Perovskite thin film inkjet printing control method based on transformer in-situ monitoring model

Abstract

This invention discloses a control method and related apparatus for inkjet printing of perovskite thin films based on a Transformer in-situ monitoring model. The method includes: inputting a time-series evolution signal into the Transformer in-situ monitoring model to capture transient nucleation signals in the early stage of crystallization and grain growth transformation signals in the middle stage, extracting multi-stage non-stationary deep features; mapping these multi-stage non-stationary deep features to conditional probability distribution parameters of device performance indicators and extracting expected values ​​as predicted performance indicators; calculating the joint quantization value corresponding to the predicted performance indicators as the predicted uncertainty quantization value; calculating the deviation between the predicted performance indicators and a preset target performance indicator threshold to obtain the performance deviation; generating an inkjet printing control compensation signal based on the predicted uncertainty quantization value; and co-optimizing and controlling the droplet volume variation coefficient and droplet position. This invention achieves adaptive control driven by the crystallization process in the inkjet printing of perovskite thin films, significantly improving film quality consistency and production line stability.

Description

Technical Field

[0001] This invention relates to the field of perovskite thin film inkjet printing control technology, specifically to a perovskite thin film inkjet printing control method, apparatus, and computing device based on the Transformer in-situ monitoring model. Background Technology

[0002] Inkjet printing, with its advantages of non-contact patterned deposition, high material utilization, flexibility, and large-scale scalability, has become a key technology for the industrial production of perovskite thin films. However, the quality of perovskite thin films prepared by inkjet printing is highly dependent on the coordinated matching of multiple dynamic and non-stationary processes during printing, such as solvent evaporation, mesophase transformation, and halogen component homogenization. Among these, the vacuum quenching crystallization stage is particularly critical, determining the grain size, orientation, defect state density, and photoelectric property uniformity of the final film. Existing production lines generally adopt an open-loop control mode of offline sampling and experience-based correction. That is, after printing, samples are taken for X-ray diffraction, scanning electron microscopy, or photoluminescence characterization, and printing parameters are manually adjusted based on the test results. This approach cannot promptly capture crystallization path drift caused by environmental fluctuations (temperature, air pressure), printhead aging, and ink batch differences, resulting in large fluctuations in yield and significant material waste.

[0003] Currently, process monitoring and control of inkjet-printed perovskite thin films mainly include: (1) Open-loop or semi-closed-loop control methods based on traditional optical feature threshold judgment, using CCD cameras to collect changes in the brightness or color of the thin film surface, and judging the crystallization endpoint or abnormality by setting a fixed threshold. The perovskite crystallization process is accompanied by complex mesophase evolution, making it impossible to distinguish between different failure modes such as "normal nucleation" and "abnormal homogenization delay", which is very easy to cause misjudgment. (2) Parameter adjustment methods based on physical models or empirical kinetic equations, establishing simplified differential equations of solvent evaporation rate and grain growth rate to adjust vacuum degree or substrate temperature, but the crystallization path of perovskite systems (such as FA / MA mixed cations, Br / I mixed halogens) is component sensitive, and is severely mismatched under complex industrial disturbances, resulting in poor control robustness.

[0004] To address the aforementioned issues, this invention proposes a perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model. The method generates an inkjet printing control compensation signal based on the quantified value of predicted uncertainty, and collaboratively optimizes and controls the droplet volume variation coefficient and droplet position to adaptively control the crystallization process driven by the perovskite thin film inkjet printing process, thereby improving film quality consistency and production line stability. Summary of the Invention

[0005] In view of the above problems, the present invention provides a control method and apparatus for inkjet printing of perovskite thin films based on the Transformer in-situ monitoring model, as well as a computing device.

[0006] According to one aspect of the present invention, a method for controlling perovskite thin film inkjet printing based on a Transformer in-situ monitoring model is provided, comprising:

[0007] During the vacuum quenching and crystallization stage of perovskite thin film inkjet printing, the time-series evolution signal of the entire crystallization cycle was simultaneously acquired by a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system.

[0008] The time-series evolution signal is input into the Transformer in-situ monitoring model to capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage. Multi-stage non-stationary deep features caused by solvent volatilization kinetics, intermediate phase transformation path and halogen homogenization process during perovskite crystallization are extracted. The multi-stage non-stationary deep features are mapped to the conditional probability distribution parameters of device performance indicators. The expected value is extracted from the conditional probability distribution parameters as the predicted performance indicator. The joint quantization value corresponding to the predicted performance indicator is calculated as the prediction uncertainty quantization value.

