A method and system for processing images with multiple layers of thickness in UV printing

Abstract

This invention discloses a method and system for processing multi-layer thickness images in UV printing. The method includes acquiring the target thickness intention, outputting a three-dimensional joint instruction tensor of droplet volume, droplet temperature, and UV light intensity via a generative adversarial network (GAN), and then performing multi-physics collaborative printing based on this tensor. At the pixel level, the method independently controls the droplet temperature of each nozzle and the UV light intensity of each pixel, achieving differentiated modulation such as high stacking, tiling, and edge barrier walls. An error map is obtained by measuring the actual thickness in situ and calculating the expected thickness using a differentiable physical proxy model. This error map is then processed through a three-layer closed loop to achieve real-time correction, GAN update, and fine-tuning of the tactile mapping model. This invention significantly improves thickness control accuracy, substrate adaptability, and printing efficiency, and can be applied to various scenarios such as 3D relief, micro-optical components, functionally graded materials, and 4D self-folding.

Description

Technical Field

[0001] This invention belongs to the field of UV inkjet printing and additive manufacturing technology, and particularly relates to a method and system for processing multi-layer thickness images in UV printing. Background Technology

[0002] UV printing technology uses a printhead to spray UV-curable ink onto a substrate surface, which is then immediately cured by UV light, enabling high-precision, multi-material image printing. In applications requiring tactile feedback or functional thickness (such as Braille, embossing, microlens arrays, and tactile panels), multi-layer stacking printing is often used to create thickness gradients. Existing technologies offer various multi-layer thickness control schemes: for example, analyzing the position of air bubbles after the previous layer is printed and reconstructing the printing path; predicting thickness based on an improved height profile model of the height difference between adjacent ink droplets; inserting blank layers in multi-layer printing to control interlayer spacing; or printing in reverse order from high to low cross-sections to eliminate the "terracing" effect. Furthermore, some schemes use multi-wavelength or variable-angle UV LED light to modulate the surface geometry of the photosensitive printing plate. These technologies have made some progress in improving the thickness controllability and surface quality of UV printing, but several limitations remain.

[0003] The common shortcomings of existing technologies are: thickness control mainly relies on droplet volume gradation or the accumulation of printed layers, lacking an active compensation mechanism for nonlinear effects during ink curing (such as diffusion flow, volume shrinkage, and interlayer solubility), resulting in a large deviation between the actual thickness and the design value. Furthermore, traditional methods typically use UV curing lamps with constant power and room temperature ink, making it impossible to differentiate the modulation of different areas within a single printed layer, and difficult to balance high thickness deposition with steep edges. In addition, existing solutions lack real-time thickness monitoring and closed-loop feedback during the printing process, allowing interlayer errors to accumulate layer by layer without correction; they also lack adaptability to different substrates (such as porous materials like textiles) or complex applications requiring functional gradients (such as hard-soft transitions and 4D self-folding). Users often rely on grayscale images or specialized software to design thickness, resulting in a high barrier to entry. Therefore, there is an urgent need for a high-precision UV printing multilayer thickness control solution capable of pixel-level multi-physics field collaborative modulation, nonlinear compensation, and self-learning capabilities. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a UV printing multilayer thickness image processing method and system, which solves the problems of lack of active compensation for ink nonlinear effects and differential modulation within a single layer, lack of real-time closed-loop feedback, and insufficient thickness control accuracy and substrate adaptability in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method and system for processing multi-layer thickness images in UV printing, comprising the following steps: S1: Obtain or generate the target thickness intent; S2: Input the target thickness intention and material parameter vector into the generator in the generative adversarial network, and the generator outputs a three-dimensional joint instruction tensor; wherein, the three-dimensional joint instruction tensor includes an ink droplet volume distribution map, an ink droplet target temperature distribution map, and a UV light intensity mask map; S3: Subsequently, multi-physics collaborative printing is performed according to the three-dimensional joint instruction tensor: the nozzle is controlled to eject ink droplets of corresponding volume according to the ink droplet volume distribution map, the temperature of the ink in each nozzle is controlled according to the ink droplet target temperature distribution map, and the spatial light modulator is controlled to selectively solidify the settled ink droplets according to the UV light intensity mask map; wherein, the ink droplet temperature and UV light intensity are co-modulated at the pixel level to form a target thickness gradient within a single printing layer; S4: Then, measure the actual thickness distribution of the printed layer and calculate the error map between the actual thickness and the expected thickness, which is calculated by substituting the three-dimensional joint instruction tensor into the differentiable physical proxy model; S5: Finally, the error maps are fed back to the three closed loops respectively: S51: First closed loop: Real-time correction of the three-dimensional joint instruction tensor of the unprinted area of ​​the current layer based on the error map; S52: Second closed loop: The error map and the corresponding three-dimensional joint instruction tensor and actual thickness data are used as training samples to update the generative adversarial network; S53: Third closed loop: Compare the error map with the target thickness intention, and adjust the model used to generate the target thickness intention.

[0006] Preferably, the method for obtaining or generating the target thickness intention in S1 includes: acquiring and mapping the tactile feature vector generated when the user operates the pressure-sensitive stylus; the tactile feature vector includes one or more of the following: vertical pressure of the pen tip, pen movement speed, cumulative dwell time at the same position, pressure change rate, pen body tilt angle, and pen body azimuth angle.

[0007] Preferably, the generative adversarial network in S2 includes a discriminator, which embeds a differentiable physical proxy model. This model is used to predict the thickness of the ink droplet after curing based on the droplet volume, droplet temperature, UV light intensity, and material parameters. The loss function of the discriminator includes adversarial loss and physical consistency loss based on the difference between the predicted thickness and the actual thickness.

[0008] Preferably, the coordinated modulation of droplet temperature and UV light intensity in S3 includes at least one of the following strategies: For regions where the target thickness is higher than the first thickness threshold, an ink droplet temperature lower than the first temperature threshold and a UV light intensity lower than the first light intensity threshold are used. For regions where the target thickness is below the second thickness threshold, an ink droplet temperature higher than the second temperature threshold and a UV light intensity higher than the second light intensity threshold are used. For edge areas with existing raised structures, a preset medium droplet temperature and UV light intensity higher than the rated value are used to form a curing barrier wall.

[0009] Preferably, the correction delay of the first closed loop in S51 is less than or equal to 5 milliseconds and is completed within the same scan line; the second closed loop in S52 is triggered after completing a preset number of layers or accumulating a preset number of pixel data, and only some parameter layers of the generative adversarial network are adjusted during the update; the third closed loop in S53 is triggered after accumulating a preset number of printing tasks, and the model for generating the target thickness intention is fine-tuned.

[0010] Preferably, the spatial light modulator in S3 is a Micro-LED array or a digital micromirror device; the resolution of the UV light intensity mask is the same as the printing resolution, and the light intensity and pulse width of each pixel are independently controllable; the peak wavelength of the UV light intensity is selected from at least one of 365nm or 395nm.

