Simulation experiments reveal that our method reduces total alignment time by 55.5% and decreases a mean fitting error by 78.8% compared to traditional one-time fitting methods. Furthermore, actual experimental results validate that our proposed method and algorithm not only reduce alignment time but also exhibit superior fitting accuracy and interference resistance.

Combining a well-trained deep potential (DP) model and a high-efficiency hybrid differential evolution (HDE) algorithm to search for the lowest-energy structures of CoN (N = 11–50) clusters.

In dynamic environments, real-time trajectory planners are required to generate smooth trajectories. However, trajectory planners based on real-time sampling often produce jerky trajectories that necessitate post-processing steps for smoothing. Existing local smoothing methods may result in trajectories that collide with obstacles due to the lack of a direct connection between the smoothing process and trajectory optimization. To address this limitation, this paper proposes a novel trajectory-smoothing method that considers obstacle constraints in real time. By introducing virtual attractive forces from original trajectory points and virtual repulsive forces from obstacles, the resultant force guides the generation of smooth trajectories. This approach enables parallel execution with the trajectory-planning process and requires low computational overhead. Experimental validation in different scenarios demonstrates that the proposed method not only achieves real-time trajectory smoothing but also effectively avoids obstacles.

Despite significant advancements in leveraging artificial intelligence (AI) for drug design, materials science, and other fields, the question of how each dataset feature influences a target metric—essential for constructing better predictive models and targeted materials design—remains largely unaddressed. In this study, we explored the application of interpretable machine learning (ML) techniques to the inverse design of two-dimensional (2D) semiconductor materials, a critical yet underexplored area within the AI4Science domain. Our approach utilized a dataset from the C2DB database, incorporating advanced feature engineering and data imputation strategies to predict material stability, a key determinant of a materials industrial and academic value. Through the calculation of Shapley additive explanation scores and counterfactual analysis, we provided a nuanced understanding of feature contributions toward material stability, enabling the targeted design of 2D semiconductors with optimized properties. This work not only fills the gap in the current literature by emphasizing the role of interpretability in materials design but also demonstrates the potential of interpretable ML in guiding the development of novel materials with enhanced performance characteristics.

Improving evacuation efficiency is a central concern in evacuation simulation research. Incorrect feedback information can affect the effectiveness of evacuation control in partially observable evacuation. Hence, we introduce a framework for evacuation guidance control, emphasizing data prediction and correction to mitigate the impact of abnormal observed data. The framework is built upon force-driven Cellular Automaton (CA) models and employs a data correction module to rectify abnormal information. Guided by this framework, we implement the data correction module’s functionality using Back Propagation (BP) neural networks. We utilize historical simulation data to train the BP network, obtaining a model for correcting abnormal data. Additionally, we integrate the data correction model with density control algorithms to facilitate pedestrian flow management in abnormal evacuation scenarios. Subsequently, we conduct two comparative simulation experiments to verify the algorithm’s effectiveness. One experiment utilizes an abnormal data replacement method, while the other employs a data correction method. The results show that the method of abnormal data replacement is simple, but it cannot improve the control efficiency in abnormal evacuation scenarios. The data correction method proposed in this paper can effectively improve evacuation efficiency and alleviate the congestion at exits caused by abnormal data, which reduces evacuation efficiency. The results are expected to provide insights into improving evacuation systems’ efficiency and personnel safety.

Classifying the sentiment of user comments on a website is a crucial task within Natural Language Processing (NLP). Conducting sentiment analysis can aid businesses in gaining a more profound comprehension and examination of users’ emotional inclinations towards products or services. This study introduces a sentiment classification model that combines the Transformer and BLSTM architectures to analyze the sentiment of user comments on movie and book websites. By incorporating the strengths of both Transformer and BLSTM, the proposed model mitigates the issue of vanishing gradient by scrutinizing inputs within a long-term context using BLSTM. It employs the multi-head attention mechanism of the Transformer to extract features and capture significant semantic details within the comments. Furthermore, the joint model combines the TF-IDF weights with the vector space, which improves the embedding process. The proposed model’s effectiveness was evaluated by categorizing the sentiment of user comments on publicly available datasets containing more than 20,000 movie and book comments. The results indicate that the proposed model is superior to LSTM and CNN in sentiment classification tasks. Moreover, the proposed approach has demonstrated significant improvements, particularly in the training set, achieving an accuracy of 93.81%.

