AI Scientific Illustration Gallery

A side-view schematic illustrates a two-lane roadway. On the right, a passenger vehicle is traveling on the carriageway, with the road shoulder situated between the carriageway and the roadside vegetation. A tree, complete with trunk, branches, and canopy, stands on the roadside, accompanied by shrubs and weeds at ground level. A dashed red line delineates the building clearance envelope, extending vertically from the road surface to a height of 4.5 m and horizontally across the shoulder. Portions of the tree canopy and roadside vegetation encroach upon the clearance envelope, representing clearance violations. These intrusion regions are labeled as “exceedance.” Dimensional annotations specify the clearance height (4.5 m), road shoulder width, and carriageway width. The overall style is a clean, technical schematic appropriate for transportation engineering publications.

Generate a schematic diagram of electrodeposition. The solution is a non-metallic electrolyte, and the working electrode is mesoporous carbon loaded with nanoparticles. Based on a large schematic diagram, further magnify to show that the doping and reduction of non-metallic ions on the surface of the nanoparticle unit cell are mainly controlled by the potential. A three-electrode system is used, and no text description is needed.

Method 2: Dynamic Coefficient Learning Based on Attention Mechanism - **Core Idea**: Introduce an attention layer to allow the model to automatically learn the weights of different tokens in emotion regulation. Implementation steps: (1) Construct Token Feature Representation: Extract the Token state H=[h1, h2, ....hn] (where n is the number of Tokens) from a certain layer of the model, and use the function vector (A - B) as the 'emotion query vector'. (2) Calculate Token-Emotion Attention Weights: Calculate the matching degree between each Token and the emotion query vector through scaled dot-product attention: ai = Softmax (huaye) where d is the dimension of hi. (3) Adaptive Injection of Function Vectors: Apply hi+ai.(A-B) to each Token state. This is the idea for my part.

Core Instruction: Create a clear, modular scientific diagram illustrating the multi-layered regulation of ERF proteins. 1. Overall Composition and Layout Style: Flat vector graphics, using clear lines and blocks of color. Academic cartoon style. Layout: Employ a core-periphery radial layout. Canvas: Horizontal rectangular canvas with a white background. 2. Central Core: ERF Protein Binding to DNA (placed in the center of the image) Draw a light gray DNA double helix fragment, placed horizontally. On the DNA, highlight the "GCC-box" sequence (5'-GCCGCC-3') in yellow. Draw the ERF protein: Represent it with a soft blue irregular shape (symbolizing the protein). Inside this shape: Highlight a core region in dark blue, labeled "AP2/ERF DNA-Binding Domain." This dark blue region should be directly embedded in and in contact with the DNA's GCC-box, indicating binding. In the upper region of the blue protein shape, draw an orange "flame" or "starburst" shape, labeled "Activation Domain (AD)." In the lower region of the blue protein shape, draw a gray "square lock" or "inhibition symbol" shape, labeled "Repression Domain (e.g., EAR)." 3. Peripheral Module One: Transcriptional Control (placed above the center) Left: Draw several small molecule icons—ethylene (two connected spheres with a double bond in the middle), ABA (a polycyclic structure), and a light icon (sun). Middle: Draw a gene structure diagram. A rightward arrow represents "ERF Gene Promoter," followed by a rectangular box (representing "ERF Gene"), and then a lightning symbol representing transcription producing "ERF mRNA." Right: Draw a ribosome icon and a small protein icon. Connection: Connect from left to right with dashed arrows: Signal molecules → Gene promoter → ERF gene → mRNA → Protein. Label the arrow: Transcriptional Control. 4. Peripheral Module Two: Post-translational Modifications (placed to the right of the center) Next to the central ERF protein, draw two dynamic processes: Phosphorylation: Draw a red circle with "P" inside, with an arrow pointing from a kinase icon (a small enzyme encapsulating ATP) to a specific location on the ERF protein. Ubiquitination: Draw a small yellow sphere labeled "Ub." Draw an E3 ligase icon connecting a series (e.g., 3) of "Ub" spheres to the ERF protein. An arrow points from the ubiquitinated ERF to a proteasome (barrel-shaped structure). Add the title above this module: Post-translational Modifications. 5. Peripheral Module Three: Protein Interaction Network (placed below the center) Centered on

