Solar-driven semiconductor photocatalytic water splitting for hydrogen production is regarded as a green and sustainable approach to address energy and environmental challenges.However,traditional wide-bandgap semiconductor photocatalysts generally suffer from insufficient visible-light response and rapid photogenerated charge carrier recombination,leading to relatively low solar-to-hydrogen conversion efficiency.Therefore,the construction of photocatalytic material systems that combine high activity with low cost has become a critical scientific issue in this field.BaTiO3(BTO)has attracted considerable attention owing to its nontoxicity,low cost,a nd robust structural stability.Nevertheless,its wide bandgap limits its visible-light response,and inefficient separation and transport of photogenerated charge carriers severely restrict its photocatalytic performance.To address these issues,this study proposes a synergistic strategy combining Nb doping with a core-shell-like heterojunction construction to systematically regulate the band structure and interfacial charge behavior of BTO,aiming to achieve low-cost and efficient photocatalytic water splitting for hydrogen production.The chemical structure and morphology of the as-prepared materials were systematically characterized using multiple techniques,followed by a comparative evaluation of photocatalytic performance and an in-depth analysis of the hydrogen production mechanism.The results demonstrate that an appropriate Nb doping level(2%)induces a negative shift in the conduction band edge and significantly improves bulk charge transport,thereby enhancing the photoreduction capability and increasing the hydrogen production rate to 1535.3 μmol·g-1·h-1,which is approximately 4.6 times that of pristine BTO.Subsequently,a Nb-BTO/CN core-shell-like heterojunction was constructed on the surface of Nb-BTO.This configuration optimizes the energy band alignment and markedly broadens the light absorption range.The formation of an S-scheme heterojunction between Nb-BTO and carbon nitride(CN),together with the built-in electric field at the interface,significantly promotes the spatial separation and directional migration of photogenerated charge carriers both within the bulk phase and across the heterointerface.In this S-scheme system,the internal electric field drives electrons from the CN conduction band to recombine with holes from the Nb-BTO valence band while preserving the highly energetic electrons in the Nb-BTO conduction band for efficient proton reduction.Importantly,this configuration preserves the maximum redox capability of the spatially separated electrons and holes.Consequently,the hydrogen production rate of the resultant composite photocatalyst reaches 2993.9 μmol·g-1·h-1,corresponding to 9.2 times and 2.0 times that of pristine BTO and Nb-BTO,respectively.The results demonstrate the synergistic effects of Nb doping and CN core-shell-like heterojunction engineering on energy band modulation and interfacial charge carrier dynamics in BTO-based photocatalysts.Specifically,Nb doping shifts the conduction band edge of BTO toward a more negative potential and improves crystallinity as well as bulk carrier transport characteristics.Meanwhile,the S-scheme heterojunction constructed with CN further refines the charge transfer kinetics and enhances the thermodynamic driving force for reduction reactions.The synergy arises from the fact that Nb doping optimizes the bulk properties of BTO,while the CN shell extends light absorption and facilitates interfacial charge separation.Collectively,these synergistic modifications substantially enhance photocatalytic hydrogen production performance.This work provides a rational material design strategy and an experimental basis for constructing high-efficiency titanate-based photocatalytic hydrogen generation systems and offers new insights into the cooperative use of elemental doping and heterojunction engineering in wide-bandgap oxide photocatalysts.
In recent years, artificial intelligence (AI) has witnessed tremendous progress, particularly in the field of engineering applications. From simple machine learning (ML) models and deep learning (DL) models equipped with complex image processing capabilities to the rapidly advancing large language models, these have demonstrated formidable capabilities in engineering applications. Microfluidics, a crucial technology in chemical synthesis and life sciences, integrated with AI has emerged as a significant trend in the field of microfluidics. The combination of AI’s powerful data processing capabilities with the high-throughput generation, high-precision controllability, and rapid reaction analysis capabilities of microfluidics provides a powerful toolkit for fields such as materials science, chemical reaction, and biomedicine. Compared to traditional manual analysis methods, AI-assisted microfluidic technology provides faster processing speeds and reduced human intervention to address the challenges in conventional microfluidics, such as reliance on researcher experience, poor experimental repeatability, and time-consuming optimization processes. Simultaneously, AI models can facilitate investigations into fundamental microfluidic principles. This represents a highly promising direction, yet literature that systematically summarizes and elaborates on this emerging interdisciplinary field remains scarce. This paper systematically reviews the research progress of AI-assisted microfluidic technology. We start by introducing the AI models, which are employed in the microfluidics domain, including four common ML models: tree-based models, support vector machines (SVM), DL, and reinforcement learning (RL). Tree models typically possess strong interpretability. SVM is a common classification model suitable for complex data. DL models are frequently used image processing models that achieve functions such as object detection and image classification by simulating the human brain via neural networks. RL is a distinct trial-and-error AI algorithm that continuously enhances its own performance through interaction with the environment. Next, we discuss the applications of AI-assisted microfluidic technology from several aspects: microfluidic droplet generation, microreactor optimization design, micro/nanomaterial synthesis, catalytic reactions, and biomedical detection. Regarding droplet microfluidics, we examine how AI models predict the generation performance and employ explainable frameworks to investigate the underlying factors governing these processes. We showcase the use of DL for flow pattern recognition and the real-time tracking of droplets and bubbles. Turning to microreactors, we analyze the integration of ML for structural design and performance optimization. In the realm of micro/nanomaterial synthesis, we explore AI-driven approaches for performance prediction and the construction of autonomous synthesis platforms. In the field of chemical reactions, we discuss the application of AI to identify optimal reaction conditions and enhance substance detection. Finally, we discuss the advancement of AI in cell sorting, high-precision detection, and the forecasting of cellular changes within the biomedical field. In conclusion, this review provides an outlook on the future development of this interdisciplinary field. We propose four perspectives: the expand of standardized protocols and highly versatile interfaces, the development of efficient label-free, semi-supervised models and high-precision models suitable for small datasets; the creation of end-to-end AI models that bypass complex feature extraction steps; and the integration of AI models with integrated microfluidic chips to develop automated platforms and realize innovative microfluidic application modes.
