A method for designing a semiconductor selective intervention smart screening agent

Abstract

This invention belongs to the field of semiconductor material design, specifically relating to a method for designing a semiconductor selective intervention smart screening agent. The method includes: constructing an intervention matrix using optical fibers to establish a multi-dimensional channel that integrates detection signal transmission, high-energy beam guidance, and screening agent delivery; designing a photoresponsive smart screening agent and pre-setting its phase transition and chemical reaction logic based on defect characteristics; real-time acquisition of micro / nano morphology and electric field data of the chip surface to map a dynamic defect probability heatmap; and extracting heatmap features based on a deep learning model to determine the corresponding screening agent flow rate, beam energy, and irradiation time in real time. This invention, through an integrated sensing and intervention architecture, achieves in-situ real-time intervention during the manufacturing process, breaking through physical limits of precision, enhancing technical robustness, and improving semiconductor device yield.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor material design, specifically relating to a method for designing intelligent screening agents for selective intervention in semiconductors. Background Technology

[0002] With the rapid evolution of power semiconductor technology, high-power-density devices such as IGBTs are increasingly used in new energy vehicles, industrial control, and power systems, placing extremely high demands on the reliability and consistency of their manufacturing processes. In the semiconductor precision manufacturing chain, timely screening and targeted intervention for lattice defects, hot spots, or microscopic damage on the chip surface are key technical steps to ensure final product yield and curb the spread of failures. Currently, the semiconductor quality control field mainly integrates high-resolution optical characterization and physical sensing methods to strive for rapid identification and marking of potential fault points at the micro-nano scale, supporting subsequent process optimization and quality rejection.

[0003] Selective intervention and intelligent screening technology for semiconductor defects has become a core area for improving flexible manufacturing capabilities. This technology aims to utilize functionalized screening media to perform physicochemical processes such as local passivation, physical encapsulation, or chemical repair on defect areas at the microscopic scale, controlling manufacturing risks at their inception. In ideal intelligent manufacturing scenarios, the distribution, activity excitation, and projection dosage of the screening agent must be highly matched with the real-time perceived defect map. This requires the system to possess extremely high spatiotemporal resolution and robustness under complex physical environments.

[0004] Existing technologies exhibit a spatiotemporal separation between the detection and intervention processes, leading to lag in semiconductor device processing. Traditional off-site repair methods are highly susceptible to introducing secondary contamination during wafer transport and may induce stress propagation of defects such as microcracks in uncontrolled environments. Conventional mechanical probes or micro-spraying equipment struggle to achieve precise alignment of micron-level trench structures under complex conditions such as high temperatures and plasma. Furthermore, the lack of dynamically controlled design based on the physicochemical properties of the screening medium easily results in localized over-etching or poor coverage uniformity. Due to the limitation of waveguide structures to low-power signal transmission, existing equipment cannot provide high-energy physical intervention within milliseconds of defect detection, resulting in insufficient precision and real-time feedback capability for dynamic failure modes such as nonlinear hotspots. Summary of the Invention

[0005] The purpose of this invention is to provide a semiconductor selective intervention intelligent screening agent design method that can solve the problems mentioned in the background art. In the current insulated-gate bipolar transistor manufacturing process, there is a significant spatiotemporal separation between the detection process and the intervention and repair process. This leads to the risk of contamination from uncontrolled environments and the propagation of microcracks induced by thermal stress during the wafer transfer to the repair equipment. Because the utilization of optical waveguide structures in traditional technical approaches is limited to signal transmission and lacks means to perform micron-level precise intervention in complex manufacturing environments, existing solutions struggle to achieve real-time response and intelligent screening of defects, limiting further improvements in semiconductor device yield. This invention aims to resolve the contradiction between high-precision positioning and robustness of intervention in complex environments by constructing an integrated optofluidic architecture that combines sensing and intervention.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for designing semiconductor selective intervention smart screening agents, comprising the following specific steps: Step 1: Construct an optical fiber array intervention matrix. Use multi-core optical fibers or hollow-core photonic crystal fibers to configure intervention probes at key workstations in the insulated gate bipolar transistor production line to establish a multi-dimensional channel that combines detection signal transmission, high-energy intervention beam guidance, and micro-fluid screening agent delivery. Step 2: Design photoresponsive functional microspheres as smart screening agents, fill the internal cavity of the fiber array interference matrix with the screening agent, and pre-determine the phase transition behavior and chemical reaction logic of the screening agent under light excitation based on the physical characteristics of semiconductor surface defects. Step 3: Map a dynamic defect probability heat map. Use the fiber optic sensor integrated in the fiber optic array intervention matrix to collect micro-nano topography data and electric field distribution data of the chip surface to be processed in real time. Generate a probability distribution matrix characterizing the severity of defects through data fusion. Step 4: Calculate the adaptive intervention dose, extract features from the defect probability heatmap based on the deep learning model, and determine the screening agent delivery rate, intervention beam energy, and irradiation duration corresponding to the defect point coordinates in real time; Step 5: Perform in-situ coaxial intervention, control the fiber array intervention matrix to physically address the defect coordinates, and project the high-energy intervention beam and the smart screening agent synchronously to the micron-level defect point through the microlens structure of the fiber end face, so as to complete the physical encapsulation or chemical repair of the defect in the in-situ state.

[0007] Preferably, in step 1, the fiber array intervention matrix is ​​made of a high-temperature resistant and plasma corrosion-resistant material, and integrates a plurality of waveguide units internally. The waveguide units are arranged in a matrix in a direction perpendicular to the chip surface, and each waveguide unit has an independent optical modulation function. The core diameter of the multi-core fiber matches the gate trench size of the insulated gate bipolar transistor. The intervention probe is integrated above the back metallization chamber or bonding stage, and a precision displacement mechanism controls the probe end face to maintain a predetermined distance from the chip surface.

[0008] Preferably, in step 1, the fiber array intervention matrix is ​​further equipped with a fiber Bragg grating sensor. The fiber Bragg grating sensor utilizes the characteristic that the grating pitch changes with environmental parameters to sense temperature fluctuations and mechanical stress changes in the production environment in real time. When the ambient temperature changes, the center wavelength of the reflected light signal shifts. The system calculates the product of this shift and a preset proportionality coefficient to obtain the current real-time temperature value, and feeds it back to the central control unit to correct the energy parameters of the intervention beam.