The exposure control of wafer motion imaging systems greatly affects image quality.Tradition-al spatial domain-based algorithms are complex and overlook differences in die features,causing uneven ex-posure.To solve this,we propose a method using frequency domain evaluation through image partition-ing.This method uses a block-based dual-threshold segmentation algorithm to adaptively segment image features.By combining high-pass filtering with a Gaussian pyramid algorithm,it efficiently extracts high-frequency information.A region-weighted evaluation function is crafted to evaluate uniform exposure ef-fects.Furthermore,a decision tree-based search algorithm with variable steps is introduced to quickly find optimal exposure parameters.Experiments show this algorithm improves image quality by 1.34%and re-duces exposure adjustment time by 61.3%,ensuring fast imaging quality in wafer motion imaging.

The increasing wafer yield and shrinking size pose challenges for real-time defect inspection using a single computer. To address this, we propose enhancing real-time performance by increasing computing devices. However, uneven load distribution due to device performance or defects variations can reduce inspection efficiency. We introduce computer cluster load balancing to solve this, utilizing a load-balancing model and an objective function. We propose the adaptive discrete quantum particle swarm optimization algorithm (ADQPSO) for efficient load balancing and implement the adaptive dynamic smooth weighted round-robin algorithm based on ADQPSO. Experimental results demonstrate that our algorithm achieves the fastest execution speed and up to a 24% improvement in performance. Our approach significantly improves real-time performance and efficiency in wafer surface defect inspection.

Contemporary intelligent robots typically necessitate the implementation of a distributed sensing network system within their digits to execute a range of intricate tasks collaboratively. In this investigation, we have introduced a novel fiber sensor founded on an enhanced fiber-loop ring-down spectral technology. This sensor employs double differential ring-down cavities to mitigate prevalent sources of interference, such as temperature and humidity. Consequently, it exhibits a robust linear response with remarkable sensitivity, quantified at 105.2 ns/N for force measurement and 53.2 ns/mm for displacement measurement, as validated by experimental outcomes. Furthermore, the sensor boasts a compact and diminutive sensing structure that can be effortlessly integrated into the robotic finger casing. Of utmost significance, it facilitates a cost-effective time-division multiplexing system for distributed measurements. Considering these distinctive advantages, this sensor lends itself to the establishment of a practical robot-finger distributed network system, characterized by superior performance attributes, encompassing high precision (up to 19 mu m for contact detection and 0.0095 N for grasping and other manipulations), relatively modest expenses, and versatile applicability across a broad spectrum of scenarios, in stark contrast to the presently prevalent systems. In light of these noteworthy merits, the developed sensor exhibits tremendous potential for integration into robotics applications.

Defects appeared in the printed circuit board (PCB) will pose a serious damage on the following procedure. Image based inspection methods have been proposed to improve the efficiency and reliability of PCB defect detection compared to manual inspection. The machine learning and deep learning detection methods are popular one, however, they are complex, time consumption and require lots of labeled samples. Thus, we conduct the PCB defect detection and classification by using the template-based algorithm. To realize an accurate registration, the region of interest (ROI) among input image is first computed by utilizing the Grab Cut method. Furthermore, to ensure the complete overlap of feature points between the test image and template image, a perspective transformation based on four vertexes calculating is introduced. Once the different shape and posture images are transformed into a uniform imaging plane, a subtraction operation is used to extract the features of various defects. Experiments on a public data set prove the efficiency of our proposed method.

