Non-destructive thickness evaluation method of perovskite inkjet printed thin films based on regression models

Abstract

The application discloses a perovskite ink-jet printing film lossless thickness evaluation method based on a regression model and a related device, wherein the method comprises the following steps: taking the film thickness reference value of each sample in a training data set as a supervision signal, updating the network parameters of a multi-modal time series deep regression network model, and saving the cross-modal attention parameters and time series feature encoding parameters in the network parameters during the training process; arranging the online dynamic optical characteristic time series vector in time sequence; obtaining the online photoluminescence dynamic characteristic vector according to the online multi-channel photoluminescence intensity time series signal; inputting the online multi-modal time series fusion characteristic vector into the multi-modal time series deep regression network model, and outputting the online film thickness prediction value; and outputting the online film thickness prediction value according to the online time series fusion characteristic vector. The application avoids the 'unreasonable extrapolation' problem of a pure data-driven model when extrapolating outside the process parameter combination, and improves the generalization stability of the model in long-term operation of a production line.

Description

Technical Field

[0001] This invention relates to the field of online quality monitoring technology for perovskite inkjet printed films, specifically to a non-destructive thickness assessment method, apparatus, and computing equipment for perovskite inkjet printed films based on a regression model. Background Technology

[0002] Roll-to-roll continuous inkjet printing has become the mainstream technology for the large-scale manufacturing of perovskite thin films due to its advantages such as high material utilization, flexible patterning, and large-area fabrication. However, the photoelectric properties of perovskite thin films are highly dependent on their thickness uniformity; thickness deviations lead to changes in carrier recombination rates, fluctuations in light absorption efficiency, and a significant decrease in device yield. Therefore, achieving online non-destructive assessment of film thickness in high-speed continuous production can ensure product quality consistency and improve production line yield.

[0003] Currently, perovskite thin film thickness measurement is mainly divided into two categories: offline and online. Offline contact / high-precision optical measurement methods, such as white light interferometers, profilometers, and ellipsometry, can provide sub-nanometer thickness accuracy and are often used as calibration benchmarks. However, they require cutting samples or damaging the film structure, making them destructive or semi-destructive tests, and thus unable to achieve full inspection. The measurement process takes several seconds to several minutes, which cannot match the conveyor belt speed of several meters per minute on roll-to-roll production lines. Online optical monitoring methods mainly include thickness measurement schemes based on the attenuation of single transmitted light intensity and indirect thickness measurement schemes based on photoluminescence intensity calibration. When relying solely on changes in transmitted light intensity, factors such as substrate transmittance fluctuations, ambient stray light, and uneven ink droplet spreading can severely interfere with measurement accuracy, resulting in a relative error in thickness prediction that often exceeds 15%. When relying solely on photoluminescence intensity, the luminescence quantum efficiency is significantly affected by excitation power drift, changes in film crystallinity, and the environment, especially with a sharp decrease in sensitivity in the thinner thickness range.

[0004] To address the aforementioned issues, this invention proposes a non-destructive thickness assessment method for perovskite inkjet-printed thin films based on a regression model. This avoids the problem of "unreasonable extrapolation" when extrapolating process parameter combinations using a purely data-driven model, thereby improving the generalization stability of the model during long-term operation on the production line. Summary of the Invention

[0005] In view of the above problems, the present invention provides a method, apparatus and computing device for non-destructive thickness evaluation of perovskite inkjet printed films based on a regression model.

[0006] According to one aspect of the present invention, a non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model is provided, comprising:

[0007] For perovskite thin films prepared by roll-to-roll continuous inkjet printing, multiple training samples covering the process parameter tolerance window and the film thickness variation range were collected. Simultaneously, the film thickness reference value, inkjet printing process parameter records, transmitted light image sequences at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signals were acquired for each training sample. The film thickness reference value was obtained using an offline white light interferometer profilometer as a supervisory label for regression modeling. The inkjet printing process parameter records included droplet volume settings, printing substrate temperature values, droplet spacing settings, and vacuum-assisted pumping rates as static process features.

[0008] The mean gray level, standard deviation of gray level space, gray level gradient amplitude statistics, and local binary pattern texture features of each frame in the transmitted light image sequence are extracted and arranged in chronological order to form a dynamic optical feature time-series vector. The peak emission intensity, decay time constant, and peak wavelength drift of each channel in the multi-channel photoluminescence intensity time-series signal are extracted to form a photoluminescence dynamics feature vector. The static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector are concatenated in a multimodal feature layer to form a multimodal temporal fusion feature vector. The multimodal temporal fusion feature vector and the corresponding thin film thickness reference value are used to form training samples to obtain a training dataset.

[0009] Using the thin film thickness reference value of each sample in the training dataset as a supervision signal, update the network parameters of the multimodal temporal deep regression network model until the determination coefficient on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold. Save the cross-modal attention parameters and temporal feature encoding parameters in the network parameters during training.

[0010] On a roll-to-roll inkjet printing production line for perovskite thin films, online inkjet printing process parameters of the film under test are acquired in real time. A transmitted light imaging device positioned before the curing station continuously acquires online transmitted light image sequences at multiple moments during film formation. A multi-channel photoluminescence monitoring device simultaneously acquires online multi-channel photoluminescence intensity time-series signals. Based on the online transmitted light image sequences, online image feature vectors for each moment are obtained and arranged chronologically to form an online dynamic optical feature time-series vector. Based on the online multi-channel photoluminescence intensity time-series signals, an online photoluminescence dynamics feature vector is obtained. The online inkjet printing process parameters are used as online static process features and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal time-series fusion feature vector.

[0011] The online multimodal temporal fusion feature vector is input into the multimodal temporal deep regression network model, which outputs an online thin film thickness prediction value. This online thin film thickness prediction value is used as the non-destructive online thickness evaluation result of the perovskite thin film under test and is fed back to the production line manufacturing execution system in real time. Specifically, the multimodal temporal deep regression network model calls the stored cross-modal attention parameters to output an online aligned multimodal feature vector, calls the temporal feature encoding parameters to perform temporal encoding on the online aligned multimodal feature vector along the time dimension, and outputs an online temporal fusion feature vector. The online thin film thickness prediction value is output based on the online temporal fusion feature vector.

[0012] In one alternative approach, the multimodal temporal deep regression network model is composed of an input feature normalization layer, a cross-modal temporal alignment module, a temporal feature encoding backbone network, a physical constraint residual correction module, and a thickness regression output head cascaded together.

[0013] The input feature normalization layer is used to receive the online multimodal temporal fusion feature vector and output the standardized multimodal feature tensor.

[0014] The cross-modal temporal alignment module is used to receive the standardized multimodal feature tensor, call the saved cross-modal attention parameters to perform learnable positional encoding of the dynamic optical feature temporal sub-vector along the time dimension, use the photoluminescence dynamic feature sub-vector as the query guide and the dynamic optical feature temporal sub-vector as the query content to perform cross-modal multi-head attention calculation, output the optical emission alignment feature sequence, map the static process feature sub-vector through a linear projection layer to a static extended vector with the same feature dimension as the alignment feature sequence, and copy and splice the static extended vector with the optical emission alignment feature sequence in the time dimension to obtain the time-aligned multimodal feature tensor;

[0015] The temporal feature encoding backbone network includes a serial structure of stacked bidirectional recurrent memory network layers and temporal attention layers. It is used to receive the time-aligned multimodal feature tensor, call the saved temporal feature encoding parameters, extract temporal dependencies of the time-aligned multimodal feature tensor along the forward and reverse time directions through the bidirectional recurrent memory network layer, and output the hidden state sequence.

[0016] The temporal attention layer is used to apply self-attention weights to the hidden state sequence along the time dimension and output an online temporal fusion feature vector;

[0017] The physical constraint residual correction module is used to receive the online time-series fusion feature vector and the standardized static process feature sub-vector, calculate the physical prior thickness estimate using a pre-fitted surrogate model between the perovskite film thickness and the ink droplet volume set value, the substrate temperature and the vacuum pumping rate, concatenate the online time-series fusion feature vector and the physical prior thickness estimate, calculate the residual correction amount through the residual network, and output the physically corrected time-series fusion feature vector.

[0018] The thickness regression output head is used to receive the physically corrected temporal fusion feature vector and map it into a scalar, and output the online thin film thickness prediction value.

[0019] In one alternative approach, acquiring multiple training samples covering the process parameter tolerance window and the range of film thickness variation further includes:

[0020] Multiple sets of process parameter combinations are generated within the tolerance window of inkjet printing process parameters. The tolerance window at least covers the allowable deviation range of each of the droplet volume setting value, the printing substrate temperature value, the droplet spacing setting value, and the vacuum-assisted pumping rate. Each set of process parameter combinations is fully or approximately fully factored within the tolerance window according to at least three levels.

[0021] For each combination of process parameters, samples are cut from multiple transverse positions along the film travel direction in the width direction in the same roll-to-roll continuous inkjet printing batch. Multiple sub-samples are cut continuously at fixed intervals along the longitudinal direction at each transverse position. Each combination of process parameters corresponds to multiple sub-samples, thus obtaining the total number of training samples.

[0022] In an alternative approach, the acquisition process of the transmitted light image sequence further includes:

[0023] During the thin film formation process, a transmission light imaging device synchronously triggers alternating or simultaneous exposure of each channel at a fixed sampling frequency to obtain a transmission light image sequence. Each frame of the transmission light image sequence is obtained by weighted fusion of multiple single-wavelength subframes at the pixel level. The weighting coefficients are pre-calibrated based on the absorption characteristics of the perovskite thin film at the corresponding wavelength. The transmission light imaging device employs a multi-wavelength light source array, which contains at least three independent controllable channels with different center wavelengths.