[0011] Preferably, in step S4, an optical non-contact displacement sensor is used to measure the actual thickness distribution. The sensor is installed on the printhead assembly and located after the curing unit. Its measurement accuracy is better than ±2 micrometers, and the lateral sampling interval matches the printing pixel interval. The method further includes: when an area with an absolute value greater than a preset threshold and a continuous area exceeding a preset value appears in the error graph, it is judged as a printing abnormality and an alarm or pause action is executed.

[0012] Preferably, the method further includes a gesture recognition step: detecting at least one of the following gestures based on the motion parameters of the pen stroke, and modifying the target thickness intention accordingly: a quick back-and-forth smearing gesture, a circular rotation pen stroke gesture, and a double-tap gesture; wherein, the quick back-and-forth smearing gesture corresponds to performing Gaussian blur, the circular rotation pen stroke gesture corresponds to generating a Gaussian-shaped protrusion at the center of the circle, and the double-tap gesture corresponds to adding a preset fixed height at the current position.

[0013] Preferably, a UV printing multi-layer thickness image processing system includes: The target thickness intent acquisition module is used to acquire or generate the target thickness intent and output the target thickness intent to the multiphysics GAN inverse compensation module; The multiphysics GAN inverse compensation module includes a generative adversarial network, which receives the target thickness intention and the input material parameter vector, and outputs a three-dimensional joint instruction tensor to the multiphysics collaborative printing execution module based on the target thickness intention and the material parameter vector; the three-dimensional joint instruction tensor includes an ink droplet volume distribution map, an ink droplet target temperature distribution map and a UV light intensity mask map. The multiphysics collaborative printing execution module includes a piezoelectric printhead array, a thermoelectric temperature control device for controlling the temperature of ink droplets, and a spatial light modulator for selective curing; the multiphysics collaborative printing execution module coordinates the control of ink droplet volume, ink droplet temperature and UV light intensity at the pixel level according to the received three-dimensional joint instruction tensor.

[0014] The in-situ thickness monitoring and feedback module is used to measure the actual thickness distribution of the printed layer, calculate the error map between the actual thickness and the expected thickness, and output the error map to the three-layer closed-loop self-learning controller. A three-layer closed-loop self-learning controller is used to receive the error map and feed it back to: a first closed loop, which corrects the 3D joint instruction tensor of the unprinted area of ​​the current layer in real time based on the error map and sends the corrected instruction to the multiphysics collaborative printing execution module; a second closed loop, which uses the error map, the corresponding 3D joint instruction tensor, and the actual thickness data as training samples to update the generative adversarial network; and a third closed loop, which compares the error map with the target thickness intent and adjusts the mapping model in the target thickness intent acquisition module.

[0015] Preferably, the target thickness intent acquisition module includes a pressure-sensitive stylus and a neural network for mapping tactile features to thickness intent; the thermoelectric temperature control device is a thermoelectric temperature control unit independently configured for each nozzle, the thermoelectric temperature control unit being a Peltier element or a resistance heating element; the spatial light modulator is a Micro-LED array or a digital micromirror device; the in-situ thickness monitoring and feedback module employs a line-scan spectral confocal displacement sensor or a laser triangulation displacement sensor; the system also includes an anomaly detection unit and a cloud database, the anomaly detection unit being used to determine printing anomalies when the error exceeds a preset threshold and the continuous area exceeds a preset value, and the cloud database being used to store data for each printing task for offline training or federated learning updates.

[0016] The technical effects and advantages of the UV printing multi-layer thickness image processing method and system of the present invention are as follows: 1. This invention outputs a three-dimensional joint instruction tensor, including an ink droplet volume distribution map, an ink droplet target temperature distribution map, and a UV light intensity mask map, through a generative adversarial network. It also combines a differentiable physical proxy model to predict the curing thickness, thereby eliminating thickness deviations caused by nonlinear factors such as ink diffusion, curing shrinkage, and interlayer mutual solubility. This allows the relative error between the actual thickness and the target thickness to be controlled within 2%. 2. This invention independently controls the droplet temperature of each nozzle (thermodynamic viscosity regulation) and the UV light intensity of each pixel (selective curing rate regulation) within a single printing layer based on the three-dimensional instruction tensor. It can simultaneously meet the differentiated needs of high-thickness stacking, low-thickness tiling and edge barrier walls in different areas, breaking through the limitations of traditional methods that rely solely on droplet volume or the number of printing layers. 3. The discriminator of this invention has an embedded differentiable physical proxy model. By constraining the output of the generator through the physical consistency loss function, the system can automatically learn and compensate for the shrinkage rate differences of different inks (white ink, color ink, varnish), changes in environmental temperature and humidity, and the influence of substrate characteristics, without the need to manually adjust printing parameters for each material or environment. 4. This invention employs a three-layer closed-loop self-learning mechanism for continuous evolution. The first closed loop (millisecond level) corrects the 3D instructions for unprinted areas in real time within the scanned rows, eliminating instantaneous deviations. The second closed loop (inter-layer / inter-task) uses incremental updates of actual printing data to generate an adversarial network, gradually improving nonlinear prediction capabilities. The third closed loop (cross-user / cross-task) fine-tunes the mapping model from tactile feedback to thickness intent based on thickness errors, achieving personalized user adaptation. This three-timescale feedback architecture continuously improves system accuracy with repeated use, significantly enhancing long-term operational stability. 5. This invention, through independent control of ink droplet temperature (low temperature inhibits penetration, high temperature promotes spreading) and spatially variable UV light intensity (local high-intensity rapid curing of the barrier wall), can effectively adapt to porous substrates such as textiles and paper, as well as curved and flexible substrates, preventing ink smudging or running. Simultaneously, by utilizing temperature-light synergistic modulation, it can achieve a curing crosslinking density gradient (hard-soft transition) of the same material or utilize shrinkage stress to drive self-folding (4D printing), expanding the functional boundaries of UV printing. 6. This invention collects tactile feature vectors (pressure, speed, dwell time, tilt angle, etc.) through a pressure-sensitive stylus and automatically maps them to thickness intentions. Combined with gesture recognition (quick smearing to achieve Gaussian smoothing, circular rotation to generate bosses, double-tap to increase fixed height), non-professional users can intuitively complete 3D thickness design without operating grayscale images or professional software, greatly reducing the threshold for use. 7. The droplet temperature and UV light intensity control in the three-dimensional command of this invention can achieve the target thickness gradient with fewer printing layers (e.g., low temperature and low light intensity promote vertical stacking and reduce the number of reprints), while real-time correction avoids repeated printing or waste due to deviation. Compared with traditional constant parameter printing, it can save more than 30% of printing time and material consumption. 8. The in-situ thickness monitoring module of this invention calculates the error map in real time. When an abnormal area with an absolute value exceeding the threshold and a continuous area exceeding the standard appears, it can distinguish the fault types such as nozzle blockage, insufficient ink or substrate contamination, and automatically alarm or suspend printing to reduce batch waste loss. 9. The invention system can upload the feature data (tactile vector, three-dimensional instructions, material parameters, actual thickness, error map) of each printing task to the cloud database. Through joint training of big data from multiple users, multiple materials, and multiple environments, it continuously optimizes and generates global parameters of adversarial networks and tactile mapping models, and then distributes them to various terminals to achieve collective intelligent evolution. Attached Figure Description