Ultrasound image segmentation is critical for accurate diagnosis and effective treatment planning.Although numerous deep learning methods have been developed for organ and lesion segmentation in ultrasound images,their performance remains limited by indistinct anatomical boundaries and high speckle noise.To solve these problems,this study improves the U-Net model and proposes a network model,MFDU-Net,based on multiscale feature extraction and frequency attention denoising.First,a multi-resolution input(MI)module is introduced at the network input end to select pooling windows of different sizes for size reduction from the feature maps extracted after one convolution of the input image.The feature maps are then overlaid with the corresponding-size encoding layers by channels for feature fusion,and the model provides input information for different encoding layers for subsequent encoding operations.Second,the convolutional encoding layer of the original U-Net is replaced with a multi-scale separation convolution(MSC)module to capture the image feature information at different scales.Unlike existing multiscale methods that use different sizes of convolution kernels in parallel,the proposed framework employs multiple feature extraction branches with the same 3×3 convolution kernel.By c hanging the number of convolutions,the same effect as changing the size of the convolution kernel is achieved,enabling multiscale feature extraction and allowing the network to adapt to targets of different sizes and positions,thus enhancing its expressive power and robustness.Finally,considering that the U-Net skip connection does not consider the direct transmission of noise and irrelevant information when fusing the encoder and decoder features,a frequency-denoising attention module is introduced in the skip connection.Drawing on the ideas of the CBAM module,frequency-space attention and frequency-channel attention modules are designed to remove noise and weight feature information,respectively.This overcomes the limitations of spatial domain feature learning commonly found in most medical image segmentation networks and reduces the impact of irrelevant information and noise on the model segmentation performance.U-Net,DeepLabv3+,Att U-Net,ACC-Unet,and FCRNet were used as comparative methods to evaluate the proposed MFDU-Net on two self-built kidney ultrasound datasets and two publicly available datasets:DDTI and ISIC2018.Considering the small number of collected images and the requirement of segmentation models for the training set size,data augmentation was performed on all the datasets used in this study.Six quantitative evaluation indicators were selected:Jaccard coefficient,recall rate,accuracy rate,Dice coefficient,HD95,and ASSD.Compared with the baseline U-Net,MFDU-Net improves the dice index by 5.47,9.66,4.83,and 3.59 percentage points,respectively,and the HD95 distance index decreased by 36.59,5.76,48.25 and 25.75,respectively,achieving good segmentation results.To verify the effectiveness of the improved modules designed in this study,ablation experiments were conducted using a kidney ultrasound dataset.The experimental results show that,compared with existing advanced medical image segmentation models,MFDU-Net demonstrates superior performance in segmenting organs and lesions.