[0024] In one alternative embodiment, the multi-channel photoluminescence monitoring device is equipped with a fiber optic spectrometer array with spectral resolution capability;

[0025] The fiber optic spectrometer array consists of multiple detection channels, each corresponding to a characteristic wavelength range where different characteristic emission peaks of the perovskite thin film are located. Each detection channel independently records the signal of photoluminescence intensity changing over time at a sampling rate higher than the characteristic frequency of thin film crystallization kinetics.

[0026] In one alternative approach, during the process of feeding back the online film thickness prediction value to the production line manufacturing execution system in real time, the production line manufacturing execution system generates a closed-loop adjustment command for inkjet printing process parameters based on the deviation between the online film thickness prediction value and the target thickness value.

[0027] Specifically, if the current online film thickness prediction value is higher than the upper limit of the thickness target value, the droplet volume setting value is reduced first, while the vacuum-assisted pumping rate is increased. The adjustment amount of reducing the droplet volume setting value and increasing the vacuum-assisted pumping rate is allocated inversely proportional to the sensitivity coefficient of each process parameter to the pre-calibrated thickness. If the current online film thickness prediction value is lower than the lower limit of the thickness target value, the droplet volume setting value is increased first and the substrate temperature is reduced.

[0028] Specifically, when the absolute deviation between the online film thickness prediction value and the target thickness value is in different order of magnitude ranges, large step size, nominal step size, or small step size adjustment are adopted respectively, and the step size coefficient is set according to the deviation order of magnitude. Furthermore, before generating the closed-loop adjustment command, the production line manufacturing execution system first calls the prediction variance value output by the multimodal time-series deep regression network model. If the prediction variance value exceeds the preset variance threshold, the closed-loop adjustment is paused and an alarm signal is issued. The online film thickness prediction values ​​of multiple consecutive sampling times after the current time are smoothed and filtered. Only when the smoothed and filtered prediction value still exceeds the range of the target thickness value, the closed-loop adjustment command is generated and its confidence level is marked as low confidence adjustment. The step size of the low confidence adjustment is further reduced and multiple sampling cycles are forcibly waited before the next adjustment is executed.

[0029] In one alternative approach, the calculation expression for the multimodal temporal fusion feature vector is:

[0030]

[0031] in, This is the normalized static process feature vector; This represents the total number of sampling time frames for optical images and photoluminescence signals during the thin film curing process. The temporal attention weights for the optical image features of the k-th frame; The dynamic optical feature vector extracted from the transmitted light image of the k-th frame; The bandwidth of the time-series Gaussian smoothing kernel; This represents the fundamental eigenvectors of photoluminescence dynamics. For the outer product operation of characteristic tensors; This is a low-frequency Fourier filter operator for photoluminescence dynamics signals; wavelength Down Real-time photoluminescence intensity at any given moment; These are the extreme threshold values ​​of luminescence intensity at the corresponding wavelengths; This represents the fluorescence decay time constant corresponding to the characteristic wavelength; This is the second-order spatiotemporal Laplace correction term for the cross-modal characteristics; For Hadama accumulation.

[0032] In one optional approach, the physical constraint residual correction module performs constraint correction for thickness prediction based on the thin film crystallization prior coupled residual regression formula, wherein the calculation expression of the thin film crystallization prior coupled residual regression formula is:

[0033]

[0034] in, This is the predicted final film thickness after constraint correction; The original predicted thickness is the output of network regression from the temporal fusion feature vector; The learning rate is used for residual correction. This is a priori thickness estimate based on the physical properties of the process. This is the positive mapping function for predicting thickness; This is a vector of real-time process parameters; This is the threshold value for the process parameter tolerance window; This is a step indicator function; For real-time prediction of the coefficient of determination; The minimum fitting accuracy threshold is preset; This is the vector of process standard reference parameters.

[0035] According to another aspect of the present invention, an online non-destructive thickness evaluation device for perovskite inkjet-printed films based on a regression model is provided, comprising:

[0036] The sample acquisition module is used to acquire multiple training samples covering the process parameter tolerance window and the film thickness variation range for perovskite films prepared by roll-to-roll continuous inkjet printing. Simultaneously, it acquires the film thickness reference value, inkjet printing process parameter records, transmitted light image sequences at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signals for each training sample. The film thickness reference value is obtained using an offline white light interferometer profilometer as a supervision label for regression modeling. The inkjet printing process parameter records include droplet volume settings, printing substrate temperature values, droplet spacing settings, and vacuum-assisted pumping rate as static process features.

[0037] The feature extraction module is used to extract the grayscale mean, grayscale spatial standard deviation, grayscale gradient amplitude statistics, and local binary pattern texture features of each frame in the transmitted light image sequence, and arrange them in chronological order to form a dynamic optical feature time-series vector; extract the luminescence intensity peak, decay time constant, and peak wavelength drift of each channel in the multi-channel photoluminescence intensity time-series signal to form a photoluminescence dynamics feature vector; and concatenate the static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector into a multi-modal feature layer to form a multi-modal temporal fusion feature vector. The multi-modal temporal fusion feature vector and the corresponding thin film thickness reference value are used to form training samples to obtain a training dataset.

[0038] The model training module is used to update the network parameters of the multimodal temporal deep regression network model using the thin film thickness reference value of each sample in the training dataset as a supervision signal, until the determination coefficient on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold, and saves the cross-modal attention parameters and temporal feature encoding parameters in the network parameters during the training process.

[0039] An online acquisition module is used to acquire online inkjet printing process parameters of the film under test in real time on a perovskite film roll-to-roll inkjet printing production line. A transmitted light imaging device positioned before the curing station continuously acquires online transmitted light image sequences at multiple moments during film formation. A multi-channel photoluminescence monitoring device simultaneously acquires online multi-channel photoluminescence intensity time-series signals. Based on the online transmitted light image sequences, online image feature vectors for each moment are obtained and arranged chronologically to form an online dynamic optical feature time-series vector. An online photoluminescence dynamics feature vector is obtained based on the online multi-channel photoluminescence intensity time-series signals. The online inkjet printing process parameters are used as online static process features and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal time-series fusion feature vector.

[0040] The prediction feedback module is used to input the online multimodal temporal fusion feature vector into the multimodal temporal deep regression network model, output the online thin film thickness prediction value, and use the online thin film thickness prediction value as the non-destructive online thickness evaluation result of the perovskite thin film to be tested, and feed it back to the production line manufacturing execution system in real time. Specifically, the multimodal temporal deep regression network model calls the saved cross-modal attention parameters to output the online aligned multimodal feature vector, calls the temporal feature encoding parameters to perform temporal encoding on the online aligned multimodal feature vector along the time dimension, and outputs the online temporal fusion feature vector; the online thin film thickness prediction value is output based on the online temporal fusion feature vector.

[0041] According to another aspect of the present invention, a computing device is provided, comprising: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other via the communication bus;

[0042] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described regression model-based non-destructive thickness evaluation method for perovskite inkjet-printed thin films.

[0043] According to the solution provided by the present invention, for perovskite thin films prepared by roll-to-roll continuous inkjet printing, multiple training samples covering the process parameter tolerance window and the film thickness variation range are collected; simultaneously, the film thickness reference value, inkjet printing process parameter record, transmitted light image sequence at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signal of each training sample are acquired; wherein, the film thickness reference value is obtained using an offline white light interferometer as a supervision label for regression modeling; the inkjet printing process parameter record includes ink droplet volume setting value, printing substrate temperature value, droplet spacing setting value, and vacuum-assisted pumping rate as static process features; and the transmitted light image sequence is extracted from each... The grayscale mean, grayscale spatial standard deviation, grayscale gradient magnitude statistics, and local binary pattern texture features of a frame image are arranged in chronological order to form a dynamic optical feature time-series vector. The peak luminescence intensity, decay time constant, and peak wavelength shift of each channel in the multi-channel photoluminescence intensity time-series signal are extracted to form a photoluminescence dynamics feature vector. The static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector are concatenated in a multi-modal feature layer to form a multi-modal temporal fusion feature vector. The multi-modal temporal fusion feature vector and the corresponding thin film thickness reference value are used to form training samples to obtain a training dataset. The thin film thickness of each sample in the training dataset is used as the training dataset. The reference value is used as a supervision signal to update the network parameters of the multimodal temporal deep regression network model until the coefficient of determination on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold. The cross-modal attention parameters and temporal feature encoding parameters from the network parameters during training are saved. On a perovskite thin film roll-to-roll inkjet printing production line, the online inkjet printing process parameters of the film under test are acquired in real time. A transmitted light imaging device positioned before the curing station continuously acquires online transmitted light image sequences at multiple moments during the film formation process. A multi-channel photoluminescence monitoring device simultaneously acquires online multi-channel photoluminescence intensity temporal signals. The results for each moment are obtained from the online transmitted light image sequences. The online image feature vectors are arranged in chronological order to form an online dynamic optical feature time-series vector; an online photoluminescence dynamics feature vector is obtained based on the online multi-channel photoluminescence intensity time-series signal; the online inkjet printing process parameters are used as online static process features, and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal temporal fusion feature vector; the online multimodal temporal fusion feature vector is input into the multimodal temporal depth regression network model to output an online thin film thickness prediction value; the online thin film thickness prediction value is used as the non-destructive online thickness evaluation result of the perovskite thin film to be tested, and is fed back to the production line manufacturing execution system in real time;The multimodal temporal deep regression network model calls the stored cross-modal attention parameters to output an online-aligned multimodal feature vector, and calls the temporal feature encoding parameters to perform temporal encoding on the online-aligned multimodal feature vector along the time dimension, outputting an online temporal fusion feature vector; based on the online temporal fusion feature vector, an online thin film thickness prediction value is output. This invention avoids the "unreasonable extrapolation" problem when extrapolating process parameter combinations in a purely data-driven model, improving the model's generalization stability during long-term production line operation.