Determining the optimal structures and clarifying the corresponding hierarchical evolution of transition metal clusters are of fundamental importance for their applications. The global optimization of clusters containing a large number of atoms, however, is a vastly challenging task encountered in many fields of physics and chemistry. In this work, a high-efficiency self-adaptive differential evolution with neighborhood search (SaNSDE) algorithm, which introduced an optimized cross-operation and an improved Basin Hopping module, was employed to search the lowest-energy structures of CoN, PtN, and FeN (N = 3-200) clusters. The performance of the SaNSDE algorithm was first evaluated by comparing our results with the parallel results collected in the Cambridge Cluster Database (CCD). Subsequently, different analytical methods were introduced to investigate the structural and energetic properties of these clusters systematically, and special attention was paid to elucidating the structural evolution with cluster size by exploring their overall shape, atomic arrangement, structural similarity, and growth pattern. By comparison with those results listed in the CCD, 13 lower-energy structures of FeN clusters were discovered. Moreover, our results reveal that the clusters of three metals had different magic numbers with superior stable structures, most of which possessed high symmetry. The structural evolution of Co, Pt, and Fe clusters could be, respectively, considered as predominantly closed-shell icosahedral, Marks decahedral, and disordered icosahedral-ring growth. Further, the formation of shell structures was discovered, and the clusters with hcp-, fcc-, and bcc-like configurations were ascertained. Nevertheless, the growth of the clusters was not simply atom-to-atom piling up on a given cluster despite gradual saturation of the coordination number toward its bulk limit. Our work identifies the general growth trends for such a wide region of cluster sizes, which would be unbearably expensive in first-principles calculations, and advances the development of global optimization algorithms for the structural prediction of clusters.

Fiber loop ring-down spectroscopy (FLRDS) is a highly sensitive spectroscopic technology. Here, a spectral processing method is proposed for low-cost fiber loop ring-down systems. In the proposed method, the amplitude modulation theory is used to establish a new ring-down spectral model, and the particle swarm optimization algorithm is introduced into FLRDS. The experimental results demonstrate that the proposed method can effectively process the ring-down spectral signals acquired from the low-cost fiber loop ring-down system. Moreover, this method demonstrates higher accuracy and stronger adaptability than the current widely used method. Owing to its excellent performance, the proposed method has considerable potential to be applied in low-cost fiber loop ring-down systems.

A composite dynamic movement primitives algorithm based on Dirichlet process clustering and Gaussian mixture model is proposed to address the problems of low efficiency of parameter estimation and insufficient generalization ability in demonstration learning. To achieve the real-time estimation of Gaussian mixture model parameters, the Dirichlet clustering algorithm based on the distance threshold is used to perform online clustering of demo trajectory points, and the Welford formula is introduced to update the parameters to improve the efficiency of parameter estimation. After obtaining the trajectory distribution characteristics, the Gaussian mixture regression trajectories are encoded by using the dynamic movement primitives to improve the trajectory generalization. To evaluate the effectiveness of the algorithm, trajectory reachability and similarity metrics are introduced to evaluate the learning generalization ability of the algorithm, and demonstration learning experiments based on handwritten letter trajectories and robot kinesthetic demonstrations are designed. Experimental results show that the average parameter estimation time of the proposed composite dynamic movement primitive algorithm is only 0.052 ms, which has the ability of fast trajectory reproduction and generalization.

Automatic detection of product surface defects is critical to product quality control in the production process. To improve the performance of surface defect detection methods based on deep learning in practical application, we propose a fast surface defect detection network called the Dense-Yolo network, which combines the strengths of DenseNet and YOLOv3. In this network, we adapt a lightweight backbone network with improved densely connected blocks to enhance the performance on shallow features utilization and detection speed. Moreover, the feature pyramid network (FPN) of YOLOv3 is improved to increase the recall of tiny defects and the overall positioning accuracy. Meanwhile, we proposed an online multi-angle template matching method based on normalized cross-correlation to locate the detection area. Furthermore, the improved template matching method improves the detection speed and reduces the adverse effects of the background. To verify the effectiveness of the above improvement, we carry out comparative experiments over two private data sets and one public data set. The obtained results have demonstrated the advantage of the proposed method over Faster-RCNN, SSD, YOLOv3 and YOLOv5s in terms of the mean average precision (mAP) and detection speed.