[0017] Figure 1 This is a flowchart of a UV printing multi-layer thickness image processing method and system proposed in this invention; Figure 2 This is a system framework diagram of a UV printing multi-layer thickness image processing method and system proposed in this invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0020] refer to Figures 1-2This invention provides a method and system for processing multi-layer thickness images in UV printing. The method includes: acquiring or generating a target thickness intention; inputting the target thickness intention and a material parameter vector into a generator of a generative adversarial network (GAN); outputting a three-dimensional joint instruction tensor containing an ink droplet volume distribution map, an ink droplet target temperature distribution map, and a UV intensity mask map; subsequently, performing multi-physics collaborative printing based on this three-dimensional joint instruction tensor; controlling the nozzles to eject ink droplets of corresponding volumes according to the ink droplet volume distribution map; independently controlling the ink temperature in each nozzle according to the ink droplet target temperature distribution map; and controlling a spatial light modulator to selectively solidify the settled ink droplets according to the UV intensity mask map, wherein the ink droplet temperature and UV intensity are correlated. This invention employs pixel-level collaborative modulation to create a target thickness gradient within a single printed layer. Then, it measures the actual thickness distribution of the printed layer and calculates an error map between this map and the expected thickness, which is derived from a 3D joint instruction tensor substituted into a differentiable physical proxy model. Finally, the error map is fed back to three closed loops: the first loop continuously corrects the 3D joint instruction tensor of the unprinted areas in the current layer; the second loop updates the generative adversarial network (GAN) using the error map, the corresponding 3D instruction tensor, and the actual thickness data; and the third loop compares the error map with the target thickness intent and adjusts the mapping model used to generate the target thickness intent. The system includes a target thickness intent acquisition module, a multiphysics GAN inverse compensation module, a multiphysics collaborative printing execution module, an in-situ thickness monitoring and feedback module, and a three-layer closed-loop self-learning controller. This invention achieves pixel-level multiphysics collaborative modulation, nonlinear compensation, and self-learning, significantly improving thickness control accuracy, material adaptability, and printing efficiency.

[0021] System overall architecture: The system of this invention comprises the following five core modules: Target thickness intent acquisition module: This module acquires or generates the target thickness intent and outputs it to the multiphysics GAN inverse compensation module. In one specific implementation, this module includes a high-precision pressure-sensitive touch panel (sampling rate not less than 200Hz), an active capacitive stylus, and a three-layer convolutional neural network for mapping tactile feature vectors to thickness intents. This module can also accept grayscale images imported from external sources as thickness intents.

[0022] The multiphysics GAN inverse compensation module includes a generative adversarial network (GAN) that receives the target thickness intention and the input material parameter vector. Based on these, it outputs a three-dimensional joint instruction tensor to the multiphysics collaborative printing execution module. This three-dimensional joint instruction tensor includes an ink droplet volume distribution map, an ink droplet target temperature distribution map, and a UV light intensity mask map. The GAN's discriminator embeds a differentiable physical proxy model, which is constructed based on the lubrication equation and the curing kinetics equation.

[0023] The multiphysics collaborative printing execution module includes a piezoelectric printhead array, a thermoelectric temperature controller for controlling droplet temperature, and a spatial light modulator for selective curing. This module collaboratively controls droplet volume, droplet temperature, and UV light intensity at the pixel level based on the received three-dimensional joint command tensor. The thermoelectric temperature controller can be a Peltier element or a resistance heating element configured independently for each nozzle; the spatial light modulator can be a Micro-LED array or a digital micromirror device.

[0024] In-situ thickness monitoring and feedback module: This module measures the actual thickness distribution of the printed layer, calculates the error map between the actual thickness and the expected thickness, and outputs the error map to the three-layer closed-loop self-learning controller. This module uses an optical non-contact displacement sensor (e.g., a line-scan spectral confocal displacement sensor or a laser triangulation displacement sensor), mounted on the printhead assembly and located after the curing unit.

[0025] A three-layer closed-loop self-learning controller receives the error map and feeds it back to the first, second, and third closed loops. The first closed loop corrects the 3D joint instruction tensor of the unprinted region in the current layer in real time based on the error map and sends the corrected instructions to the multiphysics collaborative printing execution module. The second closed loop uses the error map, the corresponding 3D joint instruction tensor, and the actual thickness data as training samples to update the generative adversarial network. The third closed loop compares the error map with the target thickness intent and adjusts the mapping model in the target thickness intent acquisition module.

[0026] Example 1: Objective: To print the raised letter "A" on an acrylic sheet (80mm×80mm), requiring a maximum thickness of 1.2mm, an edge bevel angle ≥85°, and a bottom surface smoothness Ra≤2μm. This example verifies the invention's ability to control high-thickness stacking and edge steepness.

[0027] The implementation system adopts the complete system of this invention: Target thickness intent acquisition module: using Wacom Intuos Pro pressure-sensitive touch tablet (8192 levels of pressure sensitivity, 250Hz sampling rate) and active capacitive stylus (built-in accelerometer and gyroscope), as well as a three-layer convolutional neural network (first layer 5×5 kernel, 16 channels; second layer 3×3, 32 channels; third layer 1×1, 1 channel).

[0028] Multiphysics GAN inverse compensation module: Deployed on the NVIDIA Jetson AGX Orin platform. The generator is a U-Net structure, taking as input the thickness intention (600×600, single channel) and material parameter vectors (viscosity, shrinkage rate, ambient temperature and humidity), and outputting three channels (droplet volume, droplet temperature, UV light intensity). The discriminator embeds a differentiable physical surrogate model (based on lubrication equations and curing kinetics).

[0029] Multi-physics collaborative printing execution module: piezoelectric printhead (Ricoh GEN6, 1280 nozzle, 300npi), each nozzle is independently equipped with a Peltier thermoelectric temperature control unit (accuracy ±0.5℃, response ≤10ms); spatial light modulator is a Micro-LED array (365nm, pixel pitch 42.3μm, 12-bit PWM dimming); synchronization controller is based on FPGA.

[0030] In-situ thickness monitoring and feedback module: Line scan spectral confocal displacement sensor (Precitec CHRocodile M4S, accuracy ±0.5μm, scan line frequency 10kHz), installed on the side of the printhead, after the curing unit.