[0044] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0045] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0046] Figure 1 A flowchart illustrating the non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model, according to an embodiment of the present invention, is shown.

[0047] Figure 2a The following is a timing diagram of the production line closed-loop regulation and safety arbitration based on confidence gating according to an embodiment of the present invention;

[0048] Figure 2b The following is a timing diagram of multimodal feature fusion and online inference according to an embodiment of the present invention;

[0049] Figure 3 A framework structure diagram of a perovskite thin film prepared by a continuous inkjet printing process according to an embodiment of the present invention is shown.

[0050] Figure 4 A schematic diagram of the framework of the online non-destructive thickness evaluation device for perovskite inkjet printed films based on a regression model according to an embodiment of the present invention is shown.

[0051] Figure 5 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. Detailed Implementation

[0052] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0053] Figure 1 A schematic flowchart of a regression model-based non-destructive thickness evaluation method for perovskite inkjet-printed thin films, according to an embodiment of the present invention, is shown. Specifically, as... Figure 1 , Figure 2a , Figure 2b As shown, it includes the following steps:

[0054] Step S101: For perovskite thin films prepared by roll-to-roll continuous inkjet printing, multiple training samples covering the process parameter tolerance window and the range of film thickness variation are collected; simultaneously, the film thickness reference value, inkjet printing process parameter record, transmitted light image sequence at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signal are acquired for each training sample; wherein, the film thickness reference value is obtained using an offline white light interferometer as a supervision label for regression modeling; the inkjet printing process parameter record includes the droplet volume setting value, the printing substrate temperature value, the droplet spacing setting value, and the vacuum-assisted pumping rate as static process features.

[0055] In this embodiment, the training samples are not collected under a single optimal process, but intentionally cover the process parameter tolerance window and thickness variation range. This allows the model to learn the complex relationship between process fluctuations (such as droplet volume deviation, substrate temperature drift, and vacuum pumping rate fluctuations) and the final film thickness. This ensures stable prediction accuracy even when actual production line conditions deviate from the set values, preventing the model from being effective only under ideal conditions. Transmitted light image sequences record the dynamic optical evolution during film formation (such as transmittance changes during curing), while multi-channel photoluminescence timing signals record crystallization information (such as defect state evolution and phase transition processes). This allows for a temporal understanding of the film growth process, rather than relying solely on static final-state features, significantly improving the characterization of the thickness formation mechanism. Thickness reference values ​​obtained from an offline white light interferometer are used as supervisory labels. Nanometer-level precision provides a reliable "true value" benchmark for regression modeling. The transmitted light images and photoluminescence signals used for training are non-destructive optical inspection methods, requiring no damage or contact with the film during online application. This ensures the reliability of the model training while meeting the non-contact and high-speed inspection requirements of roll-to-roll production lines.

[0056] Specifically, such as Figure 3As shown, within the adjustable range of the inkjet printing equipment, the nominal center values ​​and allowable deviation ranges (e.g., nominal values ​​±5% to ±10%) for the droplet volume setting, printing substrate temperature, droplet spacing setting, and vacuum-assisted pumping rate are determined. Each process parameter is divided into at least three levels (e.g., low, medium, and high) within its tolerance window, and multiple combinations of process parameters are generated using full factorial design or orthogonal experimental design. For each combination, corresponding perovskite thin film samples are prepared under stable roll-to-roll continuous printing conditions, ensuring sufficient film length for subsequent sampling under each process. A transmitted light imaging device is set up before the printing and curing station. The light source uses an LED array containing multiple center wavelengths (e.g., 450 nm, 550 nm, and 650 nm), and the camera is synchronously triggered with the light source. As the thin film substrate moves continuously, transmitted light images of the same position on the film are continuously captured at fixed time intervals (e.g., one frame every 0.5 seconds) until the film is completely cured. Each frame image is fused from subframes of each wavelength channel according to pre-calibrated absorption weights and saved as time-series image data. A fiber optic spectrometer array was set up at the same monitoring position for transmitted light imaging, and the detection channels were tuned to the characteristic emission peak bands of the perovskite material (such as the main emission peak and shoulder peak positions). During the thin film formation process, the emission intensity of each channel was continuously recorded over time at a sampling rate higher than the characteristic frequency of crystallization kinetics. This data was saved as a multi-channel time-series signal and strictly aligned with the transmitted light image sequence on the time axis (synchronization was achieved through a shared hardware trigger signal). During the printing process of each set of process parameters, the actual values ​​fed back by the inkjet printing control system were recorded in real time. The droplet volume setting was calculated from waveform parameters and printhead drive voltage, the substrate temperature was fed back by thermocouples, the droplet spacing setting was calculated from printhead frequency and substrate speed, and the vacuum-assisted pumping rate was fed back by a mass flow meter. These values ​​served as the static process feature vector of the sample and were correlated with dynamic optics and photoluminescence data. After cutting the thin film samples prepared under each process, a white light interferometer was used to perform surface scanning measurements on the thin film surface. Multiple measurement points were selected along the thin film travel direction and width direction to obtain the thin film thickness value at each point. The average thickness measurement values ​​of multiple sub-samples under the same combination of process parameters (or matched according to the correspondence between measurement positions and dynamic monitoring positions) are used as the thickness reference value label for that sample, serving as the supervision signal during regression model training. The static process feature vector, transmitted light image temporal features (image frame sequence arranged in chronological order), multi-channel photoluminescence intensity temporal signal, and corresponding thickness reference value label of each training sample are used as training records.

[0057] In one alternative approach, acquiring multiple training samples covering the process parameter tolerance window and the range of film thickness variation further includes:

[0058] Multiple sets of process parameter combinations are generated within the tolerance window of inkjet printing process parameters. The tolerance window at least covers the allowable deviation range of each of the droplet volume setting value, the printing substrate temperature value, the droplet spacing setting value, and the vacuum-assisted pumping rate. Each set of process parameter combinations is fully or approximately fully factored within the tolerance window according to at least three levels.

[0059] For each combination of process parameters, samples are cut from multiple transverse positions along the film travel direction in the width direction in the same roll-to-roll continuous inkjet printing batch. Multiple sub-samples are cut continuously at fixed intervals along the longitudinal direction at each transverse position. Each combination of process parameters corresponds to multiple sub-samples, thus obtaining the total number of training samples.

[0060] In this embodiment, samples are cut at multiple lateral positions along the width direction, which can capture edge effects and lateral thickness non-uniformity commonly seen in roll-to-roll large-area coating. The training data contains information on the variation of the film in the width direction. Even when the monitoring points are located at different positions along the width, accurate predictions can still be given during online application, greatly improving the adaptability to the complex flow field and drying conditions of the actual production line. Cutting multiple sub-samples continuously at fixed intervals along the direction of travel can capture thickness fluctuations caused by non-ideal factors such as batch differences in substrates, drift in ambient temperature and humidity, and slow changes in nozzle status under the same combination of process parameters. This makes the training set contain information on unavoidable random disturbances in the real production line, enhancing the robustness of the model against interference. Each combination of process parameters corresponds to multiple sub-samples. The thickness labels measured offline can reduce the impact of single-point measurement noise by statistically summarizing multiple measurements (such as mean and median), obtaining a more reliable supervision signal. At the same time, the thickness variance among multiple sub-samples can also be used as the basis for sample weights during model training, guiding the model to give differentiated learning attention to the stable process region and the high fluctuation region.

[0061] For example, the nominal process parameters for a roll-to-roll perovskite inkjet printing production line are: droplet volume 10 pL, substrate temperature 60 ℃, droplet spacing 30 μm, and vacuum pumping rate 50 L / min. The tolerance windows for each parameter are: droplet volume 9~11 pL, substrate temperature 55~65 ℃, droplet spacing 28~32 μm, and vacuum pumping rate 45~55 L / min. Droplet volumes: 9.0, 10.0, 11.0 (3 levels); substrate temperatures: 55, 60, 65 (3 levels); droplet spacing: 28, 30, 32 (3 levels); vacuum pumping rates: 45, 50, 55 (3 levels). Using a full factorial design, theoretically there are 3×3×3×3 = 81 combinations of process parameters. Considering actual printing time and material costs, an orthogonal array L9(3) is used. 4The design selected nine representative combinations, covering a balanced mix of low, medium, and high levels for each parameter, while ensuring that extreme combinations (such as all parameters simultaneously at low or high levels) were included. For each combination of process parameters, a 2-meter-long film was continuously printed. Five horizontal positions were set along the width (1.2 meters): 5 cm from the left edge, 30 cm, 60 cm (center), 90 cm, and 115 cm. A subsample was cut every 5 cm along the longitudinal direction from each horizontal position, resulting in a total of 5 (horizontal) × 40 (longitudinal) = 200 subsamples per group. Group number: Run-06 (corresponding process combination: droplet volume 9.5 pL, substrate temperature 62 ℃, droplet spacing 29 μm, vacuum pumping rate 52 L / min); lateral position: C (center position), longitudinal sequence: 12th cutting point (60 cm from the starting point), this subsample number: Run-06-C-12. The thickness value of this subsample was subsequently measured by a white light interferometer as a training label. Its corresponding actual process feedback value and dynamic optical / photoluminescence timing data were also included in the database to form a complete training sample. 9 process groups × 5 lateral positions × 40 longitudinal subsamples = 1800 training subsamples. The sample size ensures sufficient coverage of the process parameter space and encompasses the spatial thickness variation in the width and travel directions, providing a representative data foundation.