Please generate an illustration with the title: "Mechanism of Surface Tension's Influence on Capillary Penetration Behavior of Liquids in Textile Pores (for Scientific Research)". Core elements and composition suggestions for the illustration: Overall Layout: Employ a side-by-side comparison structure, separated by a dashed line in the center. The left and right sides should display identical idealized textile fiber pore models (e.g., parallel slits or staggered networks composed of several cylindrical fibers). It is recommended to use a cross-sectional view to clearly show the morphology and position of the liquid inside the pores. Left Illustration (High Surface Tension Liquid, such as pure water): Liquid Morphology: The liquid forms a large contact angle (θ > 90°) at the pore entrance, with a convex meniscus. The main body of the liquid remains outside the pore or penetrates only slightly. Key Annotations: Use arrows and the symbol "γ_LV" (or "σ") to emphasize that the liquid has high surface tension. Clearly mark the contact angle θ_H (High) at the liquid-solid-gas three-phase contact point. A simplified force analysis diagram (local magnification) can be included, showing strong cohesive forces within the liquid molecules (inward arrows) relative to weaker solid-liquid adhesion. Result Indication: Use short dashed arrows or a lower liquid column height to indicate a low or zero capillary rise height (h), with a note such as "Weak capillary driving force" or "Limited penetration." Right Illustration (Low Surface Tension Liquid, such as a solution containing surfactants or alcohol): Liquid Morphology: The liquid forms a small contact angle (θ < 90°) at the pore entrance, with a concave meniscus. The liquid has deeply penetrated into the pores and spread along the fiber surface. Key Annotations: Use arrows and the symbol "γ_LV" to indicate low surface tension. Mark the contact angle θ_L (Low) at the contact point. A local force diagram can also be included, showing weaker cohesive forces within the molecules and stronger solid-liquid adhesion. Result Indication: Use long solid arrows and a high liquid column to indicate a significant capillary rise height (h), with a note such as "Strong capillary driving force" or "Spontaneous penetration." Unified Principle Annotation: Below the figure or in a blank space, the core formula ΔP = 2γ cosθ / r (Young-Laplace equation, simplified form) can be added, where ΔP is the capillary pressure, γ is the surface tension, θ is the contact angle, and r is the effective pore radius. Use this formula to intuitively illustrate that when γ is fixed, a decrease in θ (better wettability) leads to an increase in capillary pressure ΔP, thereby driving deeper penetration. Legends and Explanations: Include legends explaining: fiber material, liquid, gas phase, contact angle, force direction, etc. A brief statement of preconditions may be included: "Under isothermal conditions," "Ignoring the influence of gravity," or "Net height after considering gravitational equilibrium."

Alzheimer's disease is characterized by cognitive decline, memory loss, and neuroinflammation, with acetylcholinesterase (AChE) and butyrylcholinesterase (BuChE) identified as therapeutic targets. This study evaluated the in vitro and in silico anticholinesterase activity, as well as the cytotoxicity, of the essential oils of Eplingiella fruticosa and Medusantha martiusii. The oils were extracted via hydrodistillation and analyzed by GC-MS. In vitro evaluation revealed that both oils exhibited greater activity against BuChE than AChE. The oils demonstrated significant inhibition and selectivity, with IC50 values ranging from 29.2 to 612.1 µg/mL. Molecular docking studies indicated strong interactions between thymol and eudesm-7(11)-en-4-ol with the target enzymes. M. martiusii exhibited higher activity and lower cytotoxicity (IC50 of 876.5 µg/mL) compared to E. fruticosa (574.3 µg/mL). These findings highlight the dual anticholinesterase activity of these essential oils, particularly M. martiusii, which demonstrated promising selectivity for BuChE and lower cytotoxicity.

Background and aim: This study evaluates the effects of the Sahatsatara formula (STF) on platelet aggregation and examines associations with the Thai birth Dhatu Chao Ruean (bDCR) element classification. Experimental procedure: Forty healthy volunteers (20 male, 20 female) were assigned to Earth, Water, Wind, or Fire groups based on their birth month according to bDCR. Participants received STF (5 capsules three times daily, 30 minutes before meals) for 7 days. Platelet aggregation was measured by light transmission aggregometry at baseline and on days 1 and 7. Platelet cyclooxygenase-1 (COX-1) and cyclooxygenase-2 (COX-2) mRNA and protein levels, as well as plasma thromboxane B2 (TXB2), were quantified using real-time polymerase chain reaction, Western blot, and enzyme-linked immunosorbent assay. Results: On day 1, collagen-induced aggregation decreased significantly in a subgroup exhibiting disaggregation patterns; values returned to baseline by day 8. Individuals classified as Earth element exhibited reduced COX-1 mRNA expression (P < 0.05).

