Control method for nanoliter droplet impact printing based on industrial vision AI compensation

Abstract

This invention discloses a nanoscale droplet impact printing control method based on industrial vision AI compensation, belonging to the field of printing control technology that integrates industrial vision and artificial intelligence. Addressing the issues of physical interference and visual delay, it synchronously generates ejection, stroboscopic, and shutter signals through an FPGA hardware trigger circuit, with a jitter deviation of less than 50 ns. Sub-pixel-level three-dimensional coordinates of the droplets are extracted using binocular vision and Gaussian surface fitting. The coordinates, electrostatic field gradient, and vibration spectrum are input into a GRU dynamic coupling prediction model. The GRU layer extracts temporal features through reset and update gates, while the physical constraint layer incorporates fluid dynamics equations to verify the trajectory and output residual correction values. The correction values ​​are directly written to the pulse generator register via an internal hard real-time adder within the FPGA, driving the platform to complete displacement compensation before droplet landing, without relying on software scheduling. This achieves hardware-level synchronization and physical fusion prediction, significantly improving droplet landing consistency under high-frequency inkjet printing.

Description

Technical Field

[0001] This invention relates to the field of printing control technology that integrates industrial vision and artificial intelligence, specifically a nanoscale droplet impact printing control method based on industrial vision AI compensation. Background Technology

[0002] With the continuous evolution of micro-nano manufacturing and precision printing technologies, nano-scale droplet printing, as a key process in fields such as flexible electronics, biomedicine, and semiconductor manufacturing, places extremely high demands on droplet landing accuracy and deposition consistency. The nano-scale droplet printing process involves complex rheological properties and microscopic physical interactions. Achieving precise arrangement of functional materials through high-frequency controlled droplet ejection is a core foundation for improving the performance of microelectronic devices and the quality of biological patterning.

[0003] The visual inspection and closed-loop control system plays a crucial role in ensuring printing accuracy. It primarily optimizes printing quality by capturing the droplet's attitude in real-time during flight and adjusting platform motion parameters. Existing visual control solutions typically utilize industrial cameras to acquire droplet images and perform geometric center analysis, aiming to dynamically correct the printing trajectory through real-time feedback of position information.

[0004] However, traditional visual correction techniques are typically based on a serial processing mode of detection and feedback, resulting in significant software-level latency and an inability to perform real-time trajectory compensation during the millisecond-level flight of droplets. Simultaneously, conventional control algorithms lack effective online sensing methods for complex industrial environmental disturbances, making it difficult to cope with inertial yaw caused by high-frequency mechanical vibrations and charge repulsion caused by static electricity accumulation in the printing chamber. Furthermore, existing control logic largely relies on linear models or single geometric extrapolations, making it difficult to capture nonlinear fluctuations in fluid dynamics. This leads to severe yaw and poor consistency at droplet landing points during high-frequency printing operations. Summary of the Invention

[0005] The purpose of this invention is to provide a nano-level droplet impact printing control method based on industrial vision AI compensation, which solves the complex physical interference such as mechanical vibration and electrostatic accumulation faced by nano-level droplet printing in industrial production environments, as well as the software-level delay and difficulty in dealing with nonlinear fluid dynamic fluctuations in traditional vision correction technology, thereby achieving high consistency and high precision control of droplet landing points under high-frequency inkjet printing conditions.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for controlling nanoscale droplet impact printing based on industrial vision AI compensation. The method includes the following specific steps:

[0008] Step 1: The droplet ejection pulse signal, strobe light source trigger signal and shutter synchronization signal are generated synchronously through the hardware trigger circuit of the field programmable gate array (FPGA). The FPGA is configured with timing control logic implemented in hardware description language to ensure that the jitter deviation between signals is less than 50ns.

[0009] Step 2: Capture at least two frames of sequential images of the droplet within its flight envelope using a binocular vision sensor. Extract the three-dimensional coordinates of the centroid of flight based on a Gaussian surface fitting algorithm. The Gaussian surface fitting calculates the sub-pixel-level geometric center by constructing the gradient distribution of the droplet edge region. The frame rate of a single camera in the binocular vision sensor is no less than 20,000 frames per second, and the observation window is set in the region from 0.5 mm to 2.0 mm after the droplet leaves the nozzle.

[0010] Step 3: Input the three-dimensional coordinates of the flight center of mass, environmental perception factors, and motion platform state vector into the Gated Recurrent Unit (GRU) dynamic coupling prediction model. The environmental perception factors include the spatial gradient distribution of the electrostatic field intensity in the printing chamber obtained in real time by a high-sensitivity electrometer, and the mechanical vibration spectrum characteristics of the motion platform obtained by a triaxial accelerometer. The GRU dynamic coupling prediction model calculates the Coulomb force by combining the pre-calibrated effective charge of the droplet with the real-time electric field intensity, and uses the mechanical vibration spectrum characteristics as an auxiliary input of the GRU layer, not directly as the force term of the droplet.

[0011] Step 4: Using the logic gate-level hard real-time adder integrated inside the FPGA, the residual correction value output by the GRU dynamic coupling prediction model is directly written into the pulse generator register, and the pulse output parameters of the motion controller are modified in real time to drive the motion platform to complete displacement compensation before the droplet lands; the total system response delay of the displacement compensation (from the acquisition of the droplet image to the start of platform displacement) is less than 2 milliseconds, of which the electronic delay is less than 1 millisecond.

[0012] In a preferred embodiment of the present invention, in step 1, the core timing unit of the pulse-driven hardware trigger circuit is configured as a field-programmable gate array (FPGA) chip, whose internal logic operating frequency is not lower than a preset frequency threshold. The FPGA chip constructs high-precision timing control logic using a hardware description language to ensure that the phase relationship between the droplet ejection pulse signal, the stroboscopic flash source trigger signal, and the shutter synchronization signal remains constant. The jitter deviation of its signal triggering is strictly controlled within a preset time deviation threshold, preventing droplet imaging blurring or position measurement deviation caused by timing drift, thereby eliminating random errors in visual detection at the physical level.

[0013] In a preferred embodiment of the present invention, in step 2, the binocular vision sensor includes two symmetrically arranged high-speed industrial cameras, with the frame rate of a single industrial camera not lower than a preset frame rate threshold. During the capture of multi-frame sequence images of the droplet, the exposure time of the industrial camera and the pulse width of the stroboscopic light source are adjusted to ensure that the motion blur of the droplet in the image is less than a preset pixel threshold. When extracting the three-dimensional coordinates of the centroid of flight at a preset accuracy level, the original image is first converted to grayscale and denoised, and then an edge localization algorithm based on Gaussian surface fitting is used. This algorithm calculates the gradient distribution of the droplet edge region, constructs a two-dimensional Gaussian function, and performs least-squares fitting to locate the sub-pixel coordinates of the droplet's geometric center. Through the binocular parallax mapping matrix, combined with pre-calibrated camera intrinsic and extrinsic parameters, the pixel coordinates of multiple dimensions are reconstructed into three-dimensional coordinates in physical space.