[0062] In an alternative approach, the acquisition process of the transmitted light image sequence further includes:

[0063] During the thin film formation process, a transmission light imaging device synchronously triggers alternating or simultaneous exposure of each channel at a fixed sampling frequency to obtain a transmission light image sequence. Each frame of the transmission light image sequence is obtained by weighted fusion of multiple single-wavelength subframes at the pixel level. The weighting coefficients are pre-calibrated based on the absorption characteristics of the perovskite thin film at the corresponding wavelength. The transmission light imaging device employs a multi-wavelength light source array, which contains at least three independent controllable channels with different center wavelengths.

[0064] In this embodiment, the response of a single wavelength of transmitted light to film thickness is often limited by the interference period within a specific thickness range, easily leading to sensitivity saturation or non-monotonic response within certain thickness ranges. By using at least three independent channels with different center wavelengths and performing pixel-level weighted fusion, the differences in absorption coefficients of perovskite materials at different wavelengths can be utilized to construct a composite optical response with high sensitivity across a wider thickness range. The weighting coefficients are pre-calibrated based on the actual absorption characteristics of the perovskite film at each corresponding wavelength. Wavelength channels with stronger absorption receive higher weights during fusion, while the weights of wavelength channels with large penetration depths and insensitivity to film thickness changes are suppressed, ensuring that each frame of transmitted light image reflects thickness information to the greatest extent. Exposure of each channel is synchronously triggered at a fixed sampling frequency to ensure that all wavelength subframes strictly correspond to the same moment in the film formation process. Changes in transmitted light at different wavelengths must be time-aligned to accurately reflect the consistent behavior of crystallization, phase transition, and thickness evolution during film curing, avoiding temporal distortion introduced by misaligned acquisition times.

[0065] Step S102: Extract the grayscale mean, grayscale spatial standard deviation, grayscale gradient amplitude statistics, and local binary pattern texture features of each frame in the transmitted light image sequence, and arrange them in chronological order to form a dynamic optical feature time-series vector; extract the luminescence intensity peak, decay time constant, and peak wavelength drift of each channel in the multi-channel photoluminescence intensity time-series signal to form a photoluminescence dynamics feature vector; concatenate the static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector into a multi-modal feature layer to form a multi-modal temporal fusion feature vector, and use the multi-modal temporal fusion feature vector and the corresponding thin film thickness reference value to form training samples to obtain a training dataset.

[0066] In this embodiment, for example, during the curing process of the training sample, a transmission light imaging device acquired 60 fused images over 30 seconds at a frequency of 2 Hz. A multi-channel photoluminescence monitoring device was configured with three detection channels (corresponding to the main emission peak of the perovskite and two defect-related peaks, respectively), simultaneously recording a 30-second time-series signal. The static process parameters of the sample were: droplet volume 9.5 pL, substrate temperature 62 ℃, droplet spacing 29 μm, vacuum pumping rate 52 L / min, and offline white light interferometry thickness reference value of 450 nm. The 10th frame image (5 seconds after curing started) yielded a grayscale mean of 142, a grayscale spatial standard deviation of 8.3, a gradient amplitude mean of 12.7, and the main peak of the LBP texture histogram was located in the uniform texture region. In frame 30 (15 seconds after curing begins), the average grayscale value drops to 118 (due to decreased transmittance from film thickening), the spatial standard deviation rises to 12.1 (due to increased surface inhomogeneity caused by crystallization), and the average gradient amplitude rises to 18.4 (due to increased grain edges). The entire 60 frames, arranged chronologically, form a 60×64-dimensional dynamic optical characteristic temporal matrix. The luminescence intensity of the main emission peak channel (550 nm) reaches its peak of 2100 au 8 seconds after curing begins, then slowly decays to 1650 au at the end of curing. A single exponential fit yields a decay time constant τ = 5.2 seconds. Simultaneously, the peak wavelength of this channel redshifts from 548 nm to 555 nm during curing, with a total drift of 7 nm. The defect-related peak channel (680 nm) has a lower luminescence intensity (peak value only 320 au), decays faster (τ = 2.8 seconds), and its peak wavelength drift is only 2 nm. A total of 9-dimensional photoluminescence dynamics feature vectors were extracted from the three channels: [2100, 5.2, 7, 320, 2.8, 2, 1850, 4.5, 4] (representing the peak intensity, attenuation constant, and wavelength shift of each channel, respectively). The multimodal temporal fusion feature vector formed by this sample includes: 4-dimensional static process features ([9.5, 62, 29, 52]), a 60×64-dimensional temporal matrix of dynamic optical features, and 9-dimensional photoluminescence dynamics features. This fused feature vector, paired with the 450 nm thickness label, was stored as a training sample in the training dataset. Other samples in the training set may exhibit completely different feature patterns. For example, the 620 nm thick film sample has a lower grayscale mean (approximately 105), a higher peak luminescence intensity (2750 au), but faster attenuation (τ=3.5 seconds), and a more significant wavelength redshift (12 nm) at frame 10. By comparing the pattern differences of the above samples, the statistical correlation between "low gray-scale mean, high emission peak, fast decay, large redshift" and "thick film" can be learned. In online production applications, for newly collected unknown samples, only the multimodal features of the same category need to be extracted and input into the regression network to quickly output the thickness prediction value, without waiting for offline measurement.

[0067] Step S103: Using the thin film thickness reference value of each sample in the training dataset as a supervision signal, update the network parameters of the multimodal temporal deep regression network model until the determination coefficient on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold. Save the cross-modal attention parameters and temporal feature encoding parameters in the network parameters during training.

[0068] In this embodiment, both a coefficient of determination not lower than a first threshold (e.g., R² ≥ 0.90) and a root mean square error not higher than a second threshold (e.g., RMSE ≤ 15 nm) are used as training stopping conditions to avoid the one-sidedness of evaluation by a single indicator. The first threshold ensures the model's ability to explain the thickness variance, while the second threshold directly constrains the absolute error magnitude of the predicted values. Together, they ensure that the model not only fits well in a statistical sense but also meets the accuracy requirements of actual engineering, preventing overfitting or underfitting. The performance of an independent validation set (rather than the training set) is used as the termination criterion for parameter updates to avoid the model over-memorizing the training data. Training stops and parameters are saved when the validation set performance meets the threshold requirements, preventing continued training from causing the validation set error to rebound (overfitting) and ensuring that the saved model parameters can maintain stable predictive performance on new data. Only cross-modal attention parameters and temporal feature encoding parameters are saved, not all parameters of the entire model. During online inference, the static process feature mapping and thickness regression head do not need to rely on the specific parameter states during the training phase and can be initialized or replaced independently. The saved cross-modal attention parameters and temporal feature encoding parameters are directly invoked during online applications and are no longer updated (frozen), ensuring the deterministic and consistent predictive behavior of the same model on different batches of films at different times. Simultaneously, freezing the parameters means that backpropagation or gradient calculations are unnecessary during online inference, significantly reducing computational latency and meeting the real-time requirements of high-speed roll-to-roll production line inspection.

[0069] Step S104: On the perovskite thin film roll-to-roll inkjet printing production line, the online inkjet printing process parameters of the film to be tested are collected in real time. A transmitted light imaging device positioned before the curing station continuously collects online transmitted light image sequences at multiple moments during the film formation process. A multi-channel photoluminescence monitoring device simultaneously collects online multi-channel photoluminescence intensity time-series signals. Based on the online transmitted light image sequences, online image feature vectors for each moment are obtained and arranged in chronological order to form an online dynamic optical feature time-series vector. Based on the online multi-channel photoluminescence intensity time-series signals, an online photoluminescence dynamics feature vector is obtained. The online inkjet printing process parameters are used as online static process features and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal time-series fusion feature vector.

[0070] In this embodiment, both transmitted light imaging and photoluminescence monitoring are non-contact optical detection methods, requiring no contact with the film surface during acquisition. Simultaneously, both types of signals can be acquired in real-time during film travel, without requiring machine stoppages or deceleration, perfectly adapting to the high-speed cycle time (typically several meters to tens of meters per minute) of roll-to-roll continuous production, avoiding the lag and destructive nature of offline sampling inspection. The transmitted light imaging device is positioned before the curing station, completing signal acquisition and thickness prediction before the film is fully cured. This allows the thickness assessment results to be output ahead of the curing completion time. When the predicted value deviates from the target, the production line manufacturing execution system can intervene in advance to adjust process parameters before film curing, preventing the continued production of defective films and material waste.

[0071] In one alternative embodiment, the multi-channel photoluminescence monitoring device is equipped with a fiber optic spectrometer array with spectral resolution capability;

[0072] The fiber optic spectrometer array consists of multiple detection channels, each corresponding to a characteristic wavelength range where different characteristic emission peaks of the perovskite thin film are located. Each detection channel independently records the signal of photoluminescence intensity changing over time at a sampling rate higher than the characteristic frequency of thin film crystallization kinetics.