[0009] Preferably, in step 2, the smart screening agent consists of a polymer shell with photosensitive properties and a functionalized core. The functionalized core contains chemical agents for local passivation or nanoparticles for enhancing thermal conductivity. The screening agent is chemically inert under normal conditions. When the high-energy interference beam irradiates the surface of the screening agent, the polymer shell undergoes chain-breaking or shape-memory recovery under the action of photon energy, releasing the internal core material. The core material undergoes physical adsorption or chemical bonding with lattice defects in the defect region, inhibiting further diffusion of defects.

[0010] Preferably, in step 2, the delivery of the smart screening agent employs a pressure-driven and photo-induced deformation coordinated control mode. A photosensitive lubricating layer of a specific thickness is coated on the inner wall of the hollow photonic crystal fiber. By adjusting the intensity of the auxiliary control light, the frictional resistance of the inner wall is changed, thereby achieving fine-tuning of the screening agent delivery rate. The screening agent maintains a uniform particle distribution during delivery, avoiding aggregation and blockage within the fiber optic channel.

[0011] Preferably, in step 3, the acquisition of the micro / nano topography data is achieved using the principle of fiber optic interference. The fiber optic sensor emits a reference beam and a probe beam. After the probe beam is reflected off the chip surface, it undergoes coherent interference with the reference beam. By detecting the shift of the interference fringes and performing phase unfolding calculations, the height undulation information of the chip surface is obtained. The acquisition of the electric field distribution data is based on the electro-optic effect. The electro-optic crystal at the end of the fiber senses the surface electric field intensity, converting the change in electric field into a change in the polarization state of the light wave, thereby indirectly measuring the non-uniform distribution of the surface voltage.

[0012] Preferably, in step 3, the generated dynamic defect probability heatmap uses a three-dimensional coordinate mapping method to establish a one-to-one correspondence between physical space coordinates and defect probability values. The heatmap generation process includes normalizing the original sensor data, removing environmental noise interference, and using interpolation algorithms to estimate the values ​​of non-sampling points, constructing a continuous defect density surface. When the defect probability at a certain coordinate point exceeds a preset threshold, the system marks it as a high-risk intervention target.

[0013] Preferably, in step 4, the deep learning model employs a convolutional neural network architecture. This architecture includes multiple convolutional layers for extracting the geometric features of the defects, and multiple pooling layers for reducing data dimensionality while maintaining feature invariance. The extracted feature vectors are input to a fully connected layer, compared with a pre-defined expert knowledge base, and output the optimal combination of intervention parameters for the defect type. This combination of intervention parameters includes at least the time-series distributions of laser pulse width, peak power, and screening agent spraying pressure.

[0014] Preferably, in step 4, a feedback compensation mechanism is introduced into the logic for calculating the adaptive intervention dose. The system monitors the temperature rise rate of the defective area in real time during the intervention process. If the temperature rise rate exceeds the preset safety limit, the duty cycle of the intervention beam is automatically reduced, and the flow rate of the screening agent is increased to utilize its latent heat of phase change for local cooling, preventing thermal damage to the surrounding good product area.

[0015] Preferably, in step 5, the in-situ coaxial intervention achieves co-optical transmission of the high-power intervention laser and the low-power detection light through a fiber optic coupler. The fiber optic coupler combines beams of different wavelengths into the same fiber, ensuring that the coordinate center of the detection point and the energy center of the intervention point are highly coincident in physical space. The microlens structure has an aspherical design, which can focus the beam into a tiny spot smaller than a predetermined diameter to match the microstructure size of the insulated gate bipolar transistor.

[0016] Preferably, in step 5, during the projection of the smart screening agent, the screening agent is pre-excited using the evanescent wave field of the optical fiber. When the screening agent flows through the tapered region at the end of the optical fiber, part of the light field transmitted within the fiber overflows to the outside of the fiber core, exchanging energy with the screening agent. This causes the screening agent molecules to enter a metastable state, triggering a repair reaction with an extremely low energy threshold at the moment of contact with the defect surface, thus improving the timeliness of the intervention.

[0017] Furthermore, the semiconductor selective intervention smart screening agent design method also includes a real-time evaluation step of the intervention results. After intervention, the system immediately re-collects the morphology and electrical signals of the area using an in-situ fiber optic sensor, comparing the data differences before and after intervention. If the defect characteristics do not disappear or decrease to a preset safety range, a secondary intervention process is automatically initiated, and the intervention strategy is adjusted based on the current evaluation results until the predetermined quality standard is met.

[0018] The semiconductor selective intervention intelligent screening agent design method, during operation, simultaneously establishes an intelligent screening database containing multiple intervention records. This database records the defect distribution patterns of each chip, the corresponding intervention parameters, and the final electrical performance test results. The system utilizes a long short-term memory network to learn from historical records, predict the potential failure trends of the current batch of chips, and optimize the screening agent ratio for subsequent chips in advance.

[0019] The method is applied to automated production lines, achieving seamless integration with upstream epitaxial growth processes and downstream thinning and metallization processes through deep integration with industrial control systems. Throughout the intervention process, the chip is kept in a vacuum or specific inert gas protective environment, eliminating operational risks associated with equipment switching and manual intervention.

[0020] For specific structures of insulated gate bipolar transistors, such as the bottom of the gate trench or the terminal protection ring region, the probe angle of the fiber array intervention matrix is ​​adjustable, and oblique projection is achieved through a multi-axis motion platform to ensure that the screening agent can penetrate deeply into the trench with a depth-to-diameter ratio, achieving all-round selective intervention.

[0021] Compared with the prior art, the present invention has the following beneficial effects: 1. In-situ, real-time intervention in the manufacturing process of insulated gate bipolar transistors (IGBTs) has been achieved. Through the design of a fiber optic array intervention matrix, the traditional spatiotemporal limitations of detection and repair have been broken, eliminating temporal redundancy in material transport, preventing stress propagation and secondary contamination of lattice defects in uncontrolled environments, and improving the manufacturing yield of semiconductor devices.

[0022] 2. Possesses intervention precision that surpasses physical limits. Utilizing the minute size characteristics of optical waveguide structures and the focusing capability of microlenses, the intervention range can be precisely controlled within the micrometer scale. This enables point-to-point precise processing of micro-trenches, gate structures, and termination rings in insulated-gate bipolar transistors, avoiding physical or thermal damage to the surrounding good-quality areas caused by traditional mechanical intervention methods.