[0014] The observation window is set in the region of 0.5 mm to 2.0 mm after the droplet leaves the nozzle, ensuring that the droplet is at the transition point between the acceleration phase and the stable flight phase, and the length of the window is sufficient to acquire at least 2 frames of images at a frame rate of 20,000 frames per second.

[0015] In a preferred embodiment of the present invention, in step 3, the environmental sensing factors are acquired in real time by multiple high-frequency sensors distributed within the printing chamber. These high-frequency sensors include a high-sensitivity electrometer and a triaxial accelerometer. The electrometer monitors the spatial gradient distribution of the electrostatic field intensity within the printing chamber, acquiring the nonlinear deflection force generated by the electrostatic field on the charged droplet. The triaxial accelerometer is mounted at the end of the scanning axis of the motion platform, capturing in real time the mechanical vibration spectrum characteristics generated by the motion platform during high-frequency reciprocating motion. This environmental sensing data, along with the three-dimensional coordinates of the droplet's center of mass and the current position, velocity, and acceleration of the motion platform, constitutes the input vector of the prediction model, providing a comprehensive physical environment background for predicting droplet trajectory deviation.

[0016] As a preferred embodiment of the present invention, in step 3, the gated recurrent unit dynamics coupling prediction model adopts an architecture that integrates deep learning and physical mechanisms. This model specifically includes a GRU layer, a physical constraint layer, and an output layer.

[0017] The GRU layer uses reset and update gate mechanisms to model the position sequence data during droplet flight using long short-term memory, capturing trend deviations caused by airflow disturbances or viscoelastic fluctuations. Specifically, the reset gate... Decide how to input new data Memories of the past Combined, in which To reset the gate output, It is the sigmoid activation function. To reset the gate weight matrix, This is the hidden state from the previous moment. The input vector at the current time. Represents vector concatenation; update gate It controls the proportion of state information from the previous moment that is retained in the current state, among which To update the gate output, To update the gate weight matrix;

[0018] The physical constraint layer incorporates simplified fluid dynamics equations. Based on the initial ejection velocity, surface tension, density of the droplets, and the electrostatic force and air resistance models provided by environmental perception factors, it performs logical verification and constraint on the output results of the GRU layer, eliminating prediction results that do not conform to physical laws.

[0019] The overall force expression for the fluid dynamics equation is:

[0020]

[0021] in For the mass of the droplet, For acceleration, The effective charge of the droplet, as determined experimentally beforehand. For electric field strength, Aerodynamic viscosity, Where is the droplet radius, For air movement speed, For the droplet velocity, the mechanical vibration spectrum features are not directly used as force terms, but are used as input features of the GRU layer to compensate for positional disturbances of the nozzle and substrate;

[0022] The physical constraint layer receives the intermediate position prediction value output by the GRU layer and calculates the Coulomb force based on the real-time electrostatic field strength. And through kinematic equations To verify the rationality of the trajectory, among which for The droplet position vector at time t, The initial position vector, The initial velocity vector, It is the acceleration vector. For time.

[0023] The output layer ultimately generates residual offset vectors of the droplet relative to the theoretically designed landing point in both the lateral and longitudinal axes. .

[0024] As a preferred embodiment of the present invention, in step 4, the logic gate-level hard real-time adder integrated within the FPGA is constructed using pure hardware logic circuits. This adder is independent of the general-purpose processor and operating system software layer, and is directly deployed in the pulse generation core module of the motion control system. Upon receiving the residual correction value calculated by the dynamic coupling prediction model of the gated loop unit, the adder, triggered by the rising edge of the clock, directly accumulates the correction value into the interpolation instruction value of the preset trajectory and writes it directly into the FPGA pulse generator register. The total system response delay (from droplet image acquisition to the start of platform displacement) is less than 2 milliseconds, of which the electronic compensation delay (from algorithm output to register writing) is less than 1 millisecond. The mechanical response time of the motion platform is guaranteed to be within 1 millisecond by the high-dynamic linear motor (for micrometer-level compensation stroke).

[0025] This hardware-level computation mechanism ensures that the compensation delay time from the generation of the algorithm prediction result to the action of the hardware actuator is less than or equal to the preset response period, and lower than the response period of the reference software closed-loop system, thereby achieving platform displacement compensation before the droplet lands. The hardware trigger circuit, the hard real-time adder, and the FPGA pulse generator register form a hardware-level closed loop: the numerical change directly written to the register by the hard real-time adder takes effect in the next clock cycle, the displacement compensation delay depends only on the propagation delay of the hardware logic gates, does not depend on software scheduling, and the total compensation delay time is less than 1 millisecond.

[0026] As a preferred embodiment of the present invention, the nanoscale droplet impact printing control method based on industrial vision AI compensation further includes a closed-loop correction process in the height direction. A laser displacement meter is used to detect the flatness of the printing substrate surface in real time, with a sampling frequency no lower than a preset sampling frequency. The laser displacement meter feeds back the detected height fluctuation data to the vertical control loop of the motion control system, correcting the nozzle height relative to the substrate in real time to maintain the consistency of droplet flight height. This vertical compensation eliminates the droplet flight time variation caused by macroscopic topological undulations of the substrate surface, further solidifying the time series reference benchmark of the gated loop unit dynamic coupling prediction model.

[0027] As a preferred embodiment of the present invention, the nano-scale droplet impact printing control method based on industrial vision AI compensation includes a system self-healing monitoring mechanism. This mechanism monitors in real time the magnitude of the landing point residual correction value calculated in step 3, which is the square root of the sum of the squares of the deviations between the lateral and longitudinal axes.

[0028] When the magnitude of the residual correction value exceeds a preset safety threshold in N consecutive printing pulses, the nozzle is determined to be in an abnormal working state. Here, N is set according to the injection frequency f. (Rounded up), and N ranges from 10 to 500; the safety threshold is related to the droplet diameter D and is set to 0.6D. This abnormal state usually corresponds to physical faults such as nozzle scaling, ink buildup, or blockage. At this time, the system automatically pauses the current printing job and triggers the nozzle cleaning procedure, including automatic vacuum ink extraction and solvent rinsing. After cleaning, the system automatically restarts the calibration process to ensure the yield of subsequent printing jobs.