Scientific-grade UAV path planning and multi-objective optimization visualization: Generation of passable grid maps based on high-resolution DEM terrain constraints, including base points, a set of task points (labeled with coordinates, time windows, and package weights), and parameters of homogeneous UAVs (full battery energy, maximum load, safety margin); The terrain modeling section displays a 30m DEM grid map, over-limit areas under a fixed above ground level (AGL) altitude superimposed with the maximum absolute altitude, and obstacle grids marked by morphological dilation (red no-fly zones with a 30m buffer zone); The path feasibility analysis module presents the LoS straight-line traversal sampling process, comparing the terrain elevation profile along the line with the flight altitude, with green indicating feasible direct flight (Euclidean distance) and orange indicating the need to detour; The constrained A* shortest path is calculated and cached on the passable map (the blue curve is the L-DEM distance path), and the KDTree radius candidate pruning illustrates the sparsification of neighboring point pairs; The clustering module uses DBSCAN based on the L-DEM path distance (≤ε) to construct a spatial cluster structure (different colored clusters), with discrete points marked separately; Initial solution generation flowchart: intra-cluster minimum incremental insertion sequence, global optimal insertion of discrete points, Split/Repair decoding to enforce load and return-to-base energy constraints; Evaluator timeline recursive model: flight-wait-service-leave time sequence bars, labeled with arrival time, waiting time, service energy consumption, and lateness delay (soft time window penalty); Energy consumption segmented bar graph: cumulative horizontal flight, hovering, and descent power changes with load; MO-ALNS solver dynamic demonstration: destruction-repair operators acting on the Pareto front solution set, cluster removal, lateness-driven deletion, same-cluster priority insertion and other operation animations, external archives maintain non-dominated solution distribution, and crowding distance clipping maintains diversity; The overall layout is a multi-panel scientific diagram, with terrain and paths in the left column, clustering and initial solutions in the middle column, and the multi-objective optimization process in the right column, with arrow flow lines, molecular formula-style formula annotations for key models (such as $ d_{ij}^{L\text{-}DEM} $, $ E_{\text{remain}} \geq E_{\text{return}} $), color legends, and scale bars, emphasizing the scientific significance of real flight distance modeling, safety verification, and multi-objective trade-off mechanisms.

Generate a vector graphic of a BC mouse.

APPROVED This document outlines the design specifications for a scientific infographic illustrating a magnetic bead purification process for MC plasmid production. The infographic should be designed for rapid comprehension while maintaining scientific accuracy. **LAYOUT STRUCTURE:** The infographic will utilize a horizontal flow diagram consisting of four main stages, clearly demarcated by large, numbered circles (1→2→3→4). Bold arrows will connect the stages to indicate the flow of the process. Each stage will incorporate simple icons and minimal text to facilitate immediate understanding. **STAGE 1: PREPARATION** * **Icon:** Test tube containing magnetic beads. * **Text:** 'Smart Binding System' * **Visual:** Small, colorful dots representing magnetic beads and wavy lines representing DNA. * **Sub-label:** 'Optimized reaction mix' * **Color scheme:** Soft blue background. **STAGE 2: TRIPLE-COMPLEX FORMATION** * **Icon:** Triple helix DNA structure wrapping around a magnetic bead. * **Text:** 'TriDNA Complex Formation' * **Visual:** Three DNA strands (red, blue, and green) forming the triple helix structure. * **Sub-label:** 'Established binding technology' * **Animation:** Arrows indicating 'specific binding'. * **Color scheme:** Light blue background. * **Reference:** Include a small reference citation.

Prior to staining, brine shrimp were washed 2-3 times in a petri dish containing 1 ml of culture water with a salinity of 55, for 5 min each time. A 50 μg aliquot of Mito-Tracker Red was prepared in DMSO to a final concentration of 1 mM as a stock solution. The prepared stock solution was mixed with 2.5 μm microplastics at a concentration of 4 × 105 items/mL at a ratio of 1:50000 to prepare the working solution. The working solution needs to be pre-warmed in a 37°C water bath before use. 1 ml of the pre-warmed working solution was taken and mixed with the brine shrimp in the petri dish and mixed well. The petri dish was wrapped in aluminum foil (protected from light) and incubated in a constant temperature light incubator for 20-30 min. Then, after compressing each brine shrimp, ensuring that the cells are not crushed, observe and photograph under a 40x fluorescence microscope.