[0023] 3. The screening process has been made intelligent and adaptive. By combining deep learning models with dynamic defect probability heat maps, the equipment can calculate and generate personalized screening agent delivery schemes and energy intervention sequences in real time based on the specific morphology and electrical characteristics of each defect point. This changes the rigid mode of traditional equipment mechanically executing preset programs and improves the level of flexible manufacturing.

[0024] 4. Enhanced technical robustness in complex production environments. The non-contact fiber optic intervention method enables stable operation under harsh conditions such as high temperature, high pressure, or plasma. Furthermore, the real-time environmental sensing achieved through fiber Bragg gratings ensures that intervention parameters can be dynamically corrected according to environmental fluctuations, guaranteeing the consistency and reliability of the processing results.

[0025] 5. A novel photoresponsive smart screening agent was introduced. Through photofluidic technology, the phase transition and activity of the screening agent were precisely controlled, achieving millisecond-level response and precise boundary control of the intervention dose, providing a novel physicochemical approach for the process self-healing of semiconductor defects. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the integrated optical fluid control architecture for sensing and intervention in this invention; Figure 3 This is a flowchart illustrating the main stages of the multi-dimensional sensing data fusion and dynamic defect probability heatmap construction in this invention. Figure 4 This is a flowchart of the adaptive intervention dosage decision logic based on a deep learning model in this invention. Figure 5 This is a flowchart outlining the main stages of in-situ coaxial intervention execution and precise projection of intelligent screening agents in this invention. Figure 6 This is a schematic diagram illustrating the multi-level interactive relationship between real-time evaluation of intervention results and intelligent database feedback optimization in this invention. Detailed Implementation

[0027] Example 1: Please refer to the appendix Figure 1 To be continued Figure 6 In this embodiment, a semiconductor selective intervention smart screening agent design method is implemented according to the following steps.

[0028] Step 1 involves constructing a fiber optic array intervention matrix. In practice, multi-core optical fibers or hollow-core photonic crystal fibers are used to configure intervention probes at key stations on the insulated-gate bipolar transistor (IGBT) production line. These key stations include the operating station above the back metallization chamber or the bonding stage. This step establishes a multi-dimensional channel that integrates detection signal transmission, high-energy intervention beam guidance, and microfluidic screening agent delivery. The fiber optic array intervention matrix is ​​made of a high-temperature resistant and plasma-corrosion resistant material, specifically a doped quartz material or ceramic matrix composite material with high thermal stability. It integrates multiple waveguide units arranged in a matrix pattern perpendicular to the chip surface.

[0029] The waveguide units are distributed in a grid pattern with equal spacing on the horizontal projection plane. Each waveguide unit has an independent optical modulation function, enabling control of the on / off state of optical signals along a specific path via a controlled optical switch. The core diameter of the multi-core optical fiber is set between 5 and 15 micrometers, a size that matches the gate trench size of an insulated gate bipolar transistor, ensuring that the intervention action can be precisely applied inside the trench. The intervention probe is driven by a high-precision 5-axis linkage micro-displacement mechanism, controlling the probe end face to maintain a predetermined distance of 10 to 50 micrometers between the probe end face and the chip surface to prevent physical collisions and ensure optical coupling efficiency.

[0030] In step 1, a fiber Bragg grating sensor is also configured inside the fiber array intervention matrix. The fiber Bragg grating sensor utilizes the grating pitch generated by its internal periodically varying refractive index distribution, and its ability to sense temperature fluctuations and mechanical stress changes in the production environment in real time, adapting to changes in environmental parameters. When the ambient temperature changes, the thermal contraction effect of the grating causes a shift in the center wavelength of the reflected light signal. The system calculates this wavelength shift by capturing the peak value change of the reflected spectrum in real time. The specific calculation logic is: the current real-time temperature value equals the reference temperature plus the quotient of the wavelength shift and a preset temperature sensitivity coefficient. The system feeds this temperature value back to the central control unit in real time to dynamically correct the energy parameters of the intervention beam in subsequent steps, compensating for light field intensity fluctuations caused by temperature drift.

[0031] Step 2 involves designing photoresponsive functional microspheres as smart screening agents. These smart screening agents are pre-filled into the internal cavities of the fiber optic array interference matrix, typically in a specific carrier fluid suspension state. Based on the physical characteristics of semiconductor surface defects, the phase transition behavior and chemical reaction logic of the screening agent under specific light excitation are pre-defined. The smart screening agent consists of a polymer shell with photosensitive properties and a functionalized core. The polymer shell uses a photosensitive polymer containing azobenzene functional groups or an o-nitrobenzyl structure. Such materials undergo photoisomerization or photolysis upon receiving ultraviolet or deep ultraviolet light of specific wavelengths.

[0032] The functionalized core contains chemical agents for local passivation, such as polyimide precursors or special silane coupling agents, or nanoparticles for enhancing thermal conductivity, such as gold nanoparticles or boron nitride nanotubes. The screening agent is chemically inert under normal conditions; due to its outer shell, the core material does not react with the external environment. When the high-energy interference beam irradiates the surface of the screening agent, the polymer shell undergoes chain-breaking reactions under specific photon energies, or exhibits shape memory recovery effects, leading to shell rupture or changes in permeability, releasing the internal core material. The core material undergoes physical adsorption with lattice defects in the defect region, or achieves chemical bonding through covalent bonding, forming a dense passivation layer or thermally conductive bridging structure at the defect location, inhibiting further diffusion or thermal failure of defects during subsequent high-voltage testing.

[0033] In step 2, the delivery of the smart screening agent employs a pressure-driven and photo-induced deformation coordinated control mode. A photosensitive lubricating layer with a thickness of 200 to 500 nanometers is coated on the inner wall of the hollow-core photonic crystal fiber. By adjusting the intensity of the auxiliary control light, the molecules in the lubricating layer undergo directional alignment changes, altering the frictional resistance of the fiber's inner wall to the fluid. The specific delivery rate control logic is as follows: the current delivery flow rate setpoint increases monotonically with the increase of the auxiliary control light intensity and is proportional to the pressure difference between the two ends of the delivery pipeline. During delivery, the screening agent maintains a uniform particle distribution by adding an appropriate amount of surfactant to the carrier fluid, preventing aggregation and blockage within the micron-sized fiber channel.