[0029] In a preferred embodiment of the present invention, the process of extracting the three-dimensional coordinates of the flight centroid in step 2 is completed within a preset observation window after the droplet leaves the nozzle. The selection of this window is based on the transition point between the acceleration phase and the stable flight phase of the nano-level droplet ejection, ensuring that the image captured by the visual sensor contains the complete morphological features of the droplet and its initial yaw trend. By sampling multiple frames of this specific region, the gated recurrent unit neural network can calculate the instantaneous rate of change of the droplet's acceleration, thereby more accurately inferring the nonlinear offset on the subsequent predetermined flight path.

[0030] In a preferred embodiment of the present invention, the displacement compensation in step 4 is achieved by superimposing commands during the motion of the motion platform. Without stopping the preset lateral and longitudinal scanning motion of the motion platform, the motion control card superimposes an auxiliary velocity vector onto the original velocity command stream via a hard real-time adder. This auxiliary velocity vector is calculated by dividing the residual correction value by the estimated remaining flight time. Through this dynamic superposition method, the motion platform maintains the overall continuity of the original working trajectory while completing path offset compensation at a preset accuracy level, avoiding the impact and vibration caused by frequent starts and stops on the mechanical system, and ensuring a smooth printing process.

[0031] As a preferred embodiment of the present invention, the gated cyclic unit dynamic coupling prediction model incorporates comparative experimental data as weights for the loss function during the training phase. The training set includes two sets of droplet landing data under the same physical environment, one with hard real-time compensation enabled and the other without. The loss function is specifically optimized for nonlinear drift under electrostatic interference environments by increasing the weight of yaw samples, thus giving the model a preset sensitivity to long-range deviations caused by electrostatics. After model training, it is deployed in an edge computing module equipped with hardware acceleration units to meet the computing power requirements of online real-time inference.

[0032] In a preferred embodiment of the present invention, the pulse width of the stroboscopic flash source trigger signal is set within a preset pulse width range. By shortening the stroboscopic flash illumination time, a clear image of the frozen droplet is formed on the photosensitive chip of a high-speed camera using high-intensity instantaneous illumination. The stroboscopic flash source employs a light-emitting diode array with a preset wavelength, whose photon energy helps to improve the contrast between the droplet edge and the background. Combined with a Gaussian surface fitting algorithm, the coordinate extraction accuracy can be improved to a preset pixel accuracy level, providing high signal-to-noise ratio input data for subsequent residual prediction.

[0033] This invention offers the following advantages: it achieves real-time compensation performance. By using a pulse-driven hardware trigger circuit composed of a field-programmable gate array (FPGA) and a logic gate-level hard real-time adder built into the motion control card, this invention breaks down the software communication barriers between sensor acquisition, host computer processing, and low-level execution in traditional vision control systems. Through hardware-level synchronization and addition operations, the total response delay from algorithm prediction to hardware execution is controlled within a preset time threshold, ensuring that the motion platform can complete physical displacement in advance within the extremely short flight time of a nanoliter droplet, thus achieving real-time compensation during flight.

[0034] It possesses strong resistance to interference from complex environments. Addressing common nonlinear interference factors in industrial production, such as high-frequency mechanical vibration and electrostatic accumulation in the printing chamber, this invention introduces a gated cyclic unit dynamic coupling prediction model. This model not only utilizes deep learning to extract time-series features of the flight trajectory but also incorporates fluid dynamics physical constraints, making the prediction results more consistent with real physical laws. By applying this invention, in environments with severe electrostatic fluctuations and mechanical vibrations, the standard deviation of droplet landing points can be significantly reduced, resulting in a predetermined increase in landing accuracy and significantly solving the problem of scattered landing points.

[0035] The system has established self-healing capabilities and high reliability. By monitoring the real-time modulus of the landing point residual correction value, the system can automatically identify and warn of the physical deterioration state of the nozzle. This intelligent discrimination logic based on deviation trends enables the system to automatically trigger the cleaning procedure before the nozzle develops severe ink buildup or clogging, greatly reducing the waste rate in industrial production and improving the continuous operation capability and yield of the production line.

[0036] This invention achieves multi-dimensional dynamic collaborative control. Through the coordinated operation of a laser displacement meter and a binocular vision sensor, it simultaneously realizes comprehensive closed-loop management of the surface topological features of the printing substrate and the spatial trajectory of the droplets. This coupling of vertical height correction and horizontal residual compensation enables nanoscale droplet impact printing to exhibit high consistency and stability when handling complex processes such as flexible electronics and biological patterning, providing solid technical support for precision printing in the field of micro-nano manufacturing. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the overall technical solution architecture of a nano-scale droplet impact printing control method based on industrial vision AI compensation, according to an embodiment of this application.

[0038] Figure 2 This is a schematic diagram of the core principle framework of the gated loop unit dynamic coupling prediction model in a nanoscale droplet impact printing control method based on industrial vision AI compensation according to an embodiment of this application.

[0039] Figure 3 This is a schematic diagram of data flow in a nano-scale droplet impact printing control method based on industrial vision AI compensation according to an embodiment of this application. Detailed Implementation

[0040] Example 1

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0042] In the nanoscale droplet impact printing control method based on industrial vision AI compensation provided by this invention, the core of the system achieves nanosecond-level deterministic timing control at the physical layer through a high-performance field-programmable gate array (FPGA). This method deeply aligns visual perception, AI model prediction, and motion control execution on the time axis through the low-level integration of hardware triggering circuits.

[0043] See Figures 1-3 The method described in this embodiment of the invention includes the following steps:

[0044] In step S1, the droplet ejection pulse signal, the stroboscopic light source trigger signal, and the shutter synchronization signal of the binocular vision sensor are synchronously generated through the hardware trigger circuit of the FPGA. The FPGA is configured with timing control logic implemented in a hardware description language to ensure that the jitter deviation between signals is less than 50ns. In the engineering implementation of this step, the pulse-driven hardware trigger circuit uses a field-programmable gate array chip with an extremely high internal clock frequency as the core timing unit. The chip internally uses a parallel, non-blocking timing generation module built with a hardware description language, and its logic operating frequency is set above 100MHz, thereby ensuring that the jitter deviation of the signal trigger is strictly locked within 50ns. This high-precision timing coordination is crucial for nanoscale droplet printing because the initial velocity of the droplet at the nozzle exit is typically between 2 m / s and 8 m / s, and even a tiny timing drift can cause positional deviations in visual imaging.

[0045] In practice, after receiving the print enable command from the main control system, the field-programmable gate array (FPGA) chip synchronously pulls up the levels of three independent output pins on the rising edge of the clock: the first pin is connected to the drive power module of the piezoelectric nozzle, triggering controlled piezoelectric ceramic deformation to eject droplets; the second pin is connected to the constant current drive circuit of the high-brightness stroboscopic light source, generating narrow pulse illumination with a pulse width between 1 and 5 microseconds to achieve "frozen" imaging of high-speed flying droplets; and the third pin is connected to the external trigger interface of the binocular vision sensor via a differential signal transmission line to precisely open the shutter for exposure. Through this hardware-level synchronization scheme, ejection, illumination, and exposure are completely synchronized on the physical timeline, eliminating the random delay caused by software interrupts in general-purpose processors.