Currently, machine learning and artificial intelligence (AI) methods are predominantly applied to the research and development of high-entropy alloys. Although high-entropy alloys exhibit superior properties and significant industrial potential, their manufacturing costs remain prohibitive, and established industrial applications are currently limited. Conversely, the vast majority of steel products are low-entropy alloys primarily composed of Fe, C, and trace amounts of Mn, Cr, Ni, Nb, and other elements. Presently, Chinese iron and steel enterprises rely largely on traditional “trial and error” methods to develop new steel grades. This traditional “research and imitation” process requires time and high research and development (R&D) expenditure. To enhance R&D efficiency, the utilization of AI methods to facilitate new product development, optimize existing process parameters of existing products, and improve product quality is a critical technological requirement for the industry. This paper proposes a novel product development model based on AI methods, utilizing IF steel R&D as an industrial case study. Given the high-dimensionality, strong coupling, and nonlinearity characteristic of industrial production data, eight machine learning models were evaluated to predict the mechanical properties of IF steel. To ensure model generalization, the data was regularized, the super-parameters were optimized, and the training set underwent five-fold cross-validation. The prediction accuracy and applicable scenarios for each model are subsequently discussed. Among the eight models for predicting the mechanical properties of interstitial-free (IF) steel, the random forest (RF) and deep neural network (DNN) models demonstrated superior accuracy, with R2 > 0.97. Furthermore, the kernel principal component analysis (KPCA) model was employed to reduce the dimensionality of high-dimensional components and process parameter features into two-dimensional principal component vectors. A material fingerprint was then established using the Gaussian mixture module (GMM) model. This fingerprinting technique facilitated the visualization of high-dimensional feature data, enabling the analysis and observation of data distributions and the identification of potential mining intervals and the spatial location of the generated data. These potential spaces represent optimal search intervals for the discovery of novel materials. Finally, a generative model based on Wasserstein auto-encoders (WAE) was proposed to explore the potential composition and process parameter spaces. The WAE model can generate thousands of potentially valuable samples from which specific candidates can be selected for industrial testing. This digital “trial and error” approach serves as a robust alternative to traditional methods, accelerating product development while significantly reducing R&D costs and durations.
During hot stamping,the sliding contact between an aluminum alloy sheet and H13 tool steel often results in severe wear and adhesive material transfer.These tribological phenomena can significantly degrade forming quality and reduce die service life.In this study,the high-temperature tribological behavior of a heated 7075 aluminum alloy sheet sliding against H13 steel was investigated under conditions designed to simulate the aluminum hot-stamping practice.Experiments were conducted using a self-developed strip-type high-temperature friction tester.Based on the actual process sequence for aluminum hot stamping,the tester was used to reproduce a combined"solution treatment-forming-quenching"cycle such that the friction pair experienced thermal and mechanical histories similar to those encountered in industrial production.The main purpose was to elucidate how normal pressure and die surface condition i nfluenced friction evolution,adhesive transfer,and the dominant wear mechanisms at elevated temperature.Three representative surface conditions of H13 steel were examined:(I)quenched and tempered(Q&T);(II)plasma nitrided after Q&T(PN);and(III)physical vapor deposition TiN coating applied after Q&T(PVD-TiN).Under different normal loads,the coefficient of friction and extent of adhesive wear/transfer were quantified and the friction and wear mechanisms were analyzed for each surface condition.The results show that as the normal pressure increases,the friction coefficient generally rises and adhesive wear becomes more severe.Higher contact pressure increases the real area of contact and promotes stronger interfacial bonding between the softened,high-temperature 7075 aluminum and steel surface,thereby accelerating sticking,tearing,and the formation of transfer layers.In contrast,both the PVD-TiN coating and plasma-nitrided layer exhibit excellent friction-reducing performance compared with the baseline Q&T surface,indicating that surface engineering is an effective approach for mitigating galling during aluminum hot stamping.At a moderate pressure of 4.5 MPa,the PVD-TiN surface produces a lower coefficient of friction than the nitrided surface.However,because the TiN coating is relatively thin,it can undergo local spallation under combined thermal effects and tangential shear.Once delamination occurs,the fresh metallic substrate is exposed,providing highly reactive sites that facilitate strong adhesion to the aluminum sheet.As a result,despite the lower measured friction coefficient,adhesive wear on the PVD-TiN die surface is more pronounced than on the nitrided die under this intermediate-pressure condition.By comparison,the plasma-nitrided layer contains nitride compound phases,such as Fe3N and Fe4N,which create a harder and more chemically stable near-surface region and can effectively suppress adhesion,thereby reducing material pickup and transfer.At a higher pressure of 5.4 MPa,the PVD-TiNTiN-modified die not only maintains a relatively low coefficient of friction but also benefits from TiN-related hard particles formed or retained at the interface.These particles can function as protective third-body constituents,reducing direct metal-to-metal contact and significantly alleviating adhesive wear.Under this high-pressure condition,the overall anti-galling performance of the PVD-TiN surface is superior to that of the nitrided surface.In contrast,the nitrided die under high pressure tends to experience plastic deformation,which generates agglomerated Fe-Cr alloy particles and worsens the interfacial contact state,ultimately destabilizing friction behavior and aggravating wear.In summary,PVD-TiN provides a more pronounced friction-reduction effect than plasma nitriding in the simulated 7075/H13 hot-stamping tribosystem and is particularly suitable for medium-to-low pressure hot-stamping applications where low friction is critical.Plasma nitriding,on the other hand,offers superior structural stability of the modified layer,making it better suited to heavy-load,high-pressure stamping scenarios in which resistance to deformation and long-term surface integrity are essential.