[0031] Three-layer closed-loop self-learning controller: integrated in an industrial control computer, the first closed loop is implemented in the FPGA (delay <1ms), and the second and third closed loops are executed asynchronously in a Python program.

[0032] Implementation steps: S1: Obtain or generate the target thickness intent.

[0033] The user writes the letter "A" on the tablet using a pressure-sensitive stylus, applying heavy pressure (3500 / 4095) at the intersection of strokes, with a slow pen speed (<10mm / s). The system collects tactile feature vectors (vertical pressure of the pen tip, pen movement speed, cumulative dwell time at the same position, pressure change rate, pen tilt angle, and pen azimuth angle), and maps them through a three-layer convolutional neural network to obtain an initial thickness intention. This intention has a resolution of 600dpi, with the letter body thickness ranging from 0.8 to 1.2mm, the edge slope width of 0.3mm, and the background thickness of 0mm.

[0034] S2: Input the target thickness intention and material parameter vector into the generator in the generative adversarial network, and the generator outputs a three-dimensional joint instruction tensor.

[0035] The material parameter vector is set as follows: transparent UV varnish, viscosity 12 mPa·s (25℃), shrinkage rate 3.2%, curing energy threshold 150 mJ / cm², ambient temperature 25℃, humidity 50%. The generator outputs a three-dimensional joint instruction tensor, which includes a droplet volume distribution map, a droplet target temperature distribution map, and a UV light intensity mask map. Typical region parameters are as follows: Letter center area: droplet volume 22 pL, droplet temperature 25℃, UV light intensity 25%; Edge transition zone: droplet volume 14 pL, droplet temperature 38℃, UV light intensity 60%; Background area: droplet volume 5 pL, droplet temperature 55℃, UV light intensity 95%; Letter outer edge (forming a solidified barrier wall): ink droplet volume 12pL, ink droplet temperature 45℃, UV light intensity 120% (pulse width 0.2ms).

[0036] S3: Perform multiphysics collaborative printing based on the 3D joint instruction tensor.

[0037] The printing speed is set to 300 mm / s, requiring 8 layers (each approximately 0.15 mm thick). During printing, based on the aforementioned three-dimensional joint instruction tensor: the nozzles are controlled to eject ink droplets of corresponding volumes according to the droplet volume distribution map; the ink temperature within each nozzle is controlled according to the droplet target temperature distribution map (via a Peltier element); and the Micro-LED array is controlled to selectively cure the settled ink droplets according to the UV intensity mask map. Droplet temperature and UV intensity are synergistically modulated at the pixel level: in areas requiring high-thickness deposition (the center of the letter), low temperature (25°C) and low light intensity (25%) are used to delay droplet curing, maintain high viscosity, and promote vertical deposition; in edge areas, medium temperature (45°C) and ultra-high light intensity (120%) are used to form a rapid-curing "barrier wall" to prevent flow; in background areas, high temperature (55°C) and high light intensity (95%) are used to quickly flatten and lock the ink droplets.

[0038] S4: Measure the actual thickness distribution of the printed layer and calculate the error graph between the actual thickness and the expected thickness.

[0039] After each layer is printed, a line-scan spectral confocal displacement sensor immediately measures the actual thickness distribution of that layer, with a lateral sampling interval of 42.3 μm (matching the printed pixels). The expected thickness is calculated by substituting the aforementioned 3D joint instruction tensor into a differentiable physical proxy model. The pixel-by-pixel error E(x,y) = expected thickness - actual thickness is calculated.

[0040] S5: Feed the error map back to the three closed loops respectively.

[0041] S51 (First Closed Loop): Based on the error map, the 3D joint instruction tensor of the unprinted areas in the current layer is corrected in real time. After the first layer is printed, if the sensor measures that the actual thickness at the letter starting point is 15μm lower than expected (error of +15μm), the first closed loop immediately corrects the unprinted areas downstream in the same layer: droplet volume increases by 5%, droplet temperature decreases by 3℃, and UV light intensity decreases by 10%. The correction delay is less than 1ms, completed within the same scan line. Deviation elimination begins with the second layer.

[0042] S52 (Second Loop): The error map, along with the corresponding 3D joint instruction tensor and actual thickness data, are used as training samples to update the generative adversarial network. An incremental update is triggered every 5 layers of printing or accumulating 1000 pixels of data. Only the convolutional parameters of the last two layers of the generator and the fully connected parameters of the last two layers of the discriminator are adjusted, and the learning rate is reduced to one-tenth of the initial pre-training learning rate.

[0043] S53 (Third Loop): The error map is compared with the intended target thickness, and the model used to generate the intended target thickness (i.e., the neural network mapping tactile features to the thickness map) is adjusted. In this embodiment, since this is the first time the system is used, the third loop is triggered after accumulating 50 printing tasks, and the mean squared error loss function is used to fine-tune all parameters of the mapping model.

[0044] Implementation Results: The final printed letter "A" had a maximum actual thickness of 1.19 mm (target 1.2 mm, relative error 0.8%), an edge bevel angle of 87° (target ≥ 85°), and a bottom surface roughness Ra = 1.7 μm (target ≤ 2 μm). Real-time correction in the first closed loop improved interlayer thickness consistency to within ±5 μm. Compared to traditional methods with constant light intensity and no temperature control (see Comparative Example 1), this embodiment improved thickness accuracy by approximately one order of magnitude, significantly improved edge steepness, and reduced the number of printing layers (8 layers vs. 12 layers in Comparative Example), resulting in a 33% increase in efficiency.

[0045] Example 2: Objective: To print a microlens array (10×10 array) with a diameter of 200 μm and a focal length of 2 mm on a glass substrate, requiring a surface profile close to a sphere and a roughness Ra≤0.5 μm. This verifies the invention's ability to precisely control micron-level thickness and fit curved surface profiles.

[0046] The implementation system is the same as in Example 1, but the target thickness intention acquisition module does not use tactile input; instead, it directly imports an external grayscale image. The material parameters in the multiphysics GAN inverse compensation module are set to low-shrinkage transparent UV varnish (shrinkage rate 1.5%). The remaining modules are the same.

[0047] Implementation steps: S1: Obtain or generate the target thickness intent.

[0048] Because the lens array requires extremely high surface precision, this example does not use tactile input, but instead directly imports a grayscale image generated by optical design software (Zemax). Grayscale values ​​of 0-255 correspond to thicknesses of 0-50 μm. This grayscale image serves as a target thickness representation, with a resolution of 600 dpi, and each microlens profile is spherical.

[0049] S2: Input the target thickness intention and material parameter vector into the generator in the generative adversarial network, and the generator outputs a three-dimensional joint instruction tensor.