[0046] In step S2, a multi-frame sequence of images of the droplet within its flight envelope is captured using a binocular vision sensor. The three-dimensional coordinates of the flight centroid are extracted based on a Gaussian surface fitting algorithm, where the Gaussian surface fitting calculates the sub-pixel-level geometric center by constructing the gradient distribution of the droplet's edge region. In this step, the binocular vision sensor consists of two symmetrically arranged high-speed industrial cameras, each with a frame rate set to at least 20,000 frames per second and a resolution of at least 1280×1024 pixels. The cameras are mounted on either side below the printhead, forming an intersection angle of approximately 30 to 45 degrees. The capture process is set within a preset observation window, typically selected within 0.5 mm to 2.0 mm of the droplet after it leaves the nozzle.

[0047] In the image processing stage, the original acquired grayscale image is first subjected to median filtering to remove electronic noise. Then, the system employs an edge localization algorithm based on Gaussian surface fitting to extract the droplet centroid. Specifically, the algorithm searches for pixels at the droplet edge and calculates the grayscale gradient distribution in a 3×3 or 5×5 neighborhood around each edge pixel. By constructing a two-dimensional Gaussian distribution function, the energy distribution of the edge region is fitted using the least squares method to find the optimal solution for the energy center, thereby improving the centroid localization accuracy to the 0.1 pixel level. With a frame rate as high as 20,000 frames per second, the droplet's flight time within a 0.5mm~2.0mm window is approximately 0.25~1.0 milliseconds, allowing for the acquisition of 3~12 frames, meeting the requirements for multi-frame sequences. After obtaining the sub-pixel two-dimensional coordinate pairs of the left and right cameras, the three-dimensional coordinate sequence of the droplet in physical space is calculated using a pre-calibrated binocular parallax mapping matrix (containing the camera's intrinsic parameter matrix, rotation matrix, and translation vector) through the principle of triangulation reconstruction.

[0048] In step S3, the three-dimensional coordinates of the flight center of mass, environmental sensing factors, and the motion platform state vector are input into the GRU dynamic coupling prediction model. The environmental sensing factors include the spatial gradient distribution of the electrostatic field intensity within the printing chamber, acquired in real-time by a high-sensitivity electrometer, and the mechanical vibration spectrum characteristics of the motion platform, acquired by a triaxial accelerometer. This step is the core logic for achieving predictive compensation. The environmental sensing factors are acquired in real-time by multi-channel high-frequency sensors distributed within the printing chamber. Specifically, the electrometer monitors the spatial gradient distribution of the electric field intensity between the nozzle and the substrate at a sampling frequency of 10 kHz, while the triaxial accelerometer is installed at the ends of the X-axis and Y-axis scanning beams of the motion platform to capture the mechanical vibration spectrum generated during the high-speed reciprocating motion of the platform, including mid-to-high frequency fluctuations from 50 Hz to 500 Hz due to resonance. The motion platform state vector includes the current command position, real-time feedback velocity, and acceleration planning parameters.

[0049] The GRU-dynamic coupling prediction model adopts an architecture that integrates deep neural networks and physical mechanics equations, specifically including a GRU layer, a physical constraint layer, and an output layer.

[0050] The GRU layer is specifically responsible for processing time series features. It uses the reset gate and update gate mechanism to model the droplet flight centroid sequence extracted in step S2 using long short-term memory.

[0051] The formula for resetting the door is ,in To reset the gate output, It is the sigmoid activation function. To reset the gate weight matrix, This is the hidden state from the previous moment. The input vector at the current time. This indicates vector concatenation.

[0052] The updated gate formula is ,in To update the gate output, To update the gate weight matrix. The GRU layer can effectively capture the subtle trajectory deviation trends of droplets caused by air disturbances and viscoelastic fluctuations during flight.

[0053] The physical constraint layer intervenes in the model calculation, with a built-in simplified fluid dynamics equation. Its overall force expression is as follows:

[0054]

[0055] in For the mass of the droplet, For acceleration, The effective charge of the droplet was pre-calibrated experimentally (calibration method: under no airflow interference, a known voltage was applied to the nozzle, the deflection distance of the droplet in a known electric field was measured, and the typical charge was calculated by reverse calculation). For electric field strength, Aerodynamic viscosity, Where is the droplet radius, For air movement speed, The velocity of the droplet is used; the mechanical vibration spectrum features are not directly used as force terms, but are used as input features of the GRU layer to compensate for positional disturbances of the nozzle and the substrate.

[0056] The physical constraint layer receives the intermediate position prediction values ​​output by the GRU layer and calculates the Coulomb force based on the real-time electrostatic field strength. And through kinematic equations To verify the rationality of the trajectory. Among them... for The droplet position vector at time t, The initial position vector, The initial velocity vector, It is the acceleration vector. For time.

[0057] The model weighted and fused the statistical predictions from the GRU layer with the mechanical verification results from the physical constraint layer, outputting the residual offset vectors of the droplet relative to the theoretically designed landing point in the X and Y axes. The final output layer generates residual correction values ​​that have been verified against physical laws.

[0058] In step S4, the residual correction value output by the GRU dynamic coupling prediction model is directly written into the pulse generator register through the logic gate circuit-level hard real-time adder integrated inside the FPGA, thereby modifying the pulse output parameters of the motion controller in real time and driving the motion platform to complete displacement compensation before the droplet lands.

[0059] The hard real-time adder is a logic gate-level hardware implementation within the field-programmable gate array (FPGA) logic of the motion control card, independent of the operating system's kernel scheduling. After the model calculates the residual correction value, this value is loaded into the adder's input register via a high-speed bus interface (such as PCIe or Gigabit Ethernet). Within the next interpolation cycle (typically 125 or 250 microseconds), the hard real-time adder directly modifies the value in the motion controller's FPGA pulse generator register. A hardware-level closed loop is formed between the hardware trigger circuit, the hard real-time adder, and the FPGA pulse generator register: the value change directly written to the register by the hard real-time adder takes effect within the next clock cycle; the displacement compensation delay depends only on the propagation delay of the hardware logic gates, not on software scheduling, and the total compensation delay time is less than 1 millisecond.