[0034] Step 3 involves mapping a dynamic defect probability heatmap.

[0035] Specifically, step 3, the generation of the defect probability heatmap, involves a refined data fusion and processing procedure. First, the original micro / nano topography data and electric field distribution data are spatially aligned and normalized. The micro / nano topography data is represented in the form of a three-dimensional point cloud, denoted as... ,in For height value, is a planar coordinate system. The electric field distribution data is represented in two-dimensional matrix form, denoted as . ,in

[0036] This refers to the normalized surface voltage or potential difference. The normalization process uses the minimum-maximum method, mapping both height and voltage data to... Interval. Data fusion is achieved through a weighted fusion function, which is designed to comprehensively reflect the combined effects of physical damage and electrical anomalies. The fused original defect characteristic values. The calculation is as follows: ; in, and These are the maximum and minimum values ​​of the current chip surface height, respectively; and

[0037] These are the maximum and minimum values ​​of the current chip surface voltage, respectively; and These are preset weighting coefficients, summing to 1. Their specific values ​​are determined based on process experience in assessing the sensitivity of morphological and electric field defects. For example, for IGBT devices, abnormal electric field distribution has a greater impact on failure; therefore, a weighting coefficient can be set... ; and This is a non-linear adjustment index used to enhance the contrast of eigenvalues, and its value is typically between 1 and 2.

[0038] Subsequently, the original eigenvalue matrix of the fused defects was analyzed. Spatial interpolation is performed. As a preferred implementation, a bicubic interpolation algorithm is used, utilizing the known values ​​of the 16 surrounding neighboring grid points, to estimate the non-sampling points through cubic polynomial fitting. The eigenvalues ​​are used to construct a continuous defect density surface with consistent resolution. Finally, the curved surface The range is mapped to The defect probability is used to form the final defect probability heatmap. The alarm threshold It is not a fixed value, but rather dynamically set based on the yield target of the current chip process node. Specifically, ,in These represent the mean and standard deviation of the current chip defect probability distribution, respectively. For safety, a factor of 3 to 5 is typically used, so that the system only marks points that deviate significantly from the mean and are statistically abnormal as high-risk intervention targets.

[0039] The system utilizes an optical fiber sensor integrated into the optical fiber array interference matrix to acquire real-time micro / nano topographic data and electric field distribution data of the chip surface to be processed. The acquisition of the micro / nano topographic data is achieved using the principle of optical fiber interference. A coupler inside the optical fiber sensor splits the coherent light emitted from the light source into a reference beam and a probe beam. The probe beam passes through the end of the optical fiber and strikes the chip surface, reflects off the chip surface, returns to the optical fiber, and coherently interferes with the reference beam. The system records the number of shifts in the interference fringes using a high-speed photodetector. The specific logic for calculating the height fluctuation is as follows: the change in chip surface height equals the product of the number of shifts in the interference fringes and the laser wavelength, divided by twice the refractive index of the medium. The system converts the discrete interference phase information into continuous surface topographic fluctuation information through phase unrolling operations.

[0040] The acquisition of the electric field distribution data is based on the electro-optic effect, utilizing an electro-optic crystal at the end of the optical fiber to sense the surface electric field intensity. When there is local voltage anomaly caused by charge accumulation or leakage current on the chip surface, the refractive index of the electro-optic crystal changes anisotropy, causing the polarization state of the probe light passing through the crystal to rotate. The system indirectly measures the non-uniform distribution of the surface voltage by detecting changes in the polarization extinction ratio. The generated dynamic defect probability heatmap uses a three-dimensional coordinate mapping method to map the horizontal and vertical coordinates of the chip to the defect probability values. The heatmap generation process includes normalizing the raw sensor data, removing random fluctuations caused by background noise, and using a bicubic interpolation algorithm to estimate the values ​​of non-sampling areas. When the defect probability value at a certain coordinate point exceeds a preset alarm threshold, the system automatically marks the area as a high-risk intervention target and locks its geometric center coordinates as the intervention point.

[0041] Step 4 involves calculating the adaptive intervention dose. This step utilizes a deep learning model to extract multi-scale features from the defect probability heatmap.

[0042] The deep learning model used in step 4 is a multi-branch fusion convolutional neural network. The input to this network is the defect probability heatmap generated in step 3. The network contains three-channel feature maps of the original morphology and electric field data, along with their corresponding original morphology. The network includes two feature extraction branches: Branch A is a standard convolutional module used to extract features from the heatmap. Extracting macroscopic defect morphology features; Branch B is a module containing dilated convolutional layers, used to extract macroscopic defect morphology features from the original high-resolution electric field distribution data. The network captures fine-grained anomaly features of the electric field gradient. The feature maps extracted from the two branches are concatenated along the channel dimension to form a fused feature map. This fused feature map is then processed through three convolutional layers and two fully connected layers. The final output layer of the network contains three neurons, each corresponding to a predicted filter agent dispensing rate. (Unit: nL / min), Interference beam energy

[0043] (Unit: mJ) and duration of irradiation (Unit: ms). The training process of the deep learning model adopts a supervised learning method. The training dataset is constructed as follows: a large number of historical chip defect heatmaps are collected as input samples, and intervention parameter combinations that have been verified by experts are recorded. As labels. Model training employs a multi-task learning loss function. : ; in, This is the mean squared error loss for traffic flow forecasting. and

[0044] These are the mean squared error losses of energy and time, respectively, and are defined in the same way. For the number of training samples in a batch,

[0045] The weights for each task's loss are used to balance the contribution of parameters with different dimensions to the total loss. During training, the Adam optimizer is used to iteratively update the network weights through backpropagation until the loss function value on the validation set converges to a stable level.

[0046] The deep learning model employs a deep convolutional neural network architecture, comprising 5 convolutional layers, 3 pooling layers, and 2 fully connected layers. The convolutional layers utilize kernels of different sizes to extract geometric edge features, texture distribution features, and contrast features of defects. The pooling layers employ max pooling to reduce the dimensionality of data processing while maintaining the spatial invariance of features. The extracted feature vectors are input to the fully connected layers, undergo nonlinear transformation, and are then compared with a pre-defined expert knowledge base for high-dimensional feature analysis.