[0050] Material parameter vector: High-transparency UV varnish, viscosity 8 mPa·s (25℃), shrinkage rate 1.5%, ambient temperature 25℃. The generator output commands automatically adjust the temperature-light intensity combination for the spherical profile. Lens center (thickness 50μm): droplet volume 18pL, droplet temperature 30℃, UV light intensity 30% (delayed curing, allowing droplets to flow into a spherical surface). Lens edge (thickness 5μm): droplet volume 6pL, droplet temperature 50℃, UV light intensity 80%; Flat region between lenses: ink droplet volume 4 pL, ink droplet temperature 55℃, UV light intensity 95%.

[0051] S3: Perform multiphysics collaborative printing based on the 3D joint instruction tensor.

[0052] Due to the total thickness of only 50μm, single-layer printing was used in this example. The printing speed was 200mm / s. During the printing process, pixel-level collaborative control was performed based on the droplet volume distribution map, the droplet target temperature distribution map, and the UV light intensity mask map. The central area has a low droplet temperature and weak light intensity, allowing the droplets sufficient time to flow and form spheres under the drive of surface tension; the edge and background areas have high temperature and high light intensity, which allows for rapid curing to maintain the contour.

[0053] S4: Measure the actual thickness distribution of the printed layer and calculate the error graph between the actual thickness and the expected thickness.

[0054] The in-situ thickness monitoring module measures the actual thickness profile after single-layer printing in real time. The expected thickness is calculated using a differentiable physical surrogate model. The calculation error map shows that the actual profile is slightly flattened (the center height is about 3 μm lower than the target).

[0055] S5: Feed the error map back to the three closed loops respectively.

[0056] S51 (First Closed Loop): Due to single-layer printing, the first closed loop performs real-time corrections on the unprinted areas within the current layer (i.e., subsequent scan lines of the area already detected by the sensor). Based on the error map, the first closed loop reduces the UV intensity of the unprinted portion of the lens's central area from 30% to 25%, and the droplet temperature from 30°C to 28°C, allowing the droplets more time to flow and flatten. After correction, the RMS error between the actual contour and the designed sphere is reduced to 0.8μm.

[0057] S52 (Second Closed Loop): After this printing is completed, the error map, the corresponding 3D instruction tensor, and the actual thickness data are stored in the cache. Since the amount of data in this printout has not reached the trigger threshold (1000 pixels), an update will not be triggered at this time.

[0058] S53 (Third Closed Loop): This is the first use in this embodiment, and the third closed loop is not triggered.

[0059] The printed microlens array, tested using a laser confocal microscope, showed a measured focal length of 2.03 mm (design value 2 mm), a focused spot diameter of 6.2 μm (theoretical value 5.8 μm), and a surface roughness Ra = 0.42 μm. The RMS error between the contour and the spherical surface was 0.8 μm, meeting the requirements for optical applications. This demonstrates that the present invention can precisely control the micron-level thickness distribution, enabling the printing of complex curved surface contours.

[0060] Example 3: Purpose of the implementation: To print a brand logo (gradient thickness, 0.5mm at its thickest and 0.05mm at its thinnest) on the surface of a cotton T-shirt, requiring a smooth thickness transition and preventing ink from seeping into the fibers and causing smudging. This verifies the adaptability of the invention to porous, flexible substrates and the convenience of gesture interaction.

[0061] The implementation system is basically the same as in Example 1, but the piezoelectric printhead in the multiphysics collaborative printing execution module is adapted to textile ink (high viscosity). The pressure-sensitive stylus in the target thickness intent acquisition module adds gesture recognition functionality. The in-situ thickness monitoring module is adaptively adjusted for flexible substrates (adding substrate fixing clamps).

[0062] Implementation steps: S1: Obtain or generate the target thickness intent.

[0063] The user traces the logo outline with a pressure-sensitive stylus and then performs a "circular rotation" gesture in the center area of ​​the logo. The system recognizes the circular rotation gesture (trajectory radius 8mm, rotation angle greater than 360°) based on the stroke's motion parameters and generates a Gaussian-shaped boss at the center of the circle (peak 0.5mm, standard deviation 5mm). The modified thickness is intended to have a smooth, gradual thickness gradient.

[0064] S2: Input the target thickness intention and material parameter vector into the generator in the generative adversarial network, and the generator outputs a three-dimensional joint instruction tensor.

[0065] Material parameter vector: Textile-specific UV white ink, viscosity 25 mPa·s (25℃), shrinkage rate 4.5%, substrate is cotton fabric (porous). The generator output instructions automatically adjust the strategy for porous substrates: High-thickness area (logo center): ink droplet volume 25pL, ink droplet temperature 22℃ (low temperature and high viscosity), UV light intensity 20%, to prevent ink from penetrating the fibers with high viscosity; Medium thickness zone: droplet volume 15 pL, droplet temperature 35℃, UV light intensity 50%; Thin edge area: ink droplet volume 8pL, ink droplet temperature 52℃, UV light intensity 90%, rapid curing and sealing.

[0066] S3: Perform multiphysics collaborative printing based on the 3D joint instruction tensor.

[0067] Due to the uneven surface of the textile, the system first performs a substrate pre-scan (using an in-situ thickness monitoring module to scan the empty substrate) before executing S3, offsetting the intended thickness according to the actual fiber height. A total of 5 layers are printed. During the printing process, the ink droplet volume, ink droplet temperature, and UV light intensity are controlled collaboratively based on the three-dimensional instruction tensor. In low-temperature, low-light-intensity areas, the ink maintains high viscosity to reduce penetration into the fibers; in high-temperature, high-light-intensity areas, the ink cures rapidly to form a sealing layer.

[0068] S4: Measure the actual thickness distribution of the printed layer and calculate the error graph between the actual thickness and the expected thickness.

[0069] The actual thickness was measured using a spectral confocal sensor after each layer was printed. Due to the softness of the substrate, a low contact force mode was used during measurement. The error graph shows localized areas where insufficient thickness was caused by fiber indentation.

[0070] S5: Feed the error map back to the three closed loops respectively.

[0071] S51 (First Closed Loop): Based on the error map, corrective instructions are given for the unprinted areas of the current layer: increase the ink droplet volume by 8% and reduce the UV light intensity by 15% in the fiber recessed areas, so that more ink can fill the recesses.

[0072] S52 (Second Closed Loop): An update is triggered every 5 layers, using error data to fine-tune the GAN, gradually adapting it to the porous properties of textiles.

[0073] S53 (Third Closed Loop): After printing the same logo multiple times, the haptic mapping model is fine-tuned so that the peak value of the Gaussian boss generated by the user's circular rotation gesture more accurately corresponds to the actual thickness.

[0074] The printed logo exhibits a smooth thickness gradient, ranging from a maximum thickness of 0.48mm to a minimum of 0.05mm, with a uniform width in the transition zone. The edges are clear with no ink bleeding, and the logo has a distinct three-dimensional feel to the touch. In the wash fastness test (30 machine washes), the thickness retention rate is 92%, superior to traditional methods (typically <70%). Gesture interaction makes the design process intuitive and efficient.