[0060] The displacement compensation is implemented by using the "overlay in motion" command of the motion platform to overlay an auxiliary compensation vector calculated from the residual correction value onto the current velocity vector without stopping or changing the current XY axis main scanning trajectory. The total system response delay (from droplet image acquisition to platform displacement) is less than 2 milliseconds, of which the electronic compensation delay (from algorithm output to register writing) is less than 1 millisecond. The motion platform uses a high dynamic response linear motor, and the mechanical response time for the micrometer-level compensation stroke is less than 1 millisecond. Since the flight time of a nano-scale droplet from the nozzle to the substrate is typically between 1 and 3 milliseconds, this millisecond-level response ensures that the system can complete the physical position adjustment of the platform "mid-flight" before the droplet lands, thereby offsetting the predicted offset.

[0061] As a further optimization, this embodiment also includes a closed-loop correction process in the height direction. A laser displacement meter is used to detect the flatness of the printing substrate surface in real time. The laser displacement meter is installed on the side of the nozzle and uses high-frequency sampling to acquire macroscopic topological undulation information of the substrate surface. The detected height fluctuation data is directly fed back to the vertical Z-axis control loop of the motion control system. The Z-axis motor corrects the absolute height of the nozzle relative to the substrate in real time to maintain a constant droplet flight height. This vertical compensation not only ensures a consistent spread pattern when the droplets impact the substrate, but more importantly, it solidifies the droplet flight time, thus providing a stable time reference for the GRU model in step S3 and avoiding prediction distortion caused by flight time fluctuations.

[0062] Furthermore, this embodiment includes a monitoring mechanism based on system self-healing. The system continuously monitors in real time the magnitude of the landing point residual correction value calculated in step S3, i.e., the length of the residual vector: .

[0063] When the modulus If a preset safety threshold is exceeded during N consecutive printing pulses, the system logic determines that the nozzle is in an abnormal physical state. Here, N is set according to the injection frequency f. (Round up), for example, when the jet frequency is 1kHz, The safety threshold is related to the droplet diameter D and is set to 0.6D. For example, for a droplet diameter of 50... The time threshold is 30 At this point, the system automatically triggers a pause command, moves the motion platform to the maintenance station, and initiates the nozzle cleaning procedure. The cleaning procedure includes starting a vacuum pump for negative pressure ink suction and using high-frequency ultrasonic vibration of piezoelectric ceramics combined with solvent rinsing to thoroughly remove accumulated ink or dried material from the nozzle orifice. After cleaning, the system automatically restarts the calibration process, confirming the residual modulus value has returned to the normal range by spraying test droplets and using binocular vision, before resuming production.

[0064] In the training phase of the GRU-dynamic coupling prediction model, this embodiment introduces special contrastive loss function weights. The training dataset was obtained experimentally, covering two sets of landing data under typical interference environments such as a 2kV electrostatic field and 50Hz mechanical vibration, with and without the hardware real-time compensation of this invention enabled. In the loss function design, higher weight coefficients were assigned to long-range, nonlinear yaw samples caused by electrostatics. The model was trained to not only predict linear displacement but also perceive complex nonlinear drift patterns. After training, the model was deployed in an embedded edge computing module equipped with a hardware tensor acceleration unit to meet the stringent real-time requirements of industrial environments.

[0065] Through the synergy of the above steps, when performing nanoscale printing tasks ranging from 100 pL to 500 pL, the system can reduce the standard deviation of droplet landing from the original 18.5 micrometers to less than 2.8 micrometers under conditions of 50Hz mechanical vibration and strong electrostatic interference.

[0066] Example 2

[0067] This embodiment focuses on the specific application and parameter configuration of the method described in this invention in the printing scenario of bioactive materials (such as cell-containing bio-inks). Bio-inks typically have high viscoelasticity and are sensitive to shear forces; the suppression of satellite droplets and their flight trajectory during the droplet breakage stage are more significantly affected by the environment.

[0068] In step S1, to meet the specific needs of bioprinting, the pulse-driven hardware trigger circuit adjusts the droplet ejection pulse signal into a multi-segment bipolar pulse waveform. Through high-precision timing using a field-programmable gate array (FPGA), the switching interval between the forward drive pulse, the reverse buffer pulse, and the residual oscillation suppression pulse is controlled to the order of 100 nanoseconds. The pulse width of the stroboscopic light source trigger signal is compressed to 800 nanoseconds to handle the potentially violent morphological pulsations that may occur during the high-speed flight of the biological droplets, ensuring the edge sharpness of the images acquired by the binocular camera.

[0069] In step S2, taking advantage of the high transparency of the biological droplets, the image processing algorithm incorporates background subtraction technology. A binocular vision sensor acquires an empty background image before printing begins. In subsequent processing, the real-time image containing the droplets is subtracted from the background image to highlight the droplet outline. When extracting the three-dimensional coordinates of the flight centroid, a Gaussian surface fitting algorithm adds feature extraction of the droplet's aspect ratio to assess the morphological stability of the biological droplets.

[0070] In step S3, the input vector of the GRU-dynamic coupling prediction model is supplemented with the real-time ink viscosity parameter obtained by the viscosity-temperature sensor. Since the bioprinting environment is typically maintained at 37 degrees Celsius, even small temperature fluctuations can cause drastic changes in viscosity, thus affecting the initial ejection velocity. The fluid dynamics equations in the physical constraint layer are supplemented with a viscoelastic correction term for non-Newtonian fluids. Based on the captured droplet centroids from the first three frames, combined with the real-time environmental electrostatic gradient, the model predicts the nonlinear yaw of the droplet as it flies towards the culture dish substrate. At this point, the mechanical vibration spectrum features in the environmental sensing factor are primarily used to filter out low-frequency airflow disturbances caused by the incubator's ventilation system.

[0071] In step S4, since biological substrates (such as hydrogel surfaces) often have complex microscopic topography, the hard real-time adder receives micron-level height feedback from a laser displacement meter while simultaneously superimposing XY-axis displacement compensation. The response period for displacement compensation is set to 200 microseconds to match the relatively low jetting frequency in bioprinting (typically between 100Hz and 500Hz). The compensation amount is accumulated in real time into the current control loop of the linear motor, directly bypassing the instruction interpolation layer of the host computer, and achieving instantaneous disturbance compensation for the scanning trajectory by modifying the underlying registers of the driver.

[0072] In this application scenario, this embodiment directly modifies the values ​​of the field-programmable gate array pulse generator using a hard real-time adder, maintaining the compensation delay time at 0.8 milliseconds. Experiments show that, for bio-ink with a viscosity of 15 mPa·s, under conditions of external low-frequency vibration, the landing accuracy is improved by more than 50% compared to the traditional vision-based closed-loop solution.

[0073] Example 3

[0074] This embodiment details the specific implementation of the present invention in inkjet printing of flexible circuit boards. In this scenario, the printing substrate is a PET film, which is prone to static electricity accumulation and has a significant repulsive effect on nano-level charged droplets.