The optimization and training strategy system is designed with the core concept of "collaborative optimization and hierarchical control," constructing a full-process control framework of "optimizer-led parameter updates, learning rate scheduling adapted to the training stage, early stopping strategy to suppress overfitting, and regularization to enhance generalization ability." The core functions and collaborative relationships of each module are as follows: Core Optimizer: Adopts the AdamW (Adam with weight decay) adaptive optimizer, responsible for the efficient updating of model parameters, balancing convergence efficiency and generalization ability through adaptive learning rates and independent weight decay mechanisms. Learning Rate Scheduling: Introduces a ReduceLROnPlateau dynamic scheduling strategy, which adaptively adjusts the learning rate based on changes in validation loss, achieving "rapid convergence in the early stage and fine-grained optimization in the later stage." Early Stopping Strategy: Monitors validation loss in real-time to prevent the model from overfitting the training data, improving training efficiency and ensuring generalization ability. Dual Regularization: Combines Dropout layers and weight decay to suppress overfitting from both the feature and parameter levels, complementarily enhancing the generalization effect. This framework is deeply adapted to the training characteristics of deep fully connected neural networks and the task requirements of stress sequence prediction. Through the hierarchical control and synergistic effects of each module, it effectively solves the core contradictions in the training process.

The researchers incorporated several fillers, including titanium dioxide, cellulose, sodium monofluorophosphate, hydroxyapatite, and silver nanoparticles. Among these, sodium monofluorophosphate was employed to inhibit demineralization by releasing fluoride, which suppressed the growth of *Streptococcus mutans*. Ramazan Rajabnia et al. demonstrated in 2016 that chitosan nanoparticles loaded with PAFs exhibited antibacterial activity. A direct contact test assessed the antibacterial properties of resin sealants with the addition of 0, 1, 2, 3, 4, and 5 weight percent chitosan. Increasing the chitosan concentration from 2 to 5 weight percent enhanced the antibacterial capabilities of the sealants. In a 2019 study by Dara Lakshmi Swetha et al., sealants containing a combination of 1% ZnO and CaF2 nanoparticles showed improved antibacterial efficacy.

This study investigates the influence of reaction temperature on the crystal structure, optical band gap, dielectric properties, electrocatalytic activity, and electrochemical behavior of ZnS nanoparticles synthesized via a hydrothermal method at 120, 140, and 160 degrees Celsius. The optical band gaps of samples Z1, Z2, and Z3, synthesized at 120 °C, 140 °C, and 160 °C, respectively, were determined to be direct, with values of 4.5, 4.36, and 4.01 eV. Dielectric characterization revealed a significant frequency-dependent dispersion in the dielectric constant, AC conductivity, and tangent loss. The ZnS nanoparticles synthesized at 160 °C exhibited enhanced optical and dielectric properties based on the overall structural, optical, and dielectric analysis. This material shows promise as a potential candidate for various applications in electronics and optoelectronics. The electrocatalytic and electrochemical properties were evaluated using cyclic voltammetry (CV) at a scan rate of 20 mV/s and galvanostatic charge-discharge (GCD) at a current density of 1 A/g.

Synthesis and characterization of a novel nanocomposite material for industrial applications in the conversion of waste cooking oil to biodiesel.

Schematic diagram illustrating the cloning of a codon-optimized sequence, synthesized based on the expressing species, into an expression vector.

Generate an illustration with a white background depicting the fabrication of a biocoated vascular graft. The schematic should clearly illustrate two main stages: 1) Preparation of the Graft Base with PVA Coating: Begin with the fabrication of PCL nanofiber grafts via electrospinning. Show the preparation of a 1% PVA solution. Illustrate the application of the PVA coating onto the PCL grafts, followed by a drying step for the PVA-coated grafts. 2) Integration of Sulfated Polysaccharides: Depict the selection and preparation of a sulfated polysaccharide solution (e.g., heparin). Illustrate two alternative methods for functionalizing the PVA-coated grafts with the sulfated polysaccharide: Dip-Coating (Physical Adsorption) and Covalent Grafting (e.g., using a cross-linking agent or an adhesive moiety like dopamine). Conclude with a final drying/curing step and sterilization. Ensure the schematic is clear, easy to follow, and visually represents the sequential steps involved in creating the biocoated vascular graft.