[0075] Example 4: Purpose of implementation: To print a flexible sensor substrate with hard edges (Young's modulus > 100 MPa) for easy fixation and a soft center (Young's modulus < 10 MPa) for skin adhesion, with a uniform overall thickness of 0.3 mm. This verifies the invention's ability to achieve a modulus gradient by co-modulating ink droplet temperature and UV light intensity to create differences in cross-linking density in different regions of the same material.

[0076] The implementation system is similar to that of Example 1, but the input to the multiphysics GAN inverse compensation module adds a "modulus intention" (hard region, soft region, transition region). The material used is tunable modulus UV polyurethane acrylate. The remaining modules are the same.

[0077] Implementation steps: S1: Obtain or generate the target thickness intent.

[0078] The intended thickness is a constant value of 0.3 mm. Simultaneously, the user defines the intended modulus through a graphical interface: the hard region is an edge ring (5 mm wide), the soft region is a central circle (20 mm in diameter), and the transition region is 2 mm wide. The system converts the intended modulus into temperature-light intensity modulation rules and integrates them into the target thickness intended thickness (as auxiliary information).

[0079] S2: Input the target thickness intention and material parameter vector into the generator in the generative adversarial network, and the generator outputs a three-dimensional joint instruction tensor.

[0080] Material parameter vector: UV polyurethane acrylate, viscosity 15 mPa·s (25℃), shrinkage 3.0%, energy density required for complete curing 200 mJ / cm², energy density for partial curing 50 mJ / cm². Generator output command: Hard area (edge): Droplet volume 12pL, droplet temperature 50℃ (low viscosity, easy to spread quickly), UV light intensity 100% (complete cross-linking, high modulus). Soft zone (center): Droplet volume 12pL, droplet temperature 25℃ (high viscosity, reduce flow), UV light intensity 20% (partial curing, low modulus).

[0081] Transition zone: droplet volume 12pL, droplet temperature linearly decreases from 50℃ to 25℃, UV light intensity linearly decreases from 100% to 20%.

[0082] S3: Perform multiphysics collaborative printing based on the 3D joint instruction tensor.

[0083] Single-layer printing (0.3mm thickness). Printing speed: 150mm / s. During printing, droplet volume, droplet temperature, and UV light intensity are controlled collaboratively according to instructions. Since the droplet volume is uniform (12pL), thickness uniformity is maintained by temperature and light intensity. In high-temperature, high-light-intensity areas, the ink cures rapidly and becomes fully cross-linked; in low-temperature, low-light-intensity areas, curing is slow and the cross-linking density is low.

[0084] S4: Measure the actual thickness distribution of the printed layer and calculate the error graph between the actual thickness and the expected thickness.

[0085] In-situ monitoring revealed a fluctuation of ±0.02 mm in the actual thickness uniformity (mainly due to differences in ink droplet spreading at different temperatures). The error map shows that the hard area is slightly thinner (due to greater ink droplet spreading at high temperatures) and the soft area is slightly thicker.

[0086] S5: Feed the error map back to the three closed loops respectively.

[0087] S51 (First Closed Loop): Based on the error map, correct the remaining areas of the current layer: increase the droplet volume by 10% in hard areas (from 12 pL to 13.2 pL) to compensate for insufficient thickness caused by spreading; reduce the droplet volume by 5% in soft areas (to 11.4 pL). After correction, the final thickness uniformity reaches 0.3 ± 0.01 mm.

[0088] S52 (Second Closed Loop): The second closed loop is triggered after this printing, and the error data is used for GAN fine-tuning, so that it learns the relationship between the droplet spreading rate and the thickness compensation amount at different temperatures.

[0089] S53 (Third Closed Loop): Not triggered.

[0090] The printed substrate, tested using a universal testing machine, showed the following results: hard region Young's modulus 112 MPa, soft region modulus 8.5 MPa, and transition region width approximately 2.2 mm (close to the design value of 2 mm). Overall thickness uniformity was 0.3 ± 0.01 mm. This substrate, used for ECG electrode patches, provides secure edge fixation and conforms to skin deformation at the center, resulting in a 40% improvement in signal acquisition quality compared to homogeneous substrates.

[0091] Example 5: Purpose of the implementation: To print a planar sheet structure, which, after UV curing, automatically folds into a three-dimensional cubic box (10mm on each side) upon release from the substrate. This verifies the ability of this invention to achieve 4D printing (deformation) by pre-setting a stress gradient based on the difference in curing shrinkage rates in different regions.

[0092] The implementation system is basically the same as in Example 1, but the material used is a high-shrinkage UV varnish (6% shrinkage). The training data for the multiphysics GAN inverse compensation module incorporates the mapping relationship between shrinkage stress and folding angle. An in-situ thickness monitoring module is used to ensure consistent thickness.

[0093] Implementation steps: S1: Obtain or generate the target thickness intent.

[0094] Design the folding hinge area (0.2mm thick, 1mm wide) and the panel area (0.5mm thick). The thickness intention was exported as a grayscale image from CAD software and imported into the system.

[0095] S2: Input the target thickness intention and material parameter vector into the generator in the generative adversarial network, and the generator outputs a three-dimensional joint instruction tensor.

[0096] Material parameter vector: High shrinkage UV varnish, viscosity 10 mPa·s (25℃), shrinkage 6%. Generator output command: Hinge area: droplet volume 8pL (thin layer), droplet temperature 55℃, UV light intensity 100% (rapid and complete curing, large shrinkage); Panel area: Droplet volume 20pL (thick layer), droplet temperature 25℃, UV light intensity 20% (slow curing, low shrinkage); This difference creates tensile residual stress at the hinge.

[0097] S3: Perform multiphysics collaborative printing based on the 3D joint instruction tensor.

[0098] A total of three layers were printed. In the hinge area, all three layers were 0.2mm thick (cumulative 0.6mm, but reduced to 0.45mm after shrinkage); in the panel area, all three layers accumulated to 0.5mm (finally approximately 0.48mm due to less shrinkage). During printing, the ink droplet volume, temperature, and light intensity were coordinated and controlled according to instructions. The high temperature and high light intensity in the hinge area caused the ink to solidify rapidly and produce large shrinkage strain; the low temperature and low light intensity in the panel area resulted in slow curing and limited shrinkage.

[0099] S4: Measure the actual thickness distribution of the printed layer and calculate the error graph between the actual thickness and the expected thickness.

[0100] In-situ monitoring revealed good thickness consistency (deviation <0.02mm). The error graph showed no obvious abnormalities.

[0101] S5: Feed the error map back to the three closed loops respectively.

[0102] S51 (First Closed Loop): In this embodiment, the first closed loop is only used to maintain thickness consistency and no large deviation occurs; S52 (Second Closed Loop): Not triggered (data volume not reached the threshold); S53 (Third Closed Loop): Not triggered.