[0075] In step S1, the controller generates a high-precision phase synchronization signal through a field-programmable gate array, ensuring that the center alignment error between the shutter of the binocular camera and the stroboscopic light source is less than 20 ns. To address the strong reflection from the PET film surface, the stroboscopic light source employs a blue light-emitting diode array with a preset wavelength of 450 nanometers. Its high photon energy effectively penetrates the thin film interference on the surface of the bio-ink, improving image contrast.

[0076] In step S2, when extracting the sub-pixel flight centroid, the system introduces a refractive index compensation parameter into the binocular parallax mapping matrix to correct the non-uniformity of air refractive index caused by the local temperature rise in the printing area. The Gaussian surface fitting algorithm is deployed on the local processor of the motion control card through a parallel computing architecture, achieving online real-time extraction of 2000 droplet centroids per second.

[0077] In step S3, the GRU-dynamic coupling prediction model is specifically enhanced to compensate for nonlinear drift under electrostatic interference. A high-sensitivity electrometer in the environmental sensing factor performs a non-contact scan 5 mm above the PET film surface to acquire a charge distribution map. The GRU layer extracts the acceleration variation trend caused by charge repulsion based on historical trajectory data. The physical constraint layer establishes a charged particle motion model under electrostatic field, performs real-time calibration against landing prediction values, and outputs the offset component. .

[0078] In step S4, the motion control card's built-in hard real-time adder uses 32-bit floating-point logic to ensure the accuracy of the compensation value. Displacement compensation is achieved through the "overlay in motion" instruction, the core of which lies in compensating the velocity vector... It is superimposed on the original uniform scanning vector in real time, where This represents the predicted remaining flight time. This dynamic overlay method avoids the flexible impact caused by sudden position corrections during high-speed movement of the robotic arm.

[0079] In the system of this embodiment, for the electrostatic environment of up to 5kV on the PET film surface, by real-time monitoring of the modulus of the landing point residual correction value, the system can automatically determine whether to activate the ionization fan for local static elimination, further demonstrating the system's self-healing and intelligent capabilities. Experimental data show that under strong electrostatic interference, the consistency of landing points is improved by approximately 68% compared to the uncompensated state.

[0080] Example 4

[0081] This embodiment describes the application of the present invention in the printing and manufacturing of micro-nano optical devices, particularly for the precision forming control of high refractive index optical polymer droplets.

[0082] In step S1, a field-programmable gate array (FPGA) serves as the timing core, refining the clock cycle to improve the interval stepping accuracy between droplet ejection and strobe illumination to 10 nanoseconds. This extremely high-precision timing control allows the system to capture the minute necking features at the moment of droplet breakage, providing initial conditions for subsequent dynamic modeling.

[0083] In step S2, the binocular vision sensor employs a high-magnification microscope objective, with its depth of field precisely controlled near the droplet's flight trajectory. During the acquisition of multiple image sequences, the system calculates the energy gradient of the images in real time. When the droplet deviates from the depth of field center, real-time focusing compensation is performed using the micro-displacement platform in step S4. The centroid extraction algorithm employs iterative Gaussian fitting, eliminating the influence of droplet deformation on center positioning through multiple regression calculations.

[0084] In step S3, the GRU-dynamic coupling prediction model was parameter-tuned for high-viscosity optical polymers. Since such inks are greatly affected by air drag during flight, the physical constraint layer introduced a dynamic equation for a variable mass point, considering the droplet mass reduction and drag coefficient changes caused by solvent evaporation. The motion platform state vector in the input vector was refined to the tracking error of the axis system, thereby eliminating known tracking errors of the mechanical axis system in the prediction.

[0085] In step S4, the hard real-time adder communicates directly with the driver's power module at the low level. After the residual correction value is input, the adder directly modifies the pulse width modulation (PWM) counter inside the FPGA, achieving sub-micron level correction of the platform position through microsecond-level current vector adjustment. The compensation delay time is compressed to less than 0.5 milliseconds.

[0086] In this embodiment, a self-healing monitoring mechanism is used to monitor the drying and clogging of optical materials. Once the residual modulus is detected to fluctuate continuously beyond a set range, the system determines that the nozzle has experienced minor drying due to solvent evaporation, and then initiates a solvent atomization and wetting process to automatically restore nozzle activity. In actual production, this approach ensures high-precision arrangement of the optical lens array, with the center position deviation of adjacent lenses controlled within 5 micrometers.

[0087] Example 5

[0088] This embodiment describes the implementation details of the present invention in the continuous production of large-area flexible electronic devices, focusing on the realization of visual AI compensation under multi-nozzle collaboration.

[0089] In step S1, the pulse-driven hardware trigger circuit has multi-channel parallel processing capabilities, enabling it to simultaneously drive more than eight piezoelectric nozzles. The field-programmable gate array (FPGA) uses multiplexing technology to assign an independent, finely adjustable timing offset to each nozzle to compensate for minor differences in physical characteristics between different nozzles. A tight mapping relationship is established between the shutter synchronization signal of the binocular vision sensor and the emission sequence of all nozzles.

[0090] In step S2, for the dense droplet swarm generated by multiple nozzles, the visual algorithm uses a deep learning-based object detection network to pre-determine the droplet envelope region of each nozzle, and then executes a Gaussian surface fitting algorithm in parallel within each independent sub-region. The extracted three-dimensional coordinates of the centroids of multiple droplets are encapsulated in time-stamped data frames and transmitted to the subsequent model.

[0091] In step S3, the GRU-dynamic coupling prediction model employs a multi-task learning architecture. The model not only outputs the landing point residuals of individual droplets but also considers the mutual interference between adjacent droplets during flight due to airflow field coupling. An environmental perception factor is added, incorporating real-time ink supply pressure fluctuations fed back from the flow meter. The physical constraint layer uses large-scale parallel computation to solve the simplified multibody dynamics model in real time, predicting the overall migration trend of the droplet swarm in complex airflow environments.

[0092] In step S4, since multi-head printing typically employs a large gantry-type motion platform, the hard real-time adder needs to handle greater inertia compensation. The motion control card sends correction commands to each driven axis via the hard real-time bus. Without changing the main motion speed of the gantry architecture, high-frequency fine-tuning during large-stroke motion is achieved by superimposing a high-frequency response piezoelectric micro-motion displacement compensation mechanism (Piezo-Stage). The compensation time response is strictly controlled within 1 millisecond, ensuring precise droplet landing in high-speed roll-to-roll production.

[0093] In this embodiment, a laser displacement meter scans the dynamic wrinkles on the surface of the flexible substrate in real time, incorporating height fluctuations into the Z-axis feedback loop in real time. Through this dynamic coordination, the consistency of printing accuracy is significantly improved when processing flexible substrates with a width of up to 1 meter.

[0094] Example 6

[0095] This embodiment details the calibration and robustness verification process of this method under extreme laboratory conditions. The experimental environment simulates a combined condition of strong electromagnetic interference and large-amplitude mechanical vibration.