[0073] In this embodiment, each detection channel independently records the signal at a sampling rate higher than the characteristic frequency of film crystallization kinetics, ensuring complete capture of transient dynamic details such as the rapid rise, peak, and decay of luminescence intensity during the curing process. If the sampling rate is lower than the crystallization kinetic frequency, key kinetic inflection points and peak information are lost, leading to distortion of features such as the extracted decay time constant, thus affecting the accuracy of thickness prediction. A fiber optic spectrometer with spectral resolution can not only record luminescence intensity but also track the peak wavelength positions of each emission peak in real time. During the curing and annealing process of perovskite films, characteristic emission peaks often undergo redshift or blueshift due to lattice strain, phase transition, or compositional segregation. This wavelength shift is an important indicator of the evolution of the film's microstructure and its final thickness, further enriching the information of the photoluminescence characteristic vector.

[0074] For example, the perovskite thin film material is a FA0.8Cs0.2PbI3 system. Its typical photoluminescence emission spectrum at room temperature exhibits three main characteristic peaks: peak A at approximately 810 nm (band-edge emission, corresponding to the perovskite α phase), shoulder peak B at approximately 770 nm (emission from residual low-bandgap phases or shallow defects), and broad peak C at approximately 850–900 nm (emission from deep-level defects or trapped states). The fiber optic spectrometer array is configured with three detection channels: channel 1 with a center wavelength of 810 nm (bandwidth 20 nm), channel 2 with a center wavelength of 770 nm (bandwidth 15 nm), and channel 3 with a center wavelength of 875 nm (bandwidth 40 nm). In previous experiments, under standard curing conditions, the emission intensity change of channel 1 was monitored. It rapidly increased to the peak value after excitation, taking approximately 3 seconds, and then slowly decayed to a stable state, taking approximately 20 seconds. The dominant frequency component of the emission intensity change rate, analyzed by FFT, was approximately 8 Hz. With a 5-fold margin, the sampling rate was set to 40 Hz, meaning one data point was recorded every 25 milliseconds, and each channel operated synchronously at a 40 Hz sampling rate. After the thin film entered the monitoring window, the excitation light source was turned on, and recording began simultaneously for each channel. Within 0–10 seconds, the emission intensity of channel 1 (810 nm main peak) rapidly increased from 200 au to a peak of 1850 au (reaching its peak around the 8th second), and then began to decay. Channel 2 (770 nm shoulder peak) had a lower emission intensity, with a peak of only 420 au, and decayed significantly faster, dropping to 50% of its peak value by the 12th second. Channel 3 (875 nm broad peak) maintained a consistently low emission intensity (peak <150 a.u.), exhibiting a gradually changing plateau between the 15th and 25th seconds, suggesting a slow filling process of deep-level defects. Channel 1 had a peak intensity of 1850 au, a decay time constant (single-exponential fitting of the decay segment from the 8th to the 30th second) of 6.2 seconds, and a peak wavelength that red-shifted from 808 nm to 815 nm throughout the curing process, with a total drift of +7 nm. Channel 2 has a peak intensity of 420 au, a decay time constant of 2.8 seconds, and a peak wavelength that drifts from 768 nm to 772 nm (a drift of +4 nm). Channel 3 has a peak intensity of 142 a.u., a decay time constant of 9.5 seconds (slow decay), and a peak wavelength that fluctuates between 875 and 885 nm (a drift of +10 nm, but with a low signal-to-noise ratio). Nine photoluminescence dynamics features were extracted from the three channels: [1850, 6.2, 7, 420, 2.8, 4, 142, 9.5, 10]. The high peak intensity and moderate decay time constant of channel 1 indicate good crystallinity and moderate thickness of the film; the relatively fast decay time constant (2.8 seconds) of channel 2 suggests a certain density of shallow defects; the low intensity and slow decay of channel 3 are generally associated with an increase in deep defects when the film is thicker. The offline thickness measurement value for samples with similar feature combinations in the training data was 510 ± 15 nm.In practical online applications, this 9-dimensional feature vector is fused with other modalities (static process + dynamic optics) and then input into the regression network to output the thickness prediction value of the film location in real time.

[0075] Step S105: Input the online multimodal temporal fusion feature vector into the multimodal temporal deep regression network model, output the online thin film thickness prediction value, and use the online thin film thickness prediction value as the non-destructive online thickness evaluation result of the perovskite thin film to be tested, and feed it back to the production line manufacturing execution system in real time; wherein, the multimodal temporal deep regression network model calls the saved cross-modal attention parameters to output the online aligned multimodal feature vector, calls the temporal feature encoding parameters to perform temporal encoding on the online aligned multimodal feature vector along the time dimension, and outputs the online temporal fusion feature vector; output the online thin film thickness prediction value according to the online temporal fusion feature vector.

[0076] In this embodiment, thickness prediction can be completed before the curing station by online acquisition of transmitted light images and photoluminescence signals. The results can be directly fed back to the Manufacturing Execution System (MES) to achieve real-time quality monitoring during the production process. The temporal feature encoding parameters (bidirectional recurrent memory network and time attention layer) encode the aligned features along the time direction, so that the online temporal fusion feature vector not only contains the current state information, but also implicitly contains the dynamic evolution information of the entire film formation process (such as crystallization rate and solvent evaporation trend), which improves the robustness to thickness fluctuations and abnormal operating conditions.

[0077] In one alternative approach, the multimodal temporal deep regression network model is composed of an input feature normalization layer, a cross-modal temporal alignment module, a temporal feature encoding backbone network, a physical constraint residual correction module, and a thickness regression output head cascaded together.

[0078] The input feature normalization layer is used to receive the online multimodal temporal fusion feature vector and output the standardized multimodal feature tensor.

[0079] The cross-modal temporal alignment module is used to receive the standardized multimodal feature tensor, call the saved cross-modal attention parameters to perform learnable positional encoding of the dynamic optical feature temporal sub-vector along the time dimension, use the photoluminescence dynamic feature sub-vector as the query guide and the dynamic optical feature temporal sub-vector as the query content to perform cross-modal multi-head attention calculation, output the optical emission alignment feature sequence, map the static process feature sub-vector through a linear projection layer to a static extended vector with the same feature dimension as the alignment feature sequence, and copy and splice the static extended vector with the optical emission alignment feature sequence in the time dimension to obtain the time-aligned multimodal feature tensor;

[0080] The temporal feature encoding backbone network includes a serial structure of stacked bidirectional recurrent memory network layers and temporal attention layers. It is used to receive the time-aligned multimodal feature tensor, call the saved temporal feature encoding parameters, extract temporal dependencies of the time-aligned multimodal feature tensor along the forward and reverse time directions through the bidirectional recurrent memory network layer, and output the hidden state sequence.

[0081] The temporal attention layer is used to apply self-attention weights to the hidden state sequence along the time dimension and output an online temporal fusion feature vector;

[0082] The physical constraint residual correction module is used to receive the online time-series fusion feature vector and the standardized static process feature sub-vector, calculate the physical prior thickness estimate using a pre-fitted surrogate model between the perovskite film thickness and the ink droplet volume set value, the substrate temperature and the vacuum pumping rate, concatenate the online time-series fusion feature vector and the physical prior thickness estimate, calculate the residual correction amount through the residual network, and output the physically corrected time-series fusion feature vector.

[0083] The thickness regression output head is used to receive the physically corrected temporal fusion feature vector and map it into a scalar, and output the online thin film thickness prediction value.

[0084] In this embodiment, the cross-modal temporal alignment module uses photoluminescence dynamics features as the attention calculation for query-guided dynamic optical image features, supplemented by learnable positional encoding. This allows it to automatically capture key time nodes with physical correlations during thin film crystallization (such as the frame with the most drastic optical texture change corresponding to a sudden change in luminescence intensity), effectively mining cross-modal co-evolution information and improving the information density of fused features. The temporal feature encoding backbone network first captures the long-term dynamic evolution laws of forward and reverse directions throughout the entire thin film formation process (such as changes in solvent evaporation rate and the temporal dependence of the crystallization induction and growth phases) through a bidirectional recurrent memory network layer (such as bidirectional LSTM / GRU). Then, an adaptive weight is applied to the hidden states at each time point through a temporal attention layer, focusing on the key moments that contribute the most to the thickness (such as the moment when the crystallization exothermic peak appears). This preserves complete temporal history information and avoids the problem of key events being submerged in long sequences, improving the sensitivity of thickness prediction to subtle process changes. The physical constraint residual correction module does not directly replace network prediction. Instead, it calculates the residual between the network output and the physical prior, thus preserving the data-driven model's ability to extract information from complex optical / luminescent signals while constraining the prediction results through physical priors. Furthermore, the magnitude of the residual correction can be tracked, providing process engineers with diagnostic criteria to determine whether the model overly relies on data fitting.

[0085] In one alternative approach, the calculation expression for the multimodal temporal fusion feature vector is:

[0086]

[0087] in, This is the normalized static process feature vector; This represents the total number of sampling time frames for optical images and photoluminescence signals during the thin film curing process. The temporal attention weights for the optical image features of the k-th frame; The dynamic optical feature vector extracted from the transmitted light image of the k-th frame; The bandwidth of the time-series Gaussian smoothing kernel; This represents the fundamental eigenvectors of photoluminescence dynamics. For the outer product operation of characteristic tensors; This is a low-frequency Fourier filter operator for photoluminescence dynamics signals; wavelength Down Real-time photoluminescence intensity at any given moment; These are the extreme threshold values ​​of luminescence intensity at the corresponding wavelengths; This represents the fluorescence decay time constant corresponding to the characteristic wavelength; This is the second-order spatiotemporal Laplace correction term for the cross-modal characteristics; For Hadama accumulation.