[0047] Furthermore, the feedback compensation mechanism in step 4 is implemented using a dynamic adjustment algorithm based on the proportional-integral-derivative (PID) control principle. Real-time temperature rise rate. The temperature of the defect point within two consecutive sampling periods The difference was calculated to obtain: ,in

[0048] This is the sampling period for the temperature sensor. The preset safe temperature rise limit. Determined based on the thermal shock tolerance threshold of the chip material (such as silicon). When At that time, the system calculates the compensation amount for the intervention parameters. This includes the adjustment amount of the beam duty cycle. With traffic compensation Calculate according to the following formula: ; in,

[0049] This represents the deviation of the current rate of temperature rise. These are the proportional, integral, and derivative gain coefficients of the PID controller, whose values ​​are obtained through system modeling and experimental calibration. For time; To map the duty cycle change to a proportional coefficient for flow rate adjustment, its physical meaning is the amount of heat that can be compensated by increasing the filter agent flow rate per unit decrease in duty cycle. The specific value is related to the specific heat capacity and latent heat of phase change of the filter agent. Adjusted duty cycle Adjusted flow rate This compensation logic is executed cyclically at millisecond intervals during the intervention process to ensure that the temperature at the defect point is stably controlled within a safe range.

[0050] The specific intervention parameter calculation process is as follows: Based on the input feature vector, the system retrieves the intervention template with the highest matching degree from the expert knowledge base and outputs the optimal combination of intervention parameters for the current defect type. The intervention parameter combination includes at least the laser pulse width (ranging from 10 nanoseconds to 500 nanoseconds), laser peak power (ranging from 1 megawatt to 50 megawatts), and the time series distribution of the screening agent spraying pressure. A feedback compensation mechanism is introduced into the calculation logic, and the system monitors the temperature rise rate of the defect area in real time during the intervention process. If the monitored temperature rise rate exceeds the safe limit of 100 degrees Celsius per second, the calculation module automatically executes derating logic: the current laser duty cycle is adjusted to 0.8 times the original set value, the screening agent flow compensation value is increased, and the latent heat generated by the phase change of the screening agent is used to absorb excess heat energy, preventing thermal stress damage to the surrounding good product area.

[0051] Step 5 involves performing in-situ coaxial intervention. The fiber optic array intervention matrix is ​​controlled to physically address the defect coordinates determined in step 3. The high-energy intervention beam and the smart screening agent are synchronously projected onto the micron-level defect point through a microlens structure on the fiber end face. This in-situ coaxial intervention is achieved using a special fiber optic coupler with wavelength division multiplexing (WDM) capability. This coupler can combine the low-power infrared light used for detection and the high-power ultraviolet laser used for intervention into the same fiber, ensuring that the coordinate center of the detection beam and the energy center of the intervention beam have a spatial overlap deviation of less than 1 micrometer.

[0052] The microlens structure features an aspherical design, enabling it to focus the emitted divergent beam into a tiny spot with a diameter of less than 5 micrometers, matching the micro-groove structure of the insulated-gate bipolar transistor. During the projection of the smart screening agent, the evanescent wave field of the optical fiber pre-excites the agent. As the screening agent flows through the tapered region at the end of the optical fiber, a portion of the high-density light field transmitted within the fiber overflows to the cladding interface outside the fiber core. When the screening agent molecules pass through this region, they exchange energy with the evanescent wave, causing electrons to transition from the ground state to an excited state, thus putting the screening agent into a metastable state. This pre-excited state allows the screening agent to trigger a repair chemical reaction with an extremely low energy threshold upon contact with the chip defect surface, improving the timeliness and depth of the intervention.

[0053] Furthermore, the method in this embodiment also includes real-time evaluation of the intervention results. Within 100 milliseconds after step 5 is completed, the system automatically starts the evaluation program. Micro-nano morphology data and electrical performance data of the intervention area are collected again using an in-situ fiber optic sensor. The system compares the data differences before and after the intervention and calculates the defect elimination rate. The defect elimination rate is equal to the difference between the defect feature value before and after the intervention, divided by the defect feature value before the intervention. If the calculated defect elimination rate is less than 95%, or the residual defect features are still higher than the preset safety threshold, the system automatically adds the coordinates back to the intervention sequence, starts the secondary intervention process, and automatically increases the laser energy density based on the current residual feature feedback until the predetermined quality judgment standard is reached.

[0054] Throughout the process, the system simultaneously establishes a smart screening database containing multiple intervention records. This database records the defect type distribution, corresponding intervention parameter sequences, historical environmental temperature and humidity curves, and final electrical breakdown characteristic test results for each chip. The system utilizes a Long Short-Term Memory (LSTM) network to perform correlation learning on the historical sequence data, analyzes the defect evolution patterns under different process batches, enables early prediction of potential failure trends in the current batch of chips, and dynamically optimizes the chemical ratio of screening agents for subsequent chips. This method is applied to automated production lines, achieving seamless integration with upstream epitaxial growth and photolithography processes, as well as downstream thinning and metallization processes. During the entire intervention process, the chips are in a sealed vacuum environment or a protective environment filled with high-purity nitrogen inert gas, avoiding the environmental pollution risks associated with traditional offline repair technologies.

[0055] For the special topologies of insulated-gate bipolar transistors (IGBTs), such as the bottom of the gate trench with a very large aspect ratio or the highly undulating termination guard ring region, the angle of the fiber array interference matrix probe in this embodiment is designed to be at... Adjustable within a certain range. Through the spatial coordinated motion of a multi-axis motion platform, oblique projection of the beam and fluid is achieved. This oblique projection path ensures that the intelligent screening agent can overcome structural barriers and penetrate deep into the bottom of the deep aspect ratio trench, achieving comprehensive, blind-spot-free selective intervention.

[0056] Example 2: Based on Example 1, this example provides a distributed collaborative intervention method to meet the rapid screening requirements of large-scale integrated IGBT wafers.

[0057] In step 1, the constructed distributed fiber array intervention matrix consists of multiple subarrays, each independently covering a quadrant region on the wafer. This distributed architecture allows the system to simultaneously perform parallel detection and intervention on multiple chips. In the waveguide unit design, this embodiment employs a composite structure based on a quartz core and polymer cladding to improve the fatigue life of the fiber under frequent bending motion. The precision displacement mechanism utilizes magnetic levitation drive technology, achieving a positioning accuracy on the order of 100 nanometers, which can meet the processing requirements of ultra-high density power devices.