[0103] Additional post-processing steps (not mandatory but used in this embodiment): After printing, heat the substrate to 60°C and hold for 5 minutes to promote stress release, then cool to room temperature. Upon peeling from the substrate, the sheet automatically folds into a cubic box.

[0104] The final folded cubic box has a side length of 10.2mm (target 10mm), a folding angle deviation of <5°, and a box size tolerance of ±0.3mm. The hinge area thickness is 0.45mm, and the panel thickness is 0.48mm. This 4D printed structure requires no external trigger and folds itself solely using UV curing shrinkage stress. Traditional UV printing cannot achieve this function.

[0105] Comparative Example 1: The implementation objective is the same as in Example 1 (printing the letter "A" on an acrylic plate with a maximum thickness of 1.2 mm), but a traditional UV printing process is used (without gradient light intensity, without droplet temperature control, without GAN compensation, and without closed-loop feedback) to illustrate the technological advancements of the present invention.

[0106] The implementation system, a traditional UV flatbed printer (e.g., a certain commercial model), is configured as follows: Piezoelectric printhead (no independent temperature control, ink room temperature 25℃); UV mercury lamp (constant power 100%, 6W / cm²). No in-situ thickness monitoring; No closed-loop feedback; No GAN compensation; The grayscale image to layer mapping uses a fixed algorithm (one layer every 0.1mm).

[0107] Implementation steps: S1 (Analogy): The operator imports a grayscale image of the letter "A" into the software. Grayscale values ​​from 0 to 255 correspond to thicknesses from 0 to 1.5 mm. Based on a maximum thickness of 1.2 mm, the software calculates that 12 layers are needed for printing. S2 (Analog): No GAN generator. Fixed printing parameters for each layer: droplet volume is divided into three levels based on grayscale (10pL, 15pL, 20pL), droplet temperature is room temperature (25℃), and UV light intensity is constant at 100%. S3 (Analog): The printhead scans at 300mm / s, and each layer is cured immediately after ink ejection by a full-power UV lamp. There is no droplet temperature control or variable light intensity. A total of 12 layers are printed. S4 (Analog): No in-situ thickness monitoring; thickness is measured using vernier calipers and an optical microscope after printing. S5 (Analog): No closed-loop feedback.

[0108] Implementation results and actual printing output: The maximum thickness is only 0.85mm (due to ink curing shrinkage and flow, the target of 1.2mm was not achieved). The edges of the letters show a severe "pancake" phenomenon (ink flows beyond the boundary), and the edge slope angle is only 45°; Obvious air bubbles appeared between the layers (due to the rapid curing and encapsulation failing to expel air); The bottom surface is uneven, with local thickness deviations of ±0.15mm; The total time taken was 50% longer than in Example 1 (12 layers vs. 8 layers, and no adaptive layering).

[0109] Compared to Examples 1-5 and Comparative Example 1, Examples 1-5 of this invention verified the technical effects of the invention in different application scenarios, while Comparative Example 1 used a traditional printing method with constant light intensity, no droplet temperature control, no GAN compensation, and no closed-loop feedback as a baseline comparison. Regarding thickness control accuracy, the traditional method (Comparative Example 1) resulted in a letter thickness of only 0.85mm for the target 1.2mm thickness due to ink curing shrinkage accumulation and flow, with a relative error as high as 29%, an edge bevel angle of only 45°, and interlayer bubbles and a local thickness deviation of ±0.15mm. In contrast, Example 1 (acrylic high-relief), through the coordinated modulation of gradient light intensity and droplet temperature, used low temperature and low light intensity to delay curing in the high-thickness area, and used ultra-high light intensity to form a barrier wall at the edge. Combined with GAN nonlinear compensation and real-time correction of the first closed loop, the actual thickness was 1.19mm (error 0.8%), the edge bevel angle was 87°, the bottom surface roughness Ra=1.7μm, and the number of printing layers was reduced by 33% compared to the comparative example. Example 2 (microlens array) achieved precise replication of micron-level spherical contours with an RMS error of only 0.8 μm; Example 3 (textile logo) successfully suppressed ink penetration on porous substrates, achieved smooth thickness gradients, and maintained a 92% water wash retention rate; Example 4 utilized temperature and light synergy to achieve modulus gradients (112 MPa in hard areas and 8.5 MPa in soft areas) on the same material; Example 5 innovatively utilized the difference in curing shrinkage stress to achieve a 4D self-folding cubic box. The common features of the above examples are: they all employ three-dimensional joint instruction tensors (droplet volume, temperature, light intensity) and multi-physics field collaborative printing, as well as a three-layer self-learning architecture consisting of S51 first closed-loop real-time correction, S52 second closed-loop GAN incremental update, and S53 third closed-loop tactile mapping fine-tuning.

[0110] In terms of printing efficiency and material utilization, the traditional method (Comparative Example 1) requires 12 layers to barely achieve a thickness of 0.85 mm due to the lack of adaptive layering and real-time correction. Ink flow causes material waste of about 40%, and the lack of in-situ monitoring leads to a high scrap rate. In contrast, Example 1 of the present invention achieves a thickness of 1.19 mm with only 8 layers, improving material utilization by about 35%. Example 2 completes a 50 μm thick microlens array with a single layer printing, eliminating the need for multiple overprints. Example 3 avoids the problem of excessive first-layer penetration common in textiles by pre-scanning the substrate and first closed-loop compensation. Example 4 simplifies the printing process by using temperature-light intensity joint control to produce the same thickness but different moduli in different areas with the same ink droplet volume (12 pL). Example 5 utilizes shrinkage stress to drive self-folding, eliminating the need for shape memory materials or external triggers required by traditional 4D printing. Furthermore, the three-layer closed-loop self-learning function (S52, S53) of this invention enables the system to continuously optimize the GAN generator and tactile mapping model based on the actual error of each print. After multiple tasks, the thickness accuracy can be further improved to within ±1%, while Comparative Example 1 has no learning ability, and the same error will be repeated. In summary, the technical solution of this invention is significantly superior to the traditional constant parameter printing method in terms of thickness control accuracy, substrate adaptability, functional diversity, and efficiency, realizing a leap from "open-loop stacking" to "intelligent closed-loop control".