[0088] In this embodiment, the second-order spacetime Laplace correction term Characterizing the second-order rate of change (i.e., the acceleration of the change trend) of dynamic optical features and photoluminescence dynamics in the spatiotemporal domain can compensate for the missing information on the drastic changes in the first-order fusion information. It has a stronger ability to identify abrupt changes in film thickness caused by local inhomogeneities in the later stage of crystallization, while making the fused features smoother and more continuous in the time dimension, avoiding feature jumps caused by sensor asynchrony or sampling jitter. The static process feature vector modulates the entire fusion result through the Hadamard product (element-by-element multiplication). Under different combinations of process parameters (such as high droplet volume, low droplet volume), the contribution weights of optical and photoluminescence features to thickness are dynamically scaled, so that the same set of optical / luminescence signals are given different interpretation scales under different process backgrounds. This makes the fused features adaptive to process conditions, and the model does not need to learn the interaction effects of process parameters.

[0089] In one optional approach, the physical constraint residual correction module performs constraint correction for thickness prediction based on the thin film crystallization prior coupled residual regression formula, wherein the calculation expression of the thin film crystallization prior coupled residual regression formula is:

[0090]

[0091] in, This is the predicted final film thickness after constraint correction; The original predicted thickness is the output of network regression from the temporal fusion feature vector; The learning rate is used for residual correction. This is a priori thickness estimate based on the physical properties of the process. This is the positive mapping function for predicting thickness; This is a vector of real-time process parameters; This is the threshold value for the process parameter tolerance window; This is a step indicator function; For real-time prediction of the coefficient of determination; The minimum fitting accuracy threshold is preset; This is the vector of process standard reference parameters.

[0092] In this embodiment, The correction intensity is dynamically adjusted based on the Euclidean distance between the current real-time process parameter vector and the standard process reference vector. When the process parameters are within the standard tolerance window (small distance), the exponential term is close to 1, and physical correction is fully involved. When the process parameters deviate from the window (large distance), the exponential term decays, and the correction effect of physical prior is weakened. This avoids incorrect correction due to extrapolation errors of the physical proxy model in extreme process regions, thus achieving adaptive control of the correction intensity. Only the coefficient of determination predicted by the model Physical correction is activated only when the accuracy exceeds a preset threshold. When the data-driven model itself has poor fitting accuracy (such as encountering abnormal working conditions outside the training distribution), the correction module will pause intervention even if there is a large deviation between the physical prior and the prediction, to prevent the introduction of potentially erroneous physical priors in unreliable areas of the model.

[0093] In one alternative approach, during the process of feeding back the online film thickness prediction value to the production line manufacturing execution system in real time, the production line manufacturing execution system generates a closed-loop adjustment command for inkjet printing process parameters based on the deviation between the online film thickness prediction value and the target thickness value.

[0094] Specifically, if the current online film thickness prediction value is higher than the upper limit of the thickness target value, the droplet volume setting value is reduced first, while the vacuum-assisted pumping rate is increased. The adjustment amount of reducing the droplet volume setting value and increasing the vacuum-assisted pumping rate is allocated inversely proportional to the sensitivity coefficient of each process parameter to the pre-calibrated thickness. If the current online film thickness prediction value is lower than the lower limit of the thickness target value, the droplet volume setting value is increased first and the substrate temperature is reduced.

[0095] Specifically, when the absolute deviation between the online film thickness prediction value and the target thickness value is in different order of magnitude ranges, large step size, nominal step size, or small step size adjustment are adopted respectively, and the step size coefficient is set according to the deviation order of magnitude. Furthermore, before generating the closed-loop adjustment command, the production line manufacturing execution system first calls the prediction variance value output by the multimodal time-series deep regression network model. If the prediction variance value exceeds the preset variance threshold, the closed-loop adjustment is paused and an alarm signal is issued. The online film thickness prediction values ​​of multiple consecutive sampling times after the current time are smoothed and filtered. Only when the smoothed and filtered prediction value still exceeds the range of the target thickness value, the closed-loop adjustment command is generated and its confidence level is marked as low confidence adjustment. The step size of the low confidence adjustment is further reduced and multiple sampling cycles are forcibly waited before the next adjustment is executed.

[0096] In this embodiment, the adjustment command is based on the pre-calibrated sensitivity coefficient of thickness to each process parameter, and adopts an inverse proportional allocation principle for multi-parameter coordinated adjustment. When the thickness is too high, the ink droplet volume is reduced first (primary factor) and the vacuum pumping rate is increased (secondary factor). The adjustment amount is allocated inversely proportional to the sensitivity, so that the contribution of parameters to thickness is synchronous and coordinated, avoiding the risk of process fluctuation caused by large adjustment of a single parameter. The step size coefficient is set in stages according to the magnitude range of the prediction deviation (large step size / nominal step size / small step size), which ensures that the production line can quickly return to the target range when the thickness deviates significantly, and avoids repeated oscillations caused by over-adjustment when approaching the target value. Before executing closed-loop adjustment, the prediction variance value is called to determine the confidence level. When the variance exceeds the threshold, the adjustment is paused and an alarm is triggered. At the same time, the prediction values ​​at multiple sampling times are smoothed and filtered. Only when the values ​​still exceed the standard after smoothing are low-confidence adjustment commands generated, avoiding false adjustments caused by instantaneous sensor noise, external light interference, or occasional prediction fluctuations of the model, and reducing ineffective adjustments and equipment wear caused by misjudgment on the production line.

[0097] According to the solution provided by the present invention, for perovskite thin films prepared by roll-to-roll continuous inkjet printing, multiple training samples covering the process parameter tolerance window and the film thickness variation range are collected; simultaneously, the film thickness reference value, inkjet printing process parameter record, transmitted light image sequence at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signal of each training sample are acquired; wherein, the film thickness reference value is obtained using an offline white light interferometer as a supervision label for regression modeling; the inkjet printing process parameter record includes ink droplet volume setting value, printing substrate temperature value, droplet spacing setting value, and vacuum-assisted pumping rate as static process features; and the transmitted light image sequence is extracted from each... The grayscale mean, grayscale spatial standard deviation, grayscale gradient magnitude statistics, and local binary pattern texture features of a frame image are arranged in chronological order to form a dynamic optical feature time-series vector. The peak luminescence intensity, decay time constant, and peak wavelength shift of each channel in the multi-channel photoluminescence intensity time-series signal are extracted to form a photoluminescence dynamics feature vector. The static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector are concatenated in a multi-modal feature layer to form a multi-modal temporal fusion feature vector. The multi-modal temporal fusion feature vector and the corresponding thin film thickness reference value are used to form training samples to obtain a training dataset. The thin film thickness of each sample in the training dataset is used as the training dataset. The reference value is used as a supervision signal to update the network parameters of the multimodal temporal deep regression network model until the coefficient of determination on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold. The cross-modal attention parameters and temporal feature encoding parameters from the network parameters during training are saved. On a perovskite thin film roll-to-roll inkjet printing production line, the online inkjet printing process parameters of the film under test are acquired in real time. A transmitted light imaging device positioned before the curing station continuously acquires online transmitted light image sequences at multiple moments during the film formation process. A multi-channel photoluminescence monitoring device simultaneously acquires online multi-channel photoluminescence intensity temporal signals. The results for each moment are obtained from the online transmitted light image sequences. The online image feature vectors are arranged in chronological order to form an online dynamic optical feature time-series vector; an online photoluminescence dynamics feature vector is obtained based on the online multi-channel photoluminescence intensity time-series signal; the online inkjet printing process parameters are used as online static process features, and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal temporal fusion feature vector; the online multimodal temporal fusion feature vector is input into the multimodal temporal depth regression network model to output an online thin film thickness prediction value; the online thin film thickness prediction value is used as the non-destructive online thickness evaluation result of the perovskite thin film to be tested, and is fed back to the production line manufacturing execution system in real time;The multimodal temporal deep regression network model calls the stored cross-modal attention parameters to output an online-aligned multimodal feature vector, and calls the temporal feature encoding parameters to perform temporal encoding on the online-aligned multimodal feature vector along the time dimension, outputting an online temporal fusion feature vector; based on the online temporal fusion feature vector, an online thin film thickness prediction value is output. This invention avoids the "unreasonable extrapolation" problem when extrapolating process parameter combinations in a purely data-driven model, improving the model's generalization stability during long-term production line operation.

[0098] Figure 4 A schematic diagram of the framework of an online non-destructive thickness evaluation device for perovskite inkjet-printed films based on a regression model, according to an embodiment of the present invention, is shown. The online non-destructive thickness evaluation device for perovskite inkjet-printed films based on a regression model includes:

[0099] The sample acquisition module 410 is used to acquire multiple training samples covering the process parameter tolerance window and the film thickness variation range for perovskite films prepared by roll-to-roll continuous inkjet printing process; simultaneously acquires the film thickness reference value, inkjet printing process parameter record, transmitted light image sequence at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signal for each training sample; wherein, the film thickness reference value is acquired using an offline white light interferometer as a supervision label for regression modeling; the inkjet printing process parameter record includes ink droplet volume setting value, printing substrate temperature value, droplet spacing setting value, and vacuum-assisted pumping rate as static process features;

[0100] The feature extraction module 420 is used to extract the grayscale mean, grayscale spatial standard deviation, grayscale gradient amplitude statistics, and local binary pattern texture features of each frame in the transmitted light image sequence, and arrange them in chronological order to form a dynamic optical feature time-series vector; extract the luminescence intensity peak, decay time constant, and peak wavelength drift of each channel in the multi-channel photoluminescence intensity time-series signal to form a photoluminescence dynamics feature vector; and concatenate the static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector into a multi-modal feature layer to form a multi-modal temporal fusion feature vector. The multi-modal temporal fusion feature vector and the corresponding thin film thickness reference value are used to form training samples to obtain a training dataset.