[0058] In step 2, the formulation of the smart screening agent was specifically adjusted. The nanoparticles in the functionalized core are silver nanorods with a plasmon resonance effect. The length distribution of these silver nanorods is set so that their surface plasmon resonance peaks match the wavelength of the intervening laser. When irradiated by the laser, the silver nanorods generate a locally enhanced electric field, further accelerating the degradation process of the polymer shell and inducing localized micro-regional high temperatures in the core material at defects, achieving instantaneous eutectic welding. This method is particularly suitable for repairing surface voids caused by improper metallization processes.

[0059] In step 3, multi-wavelength polarization detection technology is introduced when acquiring electric field distribution data. By simultaneously emitting three different wavelengths of probe light, the polarization state changes at different wavelengths are measured. The system utilizes a multi-wavelength fitting algorithm to separate the coupling information between surface potential distribution and material refractive index fluctuations, improving the signal-to-noise ratio of electric field detection. The generated dynamic defect probability heatmap adds a time dimension. By comparing the heatmap changes over three consecutive sampling periods, the system can identify dynamic defects with growth trends, such as slowly extending cracks under stress, and assign them higher intervention priority.

[0060] In step 4, the training set of the deep learning model employs an incremental learning strategy. As the production line continues to operate, the system continuously collects new defect images and intervention result samples, and uses this new data to update the model parameters online. When calculating the adaptive intervention dosage, the system comprehensively considers the wiring density around the chip. If the defect point is close to a sensitive gate drive circuit, the system automatically calculates the divergence radius of the light beam and adjusts the focal length of the microlens to concentrate the intervention energy, ensuring that stray light does not cause logic damage to nearby sensitive circuits.

[0061] In step 5, the in-situ coaxial intervention process incorporates a screening agent recovery mechanism. A miniature vacuum nozzle is positioned at the edge of the fiber array probe. Within 10 microseconds of the smart screening agent being projected, the nozzle activates to generate localized negative pressure, removing excess microspheres that did not participate in the reaction. This synchronous recovery action ensures that no non-functional residues are left on the wafer surface, maintaining a high level of cleanliness for the semiconductor device surface. Fiber Bragg gratings are used to monitor the thermal deformation at the intervention point in real time. If the detected local strain value exceeds a preset structural strength threshold, the system immediately cuts off the laser output and injects cooling fluid for forced reset.

[0062] In this embodiment, the intelligent screening database is shared with multiple production devices via the cloud. When a device identifies a new defect morphology, its generated intervention template is instantly synchronized to other nodes across the entire production line. This group learning mechanism shortens the yield improvement cycle during the new product introduction period. Regarding automated production integration, the method in this embodiment supports the processing of thinned ultrathin wafers. Through the coordinated operation of a flexible adsorption base and a fiber optic probe, precise intervention on back-side defects is achieved without fragmentation.

[0063] For the back-side field stop layer (SLS) process of insulated gate bipolar transistors (IGBTs), the method described above can dynamically adjust the laser penetration depth based on the implanted ion depth information. Specifically, the laser energy density increases logarithmically with increasing implantation depth. This deep intervention capability allows the method to not only handle surface defects but also repair subsurface lattice distortions within the chip to a certain extent, optimizing the device's soft-turn-off characteristics.

[0064] Example 3: This example focuses on describing a smart screening agent intervention scheme based on multi-wavelength synergistic excitation, which aims to achieve stepwise repair of complex defects.

[0065] In step 1, a special hollow-core photonic crystal fiber is introduced into the fiber array interference matrix. The internal cavity of this fiber not only transmits the screening agent but also acts as a miniature chemical reactor. A photocatalytic coating is applied to the inner wall of the cavity. When the screening agent flows through the fiber, it is excited by specific ultraviolet controlled light, pre-initiating a preliminary molecular activation process. A crosstalk cancellation structure is employed between the waveguide units. By filling the gaps between the fibers with a high-absorption light-shielding material, energy leakage of high-power interference light between different channels is prevented.

[0066] In step 2, the smart screening agent is designed with a double-layered structure. The inner layer contains a highly efficient reducing agent, and the outer layer contains a cross-linking curing agent. This design allows for stepwise release via light excitation at different wavelengths. For example, a 405 nm laser is used to break the outer shell, releasing the curing agent to physically fill the defect area; a 355 nm laser is used to break the inner layer, releasing the reducing agent to reduce the metal oxide defects. This stepwise intervention strategy improves the success rate of repairing complex chemical defects.

[0067] In step 3, the acquisition of micro / nano topography data incorporates a white light interferometry mode. By introducing a broadband light source, the system can measure structural features with significant height variations, extending the measurement range from 10 nanometers to 500 micrometers. The electric field distribution acquisition module adds frequency domain analysis functionality; by changing the modulation frequency of the probe light, it can acquire the dynamic response characteristics of the charge on the chip surface, accurately distinguishing between transient charge injection defects and permanent physical traps.

[0068] In step 4, the adaptive intervention dosage calculation logic integrates a fluid dynamics simulation module. Based on the three-dimensional morphology of the defect points, the system simulates the wetting behavior of the screening agent spreading on the surface in real time. The calculated spraying pressure parameters are nonlinearly compensated for as surface roughness increases. The compensated pressure value equals the base pressure value multiplied by a roughness correction factor, which is determined by the arithmetic mean deviation of the surface profile. This simulation-guided dosage distribution solves the problem of uniform coverage on uneven surfaces.

[0069] In step 5, the in-situ coaxial intervention utilizes a spatial light modulator (SLM) to shape the wavefront of the intervention beam. By altering the phase distribution of the reflected wavefront, a light spot with a specific shape can be generated, such as a ring-shaped or rectangular spot, to perfectly fit defect regions of different geometries. During the intelligent screening agent projection process, ultrasonic-assisted atomization technology is introduced. High-frequency vibrations are generated using piezoelectric ceramics integrated at the end of the optical fiber, breaking the screening agent fluid into submicron-sized droplets, further enhancing the screening agent's penetration into extremely fine trenches.