[0111] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0112] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. 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 processing multi-layer thickness images in UV printing, characterized in that, Includes the following steps: S1: Obtain or generate the target thickness intent; S2: Input the target thickness intention and material parameter vector into the generator in the generative adversarial network, and the generator outputs a three-dimensional joint instruction tensor; wherein, the three-dimensional joint instruction tensor includes an ink droplet volume distribution map, an ink droplet target temperature distribution map, and a UV light intensity mask map; S3: Subsequently, multi-physics collaborative printing is performed according to the three-dimensional joint instruction tensor: the nozzle is controlled to eject ink droplets of corresponding volume according to the ink droplet volume distribution map, the temperature of the ink in each nozzle is controlled according to the ink droplet target temperature distribution map, and the spatial light modulator is controlled to selectively solidify the settled ink droplets according to the UV light intensity mask map; wherein, the ink droplet temperature and UV light intensity are co-modulated at the pixel level to form a target thickness gradient within a single printing layer; S4: Then, measure the actual thickness distribution of the printed layer and calculate the error map between the actual thickness and the expected thickness, which is calculated by substituting the three-dimensional joint instruction tensor into the differentiable physical proxy model; S5: Finally, the error maps are fed back to the three closed loops respectively: S51: First closed loop: Real-time correction of the three-dimensional joint instruction tensor of the unprinted area of ​​the current layer based on the error map; S52: Second closed loop: The error map and the corresponding three-dimensional joint instruction tensor and actual thickness data are used as training samples to update the generative adversarial network; S53: Third closed loop: Compare the error map with the target thickness intention, and adjust the model used to generate the target thickness intention.

2. The UV printing multi-layer thickness image processing method as described in claim 1, characterized in that, The method of obtaining or generating the target thickness intention in S1 includes: acquiring and mapping the tactile feature vector generated when the user operates the pressure-sensitive stylus; the tactile feature vector includes one or more of the following: vertical pressure of the pen tip, pen movement speed, cumulative dwell time at the same position, pressure change rate, pen body tilt angle, and pen body azimuth angle.

3. The UV printing multi-layer thickness image processing method as described in claim 1, characterized in that, The generative adversarial network in S2 includes a discriminator, which embeds a differentiable physical proxy model. This model is used to predict the thickness of the ink droplet after curing based on the droplet volume, droplet temperature, UV light intensity, and material parameters. The loss function of the discriminator includes adversarial loss and physical consistency loss based on the difference between the predicted thickness and the actual thickness.

4. The UV printing multi-layer thickness image processing method as described in claim 1, characterized in that, The coordinated modulation of droplet temperature and UV light intensity in S3 includes at least one of the following strategies: For regions where the target thickness is higher than the first thickness threshold, an ink droplet temperature lower than the first temperature threshold and a UV light intensity lower than the first light intensity threshold are used. For regions where the target thickness is below the second thickness threshold, an ink droplet temperature higher than the second temperature threshold and a UV light intensity higher than the second light intensity threshold are used. For edge areas with existing raised structures, a preset medium droplet temperature and UV light intensity higher than the rated value are used to form a curing barrier wall.

5. The UV printing multi-layer thickness image processing method as described in claim 1, characterized in that, In S51, the correction delay of the first closed loop is less than or equal to 5 milliseconds and is completed within the same scan line; in S52, the second closed loop is triggered after completing a preset number of layers or accumulating a preset number of pixel data, and only some parameter layers of the generative adversarial network are adjusted during the update; in S53, the third closed loop is triggered after accumulating a preset number of printing tasks, and the model for generating the target thickness intention is fine-tuned.

6. The UV printing multi-layer thickness image processing method as described in claim 1, characterized in that, The spatial light modulator in S3 is a Micro-LED array or a digital micromirror device; the resolution of the UV light intensity mask is the same as the printing resolution, and the light intensity and pulse width of each pixel are independently controllable; the peak wavelength of the UV light intensity is selected from at least one of 365nm or 395nm.

7. The UV printing multi-layer thickness image processing method as described in claim 1, characterized in that, In step S4, an optical non-contact displacement sensor is used to measure the actual thickness distribution. The sensor is installed on the printhead assembly and located after the curing unit. Its measurement accuracy is better than ±2 micrometers, and the lateral sampling interval is matched with the printing pixel interval. The method also includes: when an area with an absolute value greater than a preset threshold and a continuous area exceeding a preset value appears in the error map, it is judged as a printing abnormality and an alarm or pause action is executed.

8. The UV printing multi-layer thickness image processing method as described in claim 1, characterized in that, It also includes a gesture recognition step: based on the motion parameters of the pen stroke, at least one of the following gestures is detected, and the target thickness intention is modified accordingly: a quick back-and-forth smearing gesture, a circular rotation pen stroke gesture, and a double-tap gesture; wherein, the quick back-and-forth smearing gesture corresponds to performing Gaussian blur, the circular rotation pen stroke gesture corresponds to generating a Gaussian-shaped protrusion at the center of the circle, and the double-tap gesture corresponds to adding a preset fixed height at the current position.

9. A UV printing multi-layer thickness image processing system, applied to the UV printing multi-layer thickness image processing method according to any one of claims 1-8, characterized in that, include: The target thickness intent acquisition module is used to acquire or generate the target thickness intent and output the target thickness intent to the multiphysics GAN inverse compensation module; The multiphysics GAN inverse compensation module includes a generative adversarial network, which receives the target thickness intention and the input material parameter vector, and outputs a three-dimensional joint instruction tensor to the multiphysics collaborative printing execution module based on the target thickness intention and the material parameter vector; the three-dimensional joint instruction tensor includes an ink droplet volume distribution map, an ink droplet target temperature distribution map and a UV light intensity mask map. The multiphysics collaborative printing execution module includes a piezoelectric printhead array, a thermoelectric temperature control device for controlling the temperature of ink droplets, and a spatial light modulator for selective curing; the multiphysics collaborative printing execution module coordinates the control of ink droplet volume, ink droplet temperature and UV light intensity at the pixel level according to the received three-dimensional joint instruction tensor. The in-situ thickness monitoring and feedback module is used to measure the actual thickness distribution of the printed layer, calculate the error map between the actual thickness and the expected thickness, and output the error map to the three-layer closed-loop self-learning controller. A three-layer closed-loop self-learning controller is used to receive the error map and feed the error map back to: the first closed loop, which corrects the three-dimensional joint instruction tensor of the unprinted area of ​​the current layer in real time according to the error map, and sends the corrected instruction to the multiphysics collaborative printing execution module; The second closed loop uses the error map and the corresponding three-dimensional joint instruction tensor and actual thickness data as training samples to update the generative adversarial network; The third closed loop compares the error map with the target thickness intention and adjusts the mapping model in the target thickness intention acquisition module.

10. The UV printing multi-layer thickness image processing system as described in claim 9, characterized in that, The target thickness intent acquisition module includes a pressure-sensitive stylus and a neural network for mapping tactile features to thickness intent; the thermoelectric temperature control device is a thermoelectric temperature control unit independently configured for each nozzle, and the thermoelectric temperature control unit is a Peltier element or a resistance heating element; the spatial light modulator is a Micro-LED array or a digital micromirror device; the in-situ thickness monitoring and feedback module adopts a line-scan spectral confocal displacement sensor or a laser triangulation displacement sensor; the system also includes an anomaly detection unit and a cloud database. The anomaly detection unit is used to determine printing anomalies when the error exceeds a preset threshold and the continuous area exceeds a preset value. The cloud database is used to store data for each printing task for offline training or federated learning updates.

Images