[0101] The model training module 430 is used to update the network parameters of the multimodal temporal deep regression network model using the thin film thickness reference value of each sample in the training dataset as a supervision signal, until the determination coefficient on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold, and to save the cross-modal attention parameters and temporal feature encoding parameters in the network parameters during the training process.

[0102] The online acquisition module 440 is used to acquire online inkjet printing process parameters of the film under test in real time on a perovskite film roll-to-roll inkjet printing production line. A transmitted light imaging device positioned before the curing station continuously acquires online transmitted light image sequences at multiple moments during film formation. A multi-channel photoluminescence monitoring device simultaneously acquires online multi-channel photoluminescence intensity time-series signals. Based on the online transmitted light image sequences, online image feature vectors for each moment are obtained and arranged in chronological order to form an online dynamic optical feature time-series vector. An online photoluminescence dynamics feature vector is obtained based on the online multi-channel photoluminescence intensity time-series signals. The online inkjet printing process parameters are used as online static process features and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal time-series fusion feature vector.

[0103] The prediction feedback module 450 is used to input the online multimodal temporal fusion feature vector into the multimodal temporal deep regression network model, output an online thin film thickness prediction value, and use the online thin film thickness prediction value as the non-destructive online thickness evaluation result of the perovskite thin film to be tested, feeding it back to the production line manufacturing execution system in real time; wherein, the multimodal temporal deep regression network model calls the saved cross-modal attention parameters to output the online aligned multimodal feature vector, calls the temporal feature encoding parameters to perform temporal encoding on the online aligned multimodal feature vector along the time dimension, and outputs an online temporal fusion feature vector; and outputs the online thin film thickness prediction value based on the online temporal fusion feature vector.

[0104] Figure 5 The diagram shows a structural schematic of an embodiment of the computing device of the present invention. The specific embodiments of the present invention do not limit the specific implementation of the computing device.

[0105] like Figure 5 As shown, the computing device may include: a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0106] The processor 502, communication interface 504, and memory 506 communicate with each other via communication bus 508. Communication interface 504 is used to communicate with other network elements, such as clients or other servers. Processor 502 executes program 510, specifically performing the relevant steps in the above-described embodiment of the non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on regression models.

[0107] Specifically, program 510 may include program code that includes computer operation instructions.

[0108] Processor 502 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention. The computing device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0109] Memory 506 is used to store program 510. Memory 506 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0110] According to the solution provided by the present invention, for perovskite thin films prepared by roll-to-roll continuous inkjet printing, multiple training samples covering the process parameter tolerance window and the film thickness variation range are collected; simultaneously, the film thickness reference value, inkjet printing process parameter record, transmitted light image sequence at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signal of each training sample are acquired; wherein, the film thickness reference value is obtained using an offline white light interferometer as a supervision label for regression modeling; the inkjet printing process parameter record includes ink droplet volume setting value, printing substrate temperature value, droplet spacing setting value, and vacuum-assisted pumping rate as static process features; and the transmitted light image sequence is extracted from each... The grayscale mean, grayscale spatial standard deviation, grayscale gradient magnitude statistics, and local binary pattern texture features of a frame image are arranged in chronological order to form a dynamic optical feature time-series vector. The peak luminescence intensity, decay time constant, and peak wavelength shift of each channel in the multi-channel photoluminescence intensity time-series signal are extracted to form a photoluminescence dynamics feature vector. The static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector are concatenated in a multi-modal feature layer to form a multi-modal temporal fusion feature vector. The multi-modal temporal fusion feature vector and the corresponding thin film thickness reference value are used to form training samples to obtain a training dataset. The thin film thickness of each sample in the training dataset is used as the training dataset. The reference value is used as a supervision signal to update the network parameters of the multimodal temporal deep regression network model until the coefficient of determination on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold. The cross-modal attention parameters and temporal feature encoding parameters from the network parameters during training are saved. On a perovskite thin film roll-to-roll inkjet printing production line, the online inkjet printing process parameters of the film under test are acquired in real time. A transmitted light imaging device positioned before the curing station continuously acquires online transmitted light image sequences at multiple moments during the film formation process. A multi-channel photoluminescence monitoring device simultaneously acquires online multi-channel photoluminescence intensity temporal signals. The results for each moment are obtained from the online transmitted light image sequences. The online image feature vectors are arranged in chronological order to form an online dynamic optical feature time-series vector; an online photoluminescence dynamics feature vector is obtained based on the online multi-channel photoluminescence intensity time-series signal; the online inkjet printing process parameters are used as online static process features, and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal temporal fusion feature vector; the online multimodal temporal fusion feature vector is input into the multimodal temporal depth regression network model to output an online thin film thickness prediction value; the online thin film thickness prediction value is used as the non-destructive online thickness evaluation result of the perovskite thin film to be tested, and is fed back to the production line manufacturing execution system in real time;The multimodal temporal deep regression network model calls the stored cross-modal attention parameters to output an online-aligned multimodal feature vector, and calls the temporal feature encoding parameters to perform temporal encoding on the online-aligned multimodal feature vector along the time dimension, outputting an online temporal fusion feature vector; based on the online temporal fusion feature vector, an online thin film thickness prediction value is output. This invention avoids the "unreasonable extrapolation" problem when extrapolating process parameter combinations in a purely data-driven model, improving the model's generalization stability during long-term production line operation.

[0111] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination of all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed can be employed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose. Furthermore, those skilled in the art will understand that although some embodiments herein include certain features included in other embodiments but not others, combinations of features from different embodiments are intended to be within the scope of the invention and form different embodiments. For example, in the following claims, any of the claimed embodiments can be used in any combination. The invention can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. Unless otherwise specified, the steps in the above embodiments should not be construed as limiting the order of execution.

Claims

1. A non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model, characterized in that, include: For perovskite thin films prepared by roll-to-roll continuous inkjet printing, multiple training samples covering the process parameter tolerance window and the range of film thickness variation were collected. Simultaneously acquire the film thickness reference value, inkjet printing process parameter record, transmitted light image sequence at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signal for each training sample; wherein, the film thickness reference value is acquired using an offline white light interferometer as a supervision label for regression modeling; the inkjet printing process parameter record includes droplet volume setting value, printing substrate temperature value, droplet spacing setting value, and vacuum-assisted pumping rate as static process features; The mean gray level, standard deviation of gray level space, gray level gradient amplitude statistics, and local binary pattern texture features of each frame in the transmitted light image sequence are extracted and arranged in chronological order to form a dynamic optical feature time-series vector. The peak emission intensity, decay time constant, and peak wavelength drift of each channel in the multi-channel photoluminescence intensity time-series signal are extracted to form a photoluminescence dynamics feature vector. The static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector are concatenated in a multimodal feature layer to form a multimodal temporal fusion feature vector. The multimodal temporal fusion feature vector and the corresponding thin film thickness reference value are used to form training samples to obtain a training dataset. Using the thin film thickness reference value of each sample in the training dataset as a supervision signal, update the network parameters of the multimodal temporal deep regression network model until the determination coefficient on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold. Save the cross-modal attention parameters and temporal feature encoding parameters in the network parameters during training. On a roll-to-roll inkjet printing production line for perovskite thin films, online inkjet printing process parameters of the film under test are acquired in real time. A transmitted light imaging device positioned before the curing station continuously acquires online transmitted light image sequences at multiple moments during film formation. A multi-channel photoluminescence monitoring device simultaneously acquires online multi-channel photoluminescence intensity time-series signals. Based on the online transmitted light image sequences, online image feature vectors for each moment are obtained and arranged chronologically to form an online dynamic optical feature time-series vector. Based on the online multi-channel photoluminescence intensity time-series signals, an online photoluminescence dynamics feature vector is obtained. The online inkjet printing process parameters are used as online static process features and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal time-series fusion feature vector. The online multimodal temporal fusion feature vector is input into the multimodal temporal deep regression network model, which outputs an online thin film thickness prediction value. This online thin film thickness prediction value is used as the non-destructive online thickness evaluation result of the perovskite thin film under test and is fed back to the production line manufacturing execution system in real time. Specifically, the multimodal temporal deep regression network model calls the stored cross-modal attention parameters to output an online aligned multimodal feature vector, calls the temporal feature encoding parameters to perform temporal encoding on the online aligned multimodal feature vector along the time dimension, and outputs an online temporal fusion feature vector. The online thin film thickness prediction value is output based on the online temporal fusion feature vector.

2. The non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model according to claim 1, characterized in that, The multimodal temporal deep regression network model consists of an input feature normalization layer, a cross-modal temporal alignment module, a temporal feature encoding backbone network, a physical constraint residual correction module, and a thickness regression output head cascaded together. The input feature normalization layer is used to receive the online multimodal temporal fusion feature vector and output the standardized multimodal feature tensor. The cross-modal temporal alignment module is used to receive the standardized multimodal feature tensor, call the saved cross-modal attention parameters to perform learnable positional encoding of the dynamic optical feature temporal sub-vector along the time dimension, use the photoluminescence dynamic feature sub-vector as the query guide and the dynamic optical feature temporal sub-vector as the query content to perform cross-modal multi-head attention calculation, output the optical emission alignment feature sequence, map the static process feature sub-vector through a linear projection layer to a static extended vector with the same feature dimension as the alignment feature sequence, and copy and splice the static extended vector with the optical emission alignment feature sequence in the time dimension to obtain the time-aligned multimodal feature tensor; The temporal feature encoding backbone network includes a serial structure of stacked bidirectional recurrent memory network layers and temporal attention layers. It is used to receive the time-aligned multimodal feature tensor, call the saved temporal feature encoding parameters, extract temporal dependencies of the time-aligned multimodal feature tensor along the forward and reverse time directions through the bidirectional recurrent memory network layer, and output the hidden state sequence. The temporal attention layer is used to apply self-attention weights to the hidden state sequence along the time dimension and output an online temporal fusion feature vector; The physical constraint residual correction module is used to receive the online time-series fusion feature vector and the standardized static process feature sub-vector, calculate the physical prior thickness estimate using a pre-fitted surrogate model between the perovskite film thickness and the ink droplet volume set value, the substrate temperature and the vacuum pumping rate, concatenate the online time-series fusion feature vector and the physical prior thickness estimate, calculate the residual correction amount through the residual network, and output the physically corrected time-series fusion feature vector. The thickness regression output head is used to receive the physically corrected temporal fusion feature vector and map it into a scalar, and output the online thin film thickness prediction value.