APPROVED This section focuses on experimental verification and model evolution. The core aspects include: 1. Evolutionary Strategy Fusion: Integrating the Analytic Hierarchy Process (AHP) with adversarial distillation heterogeneous techniques to achieve self-evolution of the anti-interference model. 2. Model Library Generation: Applying graph-driven pruning and multi-metric channel compression to generate various intelligent anti-interference decision model variants. 3. Layered Testing: Conducting dual verification of basic performance and adversarial robustness on embedded devices and GPU server platforms. The final content includes the following: 5. [Bottom Illustration/Formula]: Depicting a **"model evolution funnel"**: Original large model $\xrightarrow{Pruning/Distillation}$ Evolved Models V1/V2/V3 $\xrightarrow{Testing}$ Optimal Strategy.

Prior to staining, brine shrimp were washed 2-3 times in a petri dish containing 5 ml of culture water with a salinity of 55, for 5 min each time. A 1 mM stock solution of 50 μg Mito-Tracker Red was prepared using DMSO. The prepared stock solution was mixed at a ratio of 1:50000 to prepare a 20 nM working solution. The working solution was pre-warmed in a 37°C water bath before use. 5 ml of the pre-warmed working solution was mixed with the brine shrimp in the petri dish and mixed well. The petri dish was wrapped in tin foil (protected from light) and incubated in a constant temperature light incubator for 20-30 min. Then, each brine shrimp was squashed, ensuring that the cells were not crushed, and observed and photographed under a 40x fluorescence microscope.

Prior to staining, brine shrimp were washed 2-3 times in a petri dish containing 5 ml of culture water with a salinity of 55, for 5 min each time. Mito-Tracker Green was prepared to a final concentration of 1 mM using DMSO, which served as the stock solution. The prepared stock solution was mixed at a ratio of 1:50000 to prepare a 20 nM working solution. The working solution was pre-warmed in a 37°C water bath before use. 5 ml of the pre-warmed working solution was taken and mixed with the brine shrimp in the petri dish and mixed well. The petri dish was wrapped in tin foil (protected from light) and incubated in a constant temperature light incubator for 20-30 min. Then, each brine shrimp was squashed, ensuring that the cells were not crushed, and observed and photographed under a 40x fluorescence microscope.

Illustration depicting the Hill muscle model, comprising three elements: the Contractile Component (CC), the Parallel Elastic Component (PEC), and the Series Elastic Component (SEC). The diagram should incorporate the Titin protein structure as an integral part of the SEC. Each component should be clearly labeled, and their interactions within the muscle fiber structure should be illustrated. The muscle fiber should be represented in a simplified, anatomical style, emphasizing the distinct components, with particular attention to the elastic elements and the contractile unit.

Scientific Transition from INRAE to CEPR (2026–): Competencies and Immediate Research Focus Core Competencies and Assets: * Mechanistic Mucosal Immunology: Expertise in innate immune regulation at mucosal barriers, focusing on inflammation, tolerance, and tissue resilience. * Host-Microbe-Pathogen Interactions: Emphasis on mucus-driven constraints affecting diffusion, immune signaling, and pathogen persistence. * Integrated Infection Biology Pipelines: Ready-to-deploy experimental toolset for comprehensive analysis of immune readouts and tissue outcomes, including immunopathology and cytokine/antiviral programs. * Microbiota-Immunometabolism Integration: Linking microbial communities and metabolites to immune programs and phenotypes. * Translational Experience: Proven ability to move between in vivo, ex vivo, and advanced epithelial systems, translating findings into actionable hypotheses. * Project Building and Translation: Coordination of multidisciplinary programs and academic-industry interfaces. Structuring funnel-like R&D workflows.