[0070] This embodiment adds a post-intervention stress annealing step. After chemical repair is completed, the system controls a laser to perform localized heat treatment on the repaired area in a low-power scanning mode. This in-situ annealing process can eliminate residual internal stress generated during the chemical reaction, making the interface bonding between the repair layer and the original lattice stronger. In the real-time evaluation step, infrared thermography monitoring is added to verify whether the intervention measures have eliminated the potential for hot spots by monitoring the heating of the repair points under simulated loads.

[0071] The intelligent screening database not only records production data but also integrates Failure Mode and Effects Analysis (FMEA) logic. The system automatically performs root cause analysis on frequently occurring defect types and generates process optimization recommendations, which are then sent to the upstream control unit. For example, if a specific type of scratch is frequently detected in a certain area, the system will automatically suggest adjusting the end effector parameters of the front-end wafer handling robot. This closed-loop industrial intelligence system makes this invention not merely a repair tool, but a core brain for improving the overall semiconductor manufacturing process.

[0072] At the application level, this method exhibits extremely high adaptability to the guard ring structure of insulated gate bipolar transistors (IGBTs). The guard ring region is highly susceptible to surface leakage due to its complex potential gradient. This embodiment utilizes a fiber optic array interference matrix to precisely project a screening agent with a high dielectric constant into the gaps between the rings, and then uses laser-induced rapid curing to form a filling layer with excellent insulating properties. This not only repairs existing leakage points but also enhances the overall withstand voltage rating of the device.

[0073] Example 4: This example introduces a real-time screening and intervention scheme for extreme high-temperature production environments (such as the substrate heating state during the back metallization process).

[0074] In step 1, the fiber array intervention matrix is ​​externally encased in an active cooling water jacket. Cooling water circulates within a special stainless steel casing, ensuring that the internal fiber Bragg grating sensor is unaffected by nonlinear interference from high ambient temperatures. The end window of the intervention probe is made of sapphire, a material that is not only highly hard but also possesses excellent infrared transmittance and thermal conductivity, enabling it to withstand the impact of sputtered particles from within the chamber.

[0075] In step 2, the intelligent screening agent utilizes an inorganic ceramic precursor. Under the high energy density of a laser, this screening agent can be transformed into a high-temperature stable ceramic phase (such as silicon dioxide or aluminum nitride). This transformation process involves complex molecular pyrolysis and recombination, and its chemical reaction logic is pre-programmed into the control system. The carrier fluid for the screening agent is a perfluoropolyether with a high flash point, ensuring that spontaneous combustion or carbonization does not occur at the high-temperature operating point.

[0076] In step 3, the topography acquisition module incorporates wavelength scanning interferometry (WSI) measurement technology. By rapidly tuning the wavelength of the light source, depth scanning is achieved without mechanical movement. This is crucial for maintaining the mechanical stability of the system in high-temperature environments. Electric field data acquisition utilizes an electro-optic crystal with a high Curie temperature to ensure that the electro-optic coefficient remains constant at high temperatures, thus maintaining measurement accuracy.

[0077] In step 4, thermal radiation compensation is added to the adaptive intervention dose calculation logic. When calculating the laser peak power, the system automatically subtracts the background energy contribution due to ambient thermal radiation. The specific laser energy adjustment strategy is as follows: the current set power equals the target power minus the equivalent power of ambient radiation, and the equivalent power of ambient radiation is proportional to the fourth power of the ambient absolute temperature. This refined energy control avoids localized wafer melting caused by laser overheating in high-temperature environments.

[0078] In step 5, multiple laser beams were used in synergistic operation during in-situ coaxial intervention. In addition to the main intervention laser, two auxiliary preheating lasers were also configured to create a gentle temperature gradient around the intervention point. This thermal enclosure technology reduces the instantaneous temperature difference between the intervention point and the surrounding good-quality area, preventing stress cracks caused by thermal shock. The intelligent screening agent was projected using a pulsed spray mode, with the spray frequency strictly synchronized with the laser pulse frequency, achieving atomic-level precise matching of energy and matter.

[0079] The real-time evaluation step incorporates compositional analysis based on laser-induced breakdown spectroscopy (LIBS). By analyzing the plasma spectrum excited on the surface after intervention, the chemical bonding degree between the screening agent and the substrate is monitored in real time, ensuring that the composition ratio of the repair layer meets the preset semiconductor process requirements. The intelligent screening database records these specific physicochemical parameters at high temperatures, providing original data support for the future development of new semiconductor devices with higher temperature resistance.

[0080] In automated production line integration, the method in this embodiment communicates with the main control PLC of the production line via a high-speed fieldbus (such as EtherCAT). The transmission delay of intervention commands is controlled, ensuring that each chip can be screened and repaired before entering the next workstation during high-speed continuous production. For the back-side metallization process of insulated-gate bipolar transistors, this method can detect metal film peeling defects and achieve film re-attachment and strengthening through in-situ spraying of screening agents and laser remelting.

[0081] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for designing semiconductor selective intervention smart screening agents, characterized in that, Includes the following steps: Step 1: Construct an optical fiber array intervention matrix. Use multi-core optical fibers or hollow-core photonic crystal fibers to configure intervention probes at the workstations of the semiconductor production line to establish a multi-dimensional channel with functions of detection signal transmission, intervention beam guidance, and intelligent screening agent delivery. Step 2: Fill the internal cavity of the fiber array interference matrix with photoresponsive functional microspheres, which serve as smart screening agents, and pre-determine the phase transition behavior and chemical reaction logic of the smart screening agents under light excitation based on the physical characteristics of semiconductor surface defects. Step 3: Use the fiber optic sensor integrated in the fiber optic array intervention matrix to collect micro-nano topographic data and electric field distribution data of the surface of the chip to be processed in real time, and generate a defect probability heat map characterizing the severity of defects through data fusion. Step 4: Use a deep learning model to extract features from the defect probability heatmap and determine in real time the screening agent delivery rate, intervention beam energy, and irradiation duration corresponding to the defect point coordinates. Step 5: Control the fiber array intervention matrix to physically address the defect coordinates, and project the intervention beam and the smart screening agent synchronously to the defect point through the microlens structure of the fiber end face, so as to complete the physical encapsulation or chemical repair of the defect in situ.