[0096] In step S1, the field-programmable gate array is designed with electromagnetic compatibility (EMC) hardening, and its pulse drive hardware trigger circuit adopts differential isolation technology to ensure that the signal jitter is still maintained within 50ns in a field strength environment of 10V / m.

[0097] In step S2, the binocular vision sensor is mounted on a specially designed shock-absorbing bracket to isolate external vibrations to the maximum extent. However, some residual vibrations may still remain during the image acquisition stage. The algorithm calculates the camera's vibration components in real time by setting stationary reference objects at the image edges and subtracts them during centroid extraction using a coordinate transformation matrix, ensuring the purity of sub-pixel coordinates.

[0098] In step S3, a large amount of simulated disturbance noise data was added to the training of the GRU-dynamic coupling prediction model. A regularization term was introduced into the loss function to prevent the model from overfitting when faced with random sudden disturbances. Comparative experiments showed that when the environmental vibration acceleration reached 0.5g, the model was still able to accurately identify the force yaw characteristics of the droplet, rather than misjudging it as camera shake.

[0099] In step S4, the hard real-time adder implements logic limiting protection during the modification process of motion control commands. If the predicted residual correction value exceeds the physical response limit of the mechanical mechanism, the adder automatically performs command smoothing to prevent mechanical overshoot. This protection mechanism ensures that the system will not pose a collision risk due to erroneous predictions in extreme environments.

[0100] In this embodiment, the system's self-healing monitoring is manifested in the identification of "compensation instability". If the standard deviation of the residual modulus continues to rise over a period of time, the system will automatically reduce the printing speed and increase the visual sampling density until environmental interference returns to a controllable range, demonstrating extremely high robustness.

[0101] Example 7

[0102] This embodiment describes the application of the present invention in metal nanoparticle inkjet printing, mainly targeting the high continuity requirements of conductive wiring.

[0103] In step S1, to achieve effective spraying of the high surface tension metal slurry, the pulse waveform generated by the pulse drive circuit has a steep falling edge. The field-programmable gate array (FPGA) achieves precise control of the waveform slope through high-frequency PWM modulation.

[0104] In step S2, the specific spectral characteristics of the stroboscopic source are utilized to enhance the reflection contrast of the metal droplet surface. When processing metal droplet images, the Gaussian surface fitting algorithm adds the identification of droplet "tailing" phenomena. The centroid extraction process automatically ignores the tail filaments at the break edge, locking only the position of the main droplet to ensure accurate coordinate input.

[0105] In step S3, the GRU-dynamic coupling prediction model was weighted to account for the high-density characteristics of the metal paste. The force analysis in the physical constraint layer focused on the inertial force component. An environmental sensing factor was used to monitor the humidity inside the printing chamber in real time, as humidity directly affects the solvent evaporation rate of the paste during flight.

[0106] In step S4, the hard real-time adder feeds back the compensation result to the high-precision linear guide. Because metal wiring requires extremely high straightness, the adder not only corrects discrete landing points but also converts the residual correction value into a smooth auxiliary speed curve through continuous differential operations, thereby achieving trajectory fitting and correction for continuously printed lines.

[0107] This solution effectively addresses the common problems of conductive line breakage and short circuits in flexible electronics manufacturing. By real-time correction of landing point deviation, the center deviation of the metal slurry landing point is limited to within 15% of the droplet diameter.

[0108] Example 8

[0109] This embodiment describes a specific implementation of the present invention in the additive manufacturing of three-dimensional microstructures. Due to the need for layer-by-layer stacking, extremely high requirements are placed on the vertical landing consistency of the droplets.

[0110] In step S1, the hardware trigger circuit adds synchronization logic with the Z-axis position sensor. The FPGA will only issue a printing trigger signal when the Z-axis reaches the preset slice height plane, achieving dual hard synchronization of spatial position and time.

[0111] In step S2, the binocular vision sensor not only captures the horizontal flight trajectory of the droplet but also obtains the vertical distance of the droplet relative to the surface of the printed layer through 3D reconstruction. The Gaussian surface fitting algorithm is executed in 3D space to extract the droplet's full attitude information in space.

[0112] In step S3, the model not only predicts the planar residuals It also predicts the impact kinetic energy of droplets in the Z-axis direction. A physical constraint layer introduces a collision fluid model to evaluate the spreading behavior of droplets after impacting existing structures. An environment-aware factor in the input vector incorporates real-time monitoring of the surface charge of the printed layers to prevent misalignment of stacked layers due to repulsion of like charges.

[0113] In step S4, the hard real-time adder operates simultaneously on the XY scan axis and the Z-axis control loop. Without stopping the lateral scan, a Z-axis compensation pulse is superimposed in real time based on the predicted substrate flatness deviation. The compensation delay time is less than 1 millisecond.

[0114] Through this multi-dimensional synchronous compensation, the interlayer misalignment error of the three-dimensional microstructure is reduced to within 10 micrometers, which greatly improves the molding accuracy and mechanical strength of the microstructure.

[0115] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for controlling nanoscale droplet impact printing based on industrial vision AI compensation, characterized in that, include: The droplet ejection pulse signal, strobe light source trigger signal and shutter synchronization signal are generated synchronously through the hardware trigger circuit of the field programmable gate array (FPGA). The FPGA is equipped with timing control logic implemented by hardware description language to ensure that the jitter deviation between signals is within the preset timing tolerance range. At least two frames of sequential images of a droplet within its flight envelope are captured using a binocular vision sensor. The three-dimensional coordinates of the centroid of flight are extracted based on a Gaussian surface fitting algorithm, wherein the Gaussian surface fitting calculates the subpixel-level geometric center by constructing the gradient distribution of the droplet edge region. The frame rate of the binocular vision sensor is matched with the droplet's flight speed to obtain at least two frames of sequential images. The three-dimensional coordinates of the flight center of mass, environmental perception factors, and the state vector of the motion platform are input into the dynamic coupling prediction model of the gated cyclic unit (GRU). The environmental perception factors include the spatial gradient distribution of the electrostatic field intensity in the printing chamber, which is obtained in real time by a high-sensitivity electrometer, and the mechanical vibration spectrum characteristics of the motion platform, which are obtained by a triaxial accelerometer. The GRU dynamic coupling prediction model calculates the Coulomb force by combining the pre-calibrated effective charge of the droplet with the real-time electric field intensity, and uses the mechanical vibration spectrum characteristics as an auxiliary input of the GRU layer, not as a force term of the droplet. By using the logic gate-level hard real-time adder integrated within the FPGA, the residual correction value output by the GRU dynamic coupling prediction model is directly written into the pulse generator register, thereby modifying the pulse output parameters of the motion controller in real time and driving the motion platform to complete displacement compensation before the droplet lands.

2. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, The specific configuration process of the hardware trigger circuit is as follows: The FPGA has a parallel, non-blocking timing generation module built inside, and its logic operating frequency is set to be no less than 100MHz; After receiving the print enable command from the main control system, the FPGA synchronously pulls up the level of three independent output pins on the rising edge of the clock: The first pin is connected to the drive power module of the piezoelectric nozzle, triggering the controlled deformation of the piezoelectric ceramic to spray droplets. The second pin is connected to the constant current drive circuit of the high-brightness stroboscopic light source, generating a pulse width of... to The narrow pulse illumination between the pins performs freeze imaging on the high-speed flying droplets; the third pin is connected to the external trigger interface of the binocular vision sensor via a differential signal transmission line to open the shutter for exposure.

3. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, The binocular vision sensor includes two sets of symmetrically arranged high-speed industrial cameras, with the frame rate of a single industrial camera set to no less than 20,000 frames per second; the cameras are mounted on both sides below the print head, forming an intersection angle of 30 to 45 degrees; the image capture process is set within an observation window of 0.5 mm to 2.0 mm after the droplet leaves the nozzle. When extracting the three-dimensional coordinates of the centroid of flight, the original grayscale image is first denoised by median filtering. Then, the edge pixels of the droplet are searched and the surrounding pixels of each edge pixel are calculated. or The gray-level gradient distribution in the neighborhood is obtained by constructing a two-dimensional Gaussian function and using the least squares method to fit the surface, thereby obtaining the sub-pixel two-dimensional coordinate pairs of the left and right cameras. Finally, the three-dimensional coordinate sequence in physical space is calculated by using the pre-calibrated binocular disparity mapping matrix and the principle of triangular reconstruction.

4. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, The process of acquiring environmental perception factors is as follows: a high-sensitivity electrometer monitors the spatial gradient distribution of the electric field intensity between the nozzle and the printing substrate at a sampling frequency of 10 kHz, and obtains information on the nonlinear deflection force generated by the electrostatic field on the charged droplet; a triaxial accelerometer is installed at the end of the scanning axis of the motion platform to capture the mechanical vibration spectrum characteristics generated by the motion platform in high-speed reciprocating motion in real time. The spectrum characteristics include mid-to-high frequency fluctuations from 50 Hz to 500 Hz generated by resonance; the motion platform state vector includes the current command position, real-time feedback velocity, and acceleration planning parameters.

5. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, The GRU dynamic coupling prediction model specifically includes: a GRU layer, which is configured by resetting the gate. and Update Gate Extract time-series features of the flight trajectory to generate intermediate position prediction values; among which, To reset the gate output, To update the gate output, It is the sigmoid activation function. and This is the weight matrix. This is the hidden state from the previous moment. The input vector at the current time step, This represents a vector concatenation operation; Physical constraint layer, with built-in fluid dynamics equations: in, For the mass of the droplet, For acceleration, The effective charge of the droplet, as determined experimentally beforehand. For electric field strength, Aerodynamic viscosity, Where is the droplet radius, For air movement speed, The droplet velocity is used; the physical constraint layer receives the intermediate position prediction value output by the GRU layer and calculates the Coulomb force based on the real-time electrostatic field strength. And through the kinematic equations: To verify the rationality of the trajectory, among which, for The droplet position vector at time t, The initial position vector, The initial velocity vector, It is the acceleration vector. For time, abnormal predictions that do not conform to fluid dynamics characteristics are eliminated; The output layer generates residual correction values ​​that have been verified by physical laws. This is represented by the offset vector components of the droplet relative to the theoretically designed landing point in the lateral and longitudinal axes.

6. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, The specific process of the hard real-time adder superimposing instructions is as follows: The hard real-time adder is deployed in the pulse generation core module of the motion control system, independent of the operating system software layer; when the residual correction value is received, the hard real-time adder directly adds the residual correction value to the interpolation instruction value of the preset trajectory under the trigger of the rising edge of the clock. The hard real-time adder adjusts the pulse output frequency or the number of pulses in real time by modifying the values ​​in the FPGA pulse generator register, so that the motion platform completes physical displacement within a preset response period. The response period is set to be less than or equal to the flight time of the droplet from the nozzle to the printing substrate, ensuring that physical compensation for trajectory yaw is completed before the droplet contacts the substrate.

7. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, The dynamic execution process of displacement compensation is as follows: while maintaining the continuity of the original lateral and longitudinal scanning motion trajectory, the motion platform executes motion superposition commands through the motion control card; the hard real-time adder superimposes an auxiliary velocity vector into the original velocity command stream, which is calculated by dividing the residual correction value by the expected remaining flight time; the motion platform moves according to the superimposed composite velocity vector, and completes path offset compensation at a preset accuracy level to avoid the impact and vibration caused by frequent start and stop of the mechanical mechanism, and maintain the dynamic stability of the droplet landing process.

8. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, It also includes a closed-loop correction step in the height direction: the flatness of the printing substrate surface is detected in real time using a laser displacement meter, with a sampling frequency configured to be no less than 10kHz; the laser displacement meter feeds back the detected substrate height fluctuation data to the vertical control loop of the motion control system, driving the vertical motion axis in real time to correct the absolute height of the nozzle relative to the substrate surface, maintaining the consistency of droplet flight height; the compensation of the vertical control loop eliminates the droplet flight time variation caused by macroscopic topological undulations of the substrate surface, providing a constant time series reference benchmark for the GRU dynamic coupling prediction model.

9. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, It also includes a system self-healing monitoring step: real-time monitoring of the magnitude of the residual correction value, wherein the magnitude is expressed as the square root of the sum of the squares of the lateral axis deviation component and the longitudinal axis deviation component; when the magnitude is continuously... If the nozzle is found to be in an abnormal working state when the number of printing pulses exceeds a preset safety threshold, then... Based on the injection frequency f, set as follows: (Round up), and The range of values ​​is 500, the abnormal working state includes physical faults caused by nozzle scaling, ink dripping, or blockage; after determining that an abnormal working state has been established, the system automatically suspends the current printing task and drives the motion platform to the maintenance station to trigger the nozzle cleaning program; the nozzle cleaning program includes automatic vacuum ink suction operation and solvent rinsing operation, and after cleaning is completed, the system restarts the calibration process.

10. The nanoscale droplet impact printing control method based on industrial vision AI compensation according to claim 1, characterized in that, The hardware triggering circuit, the hard real-time adder, and the FPGA pulse generator register form a hardware-level closed loop: the value change written directly by the hard real-time adder to the FPGA pulse generator register takes effect in the next clock cycle, and the displacement compensation delay of the motion platform depends only on the propagation delay of the hardware logic gates and does not depend on software scheduling.

Images