Drawing a network structure diagram of the algorithm. The following is a description of the algorithm: To improve the cooperative control ability of the aerial-underwater amphibious drone in the air-water heterogeneous environment, this paper introduces a structured masked attention mechanism (SMA) based on the MADDPG framework to construct a central Critic network, so that information can be interactively targeted. The input of each agent's central Q-value function not only includes its own state and action information, but also aggregates the information of other drones jointly according to the attention weight, specifically expressed as QSMAi(O,A)=fei(oi,ai),Ci, where (O,A) represents the set of observations and actions of all drones, and ei(oi,ai) is a nonlinear mapping that encodes its own observation oi and action ai into a feature vector. Each type of drone only allocates attention to the actual associated objects, allocates attention weights only to related objects, and improves information filtering ability through softmax normalization. The attention distribution of underwater followers is only for formation leaders, and the attention distribution of aerial followers is for formation leaders and other aerial followers. The formation leader aggregates all follower information. The information aggregation of each agent i can be uniformly modeled as Ci=j∈FiαijReLU(vj), where ReLU(⋅) is a nonlinear activation function and vj is the value vector of drone j. The central Critic defines the attention object set Fi (mask operation) for each drone i according to its task division. The weight aij is adaptively allocated according to the softmax mechanism according to the actual system scenario. The calculation method is αij=exp(qi)⊤kj/dkl∈Fiexp(qi)⊤kl/dk, where qi and kj are the corresponding query and key vectors. The attention score of non-Fi members is set to -∞ through a mask to achieve structured filtering and no weight is allocated on it. In this way, the central Critic effectively aggregates the key information in Fi in the attention module, and at the same time ensures that various types of aircraft can accurately focus on the effective data under the division of labor link during observation and decision-making. In training, a batch experience replay mechanism is used to continuously store (o, a, ri, o') obtained during the environmental interaction process into the experience pool D, and parameters are updated regularly by sampling from D. The value network parameter ωi is optimized through the temporal difference loss. The loss function is Loss(ωi)=E(o,a,ri,o')∼DQSMAi(O,A;ωi)−y2, where the target value y is calculated by the target value network QSMAi and the target strategy network μθj'. y=ri+γQSMAi'(O',A';ωi'), A'=(a1',...,aN'), aj'=μθj'(oj'), where ri is the reward value of drone i, γ is the discount factor, and ωi' and θj' are the target value network and target strategy network parameters, respectively. The strategy network μθi is updated from the main network parameters in a delayed manner through a soft update mechanism. The update goal is to maximize the long-term cumulative reward, and its gradient form is ∇θiJ(μθi)=E(O,A)∼D∇θiμθi(oi)∇aiQSMAi(O,A;ωi).

Drawing a network structure diagram of the algorithm. The following is a description of the algorithm: To improve the coordinated control ability of the aerial-underwater amphibious drone in the air-water heterogeneous environment, this paper introduces a structured masked attention mechanism (SMA) based on the MADDPG framework to construct a central Critic network, enabling targeted information interaction. The input of each agent's central Q-value function includes not only its own state and action information but also the information of other drones, which is jointly aggregated according to attention weights. Specifically, it is expressed as QSMAi(O,A)=fei(oi,ai),Ci, where (O,A) represents the set of observations and actions of all drones, and ei(oi,ai) is a nonlinear mapping that encodes its own observation oi and action ai into a feature vector. Each type of drone only allocates attention to the actual associated objects, allocates attention weights only to relevant objects, and improves information filtering ability through softmax normalization. The attention distribution of underwater followers is only for formation leaders, the attention distribution of aerial followers is for formation leaders and other aerial followers, and the formation leader aggregates all follower information. The information aggregation of each agent i can be uniformly modeled as Ci=j∈FiαijReLU(vj), where ReLU(⋅) is a nonlinear activation function and vj is the value vector of drone j. The central Critic defines the set of attention objects Fi (mask operation) for each drone i according to its task division. The weight aij is adaptively allocated according to the actual system scenario according to the softmax mechanism. Its calculation method is αij=exp(qi)⊤kj/dkl∈Fiexp(qi)⊤kl/dk, where qi and kj are the corresponding query and key vectors. The attention score of non-Fi members is set to -∞ through a mask to achieve structured filtering and no weight is allocated on it. In this way, the central Critic effectively aggregates the key information in Fi in the attention module, while ensuring that various types of aircraft can accurately focus on the effective data under the division of labor link during observation and decision-making. In training, a batch experience replay mechanism is used to continuously store (o, a, ri, o') obtained during environmental interaction into the experience pool D, and parameters are updated regularly by sampling from D. The value network parameter ωi is optimized through the temporal difference loss. The loss function is Loss(ωi)=E(o,a,ri,o')∼DQSMAi(O,A;ωi)−y2, where the target value y is calculated by the target value network QSMAi' and the target strategy network μθj'. y=ri+γQSMAi'(O',A';ωi'), A'=(a1',...,aN'), aj'=μθj'(oj'), where ri is the reward value of drone i, γ is the discount factor, and ωi' and θj' are the target value network and target strategy network parameters, respectively. The strategy network μθi is updated from the main network parameters in a delayed manner through a soft update mechanism. The update goal is to maximize the long-term cumulative reward, and its gradient form is ∇θiJ(μθi)=E(O,A)∼D∇θiμθi(oi)∇aiQSMAi(O,A;ωi).