2. The semiconductor selective intervention smart screening agent design method according to claim 1, characterized in that, In step 1, the fiber array interference matrix is ​​made of a material that is resistant to high temperature and plasma corrosion, and it integrates multiple waveguide units. The waveguide units are arranged in a matrix in a direction perpendicular to the chip surface, and each waveguide unit has an independent optical modulation function. The core diameter of the multi-core optical fiber is matched with the gate trench size of the semiconductor device; the intervention probe is integrated above the back metallization chamber or bonding stage, and the end face of the intervention probe is controlled by a precision displacement mechanism to maintain a preset gap distance with the chip surface.

3. The semiconductor selective intervention smart screening agent design method according to claim 2, characterized in that, In step 1, the fiber array intervention matrix is ​​also equipped with a fiber Bragg grating sensor. The fiber Bragg grating sensor utilizes the grating pitch generated by its internal periodically varying refractive index distribution, and its characteristic of changing with environmental parameters, to sense temperature fluctuations and mechanical stress changes in the production environment in real time. When the ambient temperature changes, the center wavelength of the reflected light signal shifts. The system calculates the product of the shift in the center wavelength and a preset proportional coefficient to obtain the current real-time temperature value, and feeds it back to the central control unit to correct the energy parameters of the intervention beam.

4. The semiconductor selective intervention smart screening agent design method according to claim 3, characterized in that, In step 2, the smart screening agent consists of a polymer shell with photosensitive properties and a functionalized core. The polymer shell is made of a photosensitive polymer containing photosensitive functional groups. When the photosensitive polymer is irradiated with interference light of a preset wavelength, it undergoes a photo-induced isomerization reaction or a photo-chain scission reaction. The functionalized core contains chemical agents for local passivation or nanoparticles for enhancing thermal conductivity. When the intervention beam irradiates the surface of the smart screening agent, the polymer shell releases its internal core material under the action of photon energy. The core material undergoes physical adsorption or chemical bonding with the lattice defects in the defect region, thereby inhibiting the further diffusion of defects.

5. The semiconductor selective intervention smart screening agent design method according to claim 4, characterized in that, In step 2, the delivery of the smart screening agent adopts a pressure-driven and photo-induced deformation coordinated control mode. A photosensitive lubricating layer of a preset thickness is coated on the inner wall of the hollow photonic crystal fiber. By adjusting the intensity of the auxiliary control light, the frictional resistance of the inner wall of the hollow photonic crystal fiber is changed, thereby achieving fine adjustment of the delivery rate of the smart screening agent. During the delivery process, the intelligent screening agent maintains a uniform particle distribution by adding a preset surfactant to the carrier fluid, thus avoiding aggregation and blockage in the multidimensional channels.

6. The semiconductor selective intervention smart screening agent design method according to claim 5, characterized in that, In step 3, the acquisition of the micro-nano morphology data is achieved using the principle of fiber optic interferometry. The fiber optic sensor emits a reference light and a probe light, and the probe light, after being reflected off the chip surface, coherently interferes with the reference light. By detecting the amount of movement of the interference fringes and performing phase unfolding calculations, the height undulation information of the chip surface can be obtained. The change in chip surface height is equal to the product of the number of shifts in the interference fringes and the laser wavelength, divided by twice the refractive index of the medium; The acquisition of the electric field distribution data is based on the electro-optic effect. The electro-optic crystal at the end of the optical fiber senses the surface electric field intensity and converts the change in electric field into a change in the polarization state of light waves, thereby indirectly measuring the non-uniform distribution of surface voltage.

7. The semiconductor selective intervention smart screening agent design method according to claim 6, characterized in that, In step 3, the generated defect probability heatmap uses a three-dimensional coordinate mapping method to map physical space coordinates to defect probability values ​​one by one. The process of generating the defect probability heatmap includes normalizing the original sensor data, removing environmental noise interference, and using an interpolation algorithm to numerically estimate the non-sampling points, thereby constructing a continuous defect density surface. When the probability of a defect at a certain coordinate point exceeds the preset alarm threshold, the system marks it as a high-risk intervention target and locks its geometric center coordinates as the intervention landing point.

8. The semiconductor selective intervention smart screening agent design method according to claim 7, characterized in that, In step 4, the deep learning model adopts a convolutional neural network architecture; The convolutional neural network architecture includes a plurality of convolutional layers for extracting the geometric features of defects, and a plurality of pooling layers for reducing data dimensionality and maintaining feature space invariance. The extracted feature vectors are input into a fully connected layer, compared with a pre-defined expert knowledge base in high-dimensional features, and output the optimal combination of intervention parameters for the defect type. The adaptive intervention dosage calculation logic incorporates a feedback compensation mechanism to monitor the temperature rise rate of the defective area in real time during the intervention process. If the temperature rise rate exceeds the preset safety limit, the duty cycle of the intervention beam is automatically reduced, and the flow rate of the smart screening agent is increased.

9. The semiconductor selective intervention smart screening agent design method according to claim 8, characterized in that, In step 5, the in-situ coaxial intervention achieves the transmission of high-power intervention laser and low-power detection light through the same optical path via an optical fiber coupler, ensuring that the coordinate center of the detection point and the energy center of the intervention point coincide in physical space. The microlens structure has an aspherical design, which can focus the light beam to a focused spot that matches the microstructure size of the semiconductor device; During the projection of the smart screening agent, the evanescent wave field of the optical fiber is used to pre-excite the smart screening agent, causing the smart screening agent molecules to enter a metastable state, thereby triggering a repair reaction at a preset energy threshold at the moment of contact with the defect surface.

10. A semiconductor selective intervention smart screening agent design method according to any one of claims 1 to 9, characterized in that, The method further includes the following steps: Step 6: Real-time evaluation of the intervention results. After the intervention is completed, the morphological data and electrical performance data of the intervention area are collected again by the in-situ fiber optic sensor, and the defect elimination rate is calculated by comparing the data differences before and after the intervention. The defect elimination rate is equal to the difference between the defect feature value before intervention and the defect feature value after intervention, and the ratio of the defect feature value before intervention. Step 7: If the defect elimination rate is lower than the preset quality judgment standard, the secondary intervention process will be automatically started, and the intervention strategy will be adjusted according to the evaluation results. Step 8: Simultaneously establish a smart screening database to record the defect distribution patterns of chips, corresponding intervention parameters, and final electrical performance test results. Use a long short-term memory network to learn from historical records and predict the potential failure trends of the current batch of chips.

Images