3. The non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model according to claim 1, characterized in that, The collection of multiple training samples covering the process parameter tolerance window and the film thickness variation range further includes: Multiple sets of process parameter combinations are generated within the tolerance window of inkjet printing process parameters. The tolerance window at least covers the allowable deviation range of each of the droplet volume setting value, the printing substrate temperature value, the droplet spacing setting value, and the vacuum-assisted pumping rate. Each set of process parameter combinations is fully or approximately fully factored within the tolerance window according to at least three levels. For each combination of process parameters, samples are cut from multiple transverse positions along the film travel direction in the width direction in the same roll-to-roll continuous inkjet printing batch. Multiple sub-samples are cut continuously at fixed intervals along the longitudinal direction at each transverse position. Each combination of process parameters corresponds to multiple sub-samples, thus obtaining the total number of training samples.

4. The non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model according to claim 1, characterized in that, The acquisition process of the transmitted light image sequence further includes: During the thin film formation process, a transmission light imaging device synchronously triggers alternating or simultaneous exposure of each channel at a fixed sampling frequency to obtain a transmission light image sequence. Each frame of the transmission light image sequence is obtained by weighted fusion of multiple single-wavelength subframes at the pixel level. The weighting coefficients are pre-calibrated based on the absorption characteristics of the perovskite thin film at the corresponding wavelength. The transmission light imaging device employs a multi-wavelength light source array, which contains at least three independent controllable channels with different center wavelengths.

5. The non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model according to claim 1, characterized in that, The multi-channel photoluminescence monitoring device is equipped with a fiber optic spectrometer array with spectral resolution capability; The fiber optic spectrometer array consists of multiple detection channels, each corresponding to a characteristic wavelength range where different characteristic emission peaks of the perovskite thin film are located. Each detection channel independently records the signal of photoluminescence intensity changing over time at a sampling rate higher than the characteristic frequency of thin film crystallization kinetics.

6. The non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model according to claim 1, characterized in that, During the process of feeding back the online film thickness prediction value to the production line manufacturing execution system in real time, the production line manufacturing execution system generates a closed-loop adjustment command for the inkjet printing process parameters based on the deviation between the online film thickness prediction value and the thickness target value. Specifically, if the current online film thickness prediction value is higher than the upper limit of the thickness target value, the droplet volume setting value is reduced first, while the vacuum-assisted pumping rate is increased. The adjustment amount of reducing the droplet volume setting value and increasing the vacuum-assisted pumping rate is allocated inversely proportional to the sensitivity coefficient of each process parameter to the pre-calibrated thickness. If the current online film thickness prediction value is lower than the lower limit of the thickness target value, the droplet volume setting value is increased first and the substrate temperature is reduced. Specifically, when the absolute deviation between the online film thickness prediction value and the target thickness value is in different order of magnitude ranges, large step size, nominal step size, or small step size adjustment are adopted respectively, and the step size coefficient is set according to the deviation order of magnitude. Furthermore, before generating the closed-loop adjustment command, the production line manufacturing execution system first calls the prediction variance value output by the multimodal time-series deep regression network model. If the prediction variance value exceeds the preset variance threshold, the closed-loop adjustment is paused and an alarm signal is issued. The online film thickness prediction values ​​of multiple consecutive sampling times after the current time are smoothed and filtered. Only when the smoothed and filtered prediction value still exceeds the range of the target thickness value, the closed-loop adjustment command is generated and its confidence level is marked as low confidence adjustment. The step size of the low confidence adjustment is further reduced and multiple sampling cycles are forcibly waited before the next adjustment is executed.

7. The non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model according to claim 1, characterized in that, The calculation expression for the multimodal temporal fusion feature vector is as follows: in, This is the normalized static process feature vector; This represents the total number of sampling time frames for optical images and photoluminescence signals during the thin film curing process. The temporal attention weights for the optical image features of the k-th frame; The dynamic optical feature vector extracted from the transmitted light image of the k-th frame; The bandwidth of the time-series Gaussian smoothing kernel; This represents the fundamental eigenvectors of photoluminescence dynamics. For the outer product operation of feature tensors; This is a low-frequency Fourier filter operator for photoluminescence dynamics signals; wavelength Down Real-time photoluminescence intensity at any given moment; These are the extreme threshold values ​​of luminescence intensity at the corresponding wavelengths; This represents the fluorescence decay time constant corresponding to the characteristic wavelength; This is the second-order spatiotemporal Laplace correction term for the cross-modal characteristics; For Hadama accumulation.

8. The non-destructive thickness evaluation method for perovskite inkjet-printed thin films based on a regression model according to claim 2, characterized in that, The physical constraint residual correction module performs constraint correction for thickness prediction based on the thin film crystallization prior coupled residual regression formula, wherein the calculation expression of the thin film crystallization prior coupled residual regression formula is: in, This is the predicted final film thickness after constraint correction; The original predicted thickness is the output of network regression from the temporal fusion feature vector; The learning rate is used for residual correction. This is a priori thickness estimate based on the physical properties of the process. This is the positive mapping function for predicting thickness; This is a vector of real-time process parameters; This is the threshold value for the process parameter tolerance window; This is a step indicator function; For real-time prediction of the coefficient of determination; The minimum fitting accuracy threshold is preset; This is the vector of process standard reference parameters.

9. An online non-destructive thickness evaluation device for perovskite inkjet-printed films based on a regression model, characterized in that, include: The sample acquisition module is used to collect multiple training samples covering the process parameter tolerance window and the film thickness variation range for perovskite films prepared by roll-to-roll continuous inkjet printing process. Simultaneously acquire the film thickness reference value, inkjet printing process parameter record, transmitted light image sequence at multiple moments during film formation, and multi-channel photoluminescence intensity time-series signal for each training sample; wherein, the film thickness reference value is acquired using an offline white light interferometer as a supervision label for regression modeling; the inkjet printing process parameter record includes droplet volume setting value, printing substrate temperature value, droplet spacing setting value, and vacuum-assisted pumping rate as static process features; The feature extraction module is used to extract the grayscale mean, grayscale spatial standard deviation, grayscale gradient amplitude statistics, and local binary pattern texture features of each frame in the transmitted light image sequence, and arrange them in chronological order to form a dynamic optical feature time-series vector; extract the luminescence intensity peak, decay time constant, and peak wavelength drift of each channel in the multi-channel photoluminescence intensity time-series signal to form a photoluminescence dynamics feature vector; and concatenate the static process features, the dynamic optical feature time-series vector, and the photoluminescence dynamics feature vector into a multi-modal feature layer to form a multi-modal temporal fusion feature vector. The multi-modal temporal fusion feature vector and the corresponding thin film thickness reference value are used to form training samples to obtain a training dataset. The model training module is used to update the network parameters of the multimodal temporal deep regression network model using the thin film thickness reference value of each sample in the training dataset as a supervision signal, until the determination coefficient on the independent validation set is not lower than the first threshold and the root mean square error is not higher than the second threshold, and saves the cross-modal attention parameters and temporal feature encoding parameters in the network parameters during the training process. An online acquisition module is used to acquire online inkjet printing process parameters of the film under test in real time on a perovskite film roll-to-roll inkjet printing production line. A transmitted light imaging device positioned before the curing station continuously acquires online transmitted light image sequences at multiple moments during film formation. A multi-channel photoluminescence monitoring device simultaneously acquires online multi-channel photoluminescence intensity time-series signals. Based on the online transmitted light image sequences, online image feature vectors for each moment are obtained and arranged chronologically to form an online dynamic optical feature time-series vector. An online photoluminescence dynamics feature vector is obtained based on the online multi-channel photoluminescence intensity time-series signals. The online inkjet printing process parameters are used as online static process features and combined with the online dynamic optical feature time-series vector and the online photoluminescence dynamics feature vector to form an online multimodal time-series fusion feature vector. The prediction feedback module is used to input the online multimodal temporal fusion feature vector into the multimodal temporal deep regression network model, output the online thin film thickness prediction value, and use the online thin film thickness prediction value as the non-destructive online thickness evaluation result of the perovskite thin film to be tested, and feed it back to the production line manufacturing execution system in real time. Specifically, the multimodal temporal deep regression network model calls the saved cross-modal attention parameters to output the online aligned multimodal feature vector, calls the temporal feature encoding parameters to perform temporal encoding on the online aligned multimodal feature vector along the time dimension, and outputs the online temporal fusion feature vector; the online thin film thickness prediction value is output based on the online temporal fusion feature vector.

10. A computing device, comprising: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-described regression model-based non-destructive thickness evaluation method for perovskite inkjet-printed thin films.

Images