[0009] The predicted performance index is deviated from the preset target performance index threshold to obtain the performance deviation amount. An inkjet printing control compensation signal is generated based on the predicted uncertainty quantification value. If the predicted uncertainty quantification value is lower than the preset uncertainty tolerance upper limit, the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value of each inkjet printhead are calculated based on the dynamic confidence weight determined by the performance deviation amount and the predicted uncertainty quantification value, and an inkjet printing control compensation signal is generated. If the predicted uncertainty quantification value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantification value converges to below the uncertainty tolerance upper limit.

[0010] The inkjet printing control compensation signal is sent to the embedded motion controller of the inkjet printing equipment in real time. The amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead are adjusted according to the piezoelectric drive voltage compensation value. The spatial synchronization interpolation parameters of the ejection trigger phase of each inkjet printhead and the substrate motion trajectory are dynamically adjusted according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold.

[0011] In an alternative approach, the simultaneous acquisition of time-series evolution signals throughout the crystallization cycle using a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system further includes:

[0012] At the transient zero moment of the start of the vacuum quenching and crystallization stage, the embedded industrial control computer simultaneously issues an external trigger synchronous acquisition command to the time-resolved photoluminescence spectrometer and the diffuse reflectance imaging system. The external trigger synchronous acquisition command contains a preset sampling frequency and a sampling time window length. The sampling time window length is pre-calibrated based on the saturated vapor pressure of the perovskite precursor solvent and the pumping speed of the vacuum chamber to ensure coverage of the entire crystallization cycle from supersaturated nucleation to solid film formation.

[0013] The time-resolved photoluminescence spectrometer, in response to the external trigger synchronous acquisition command, applies pulsed laser excitation to the surface of the perovskite thin film at a preset sampling frequency. A time-correlated single-photon counting process is performed by the single-photon counting module, outputting a photoluminescence intensity time-series sequence continuously recorded at nanosecond-level time resolution. Simultaneously, the diffuse reflection imaging system, in response to the same external trigger synchronous acquisition command, activates a CMOS camera to continuously capture frames of the thin film surface under broadband illumination, acquiring a diffuse reflection image sequence. An image preprocessing module calculates the grayscale mean for each pixel region of each frame image, outputting a diffuse reflection intensity time-series sequence aligned with the timestamp of the photoluminescence intensity time-series sequence.

[0014] The photoluminescence intensity time series and the diffuse reflection intensity time series are synchronized by bilinear interpolation and spatial coordinate matching using a unified timestamp index, forming a multimodal synchronization time series signal matrix as the time series evolution signal for the entire crystallization cycle; wherein, each row of the multimodal synchronization time series signal matrix corresponds to a discrete sampling time, and each column carries the photoluminescence intensity value and diffuse reflection intensity value of the discrete sampling time, respectively.

[0015] In one alternative approach, the Transformer in-situ monitoring model includes:

[0016] The input embedding layer is used to perform local temporal feature extraction and dimensional projection on the multimodal signal vector at each sampling time in the multimodal synchronous temporal signal matrix through a one-dimensional convolutional layer to generate an embedded feature vector; a learnable positional encoding vector is superimposed on the embedded feature vector to inject temporal order information, and the output is a feature embedding sequence with positional encoding that preserves global temporal dependencies.

[0017] A multi-head self-attention encoder is used to calculate the attention weight matrix between the query vector at each time step and the key vector at all time steps of the feature embedding sequence; wherein, the attention weight matrix adaptively assigns a higher aggregation weight to the early transient nucleation signal of crystallization than that of the mid-term grain growth transformation signal based on the correlation distribution of the photoluminescence intensity time sequence and the diffuse reflection intensity time sequence on short-range exciton relaxation and long-range solvent diffusion, so as to amplify the key crystallization events and obtain a context-enhanced feature sequence;

[0018] The temporal feature aggregation module is used to receive the context-enhanced feature sequence, perform max pooling and average pooling operations in parallel along the time dimension, and concatenate the dual pooling results along the feature channel dimension to form a fixed-dimensional feature vector. The fixed-dimensional feature vector encodes the multi-stage non-stationary deep features of the entire perovskite crystallization process, including early solvent volatilization feature components, mid-term feature components dominated by the intermediate phase transformation path, and late-term feature components dominated by the halogen homogenization process.

[0019] The probability prediction head module is used to receive the fixed-dimensional feature vector, and after dimensionality reduction and activation layer by layer through a multi-layer fully connected network, output two parallel branches. One branch generates the mean parameter of the device performance index through linear mapping, and the other branch generates a positive variance parameter through Softplus activation. The conditional probability distribution parameters of the device performance index are constructed based on the mean parameter and the positive variance parameter.

[0020] In one alternative approach, adjusting the amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead according to the piezoelectric drive voltage compensation value further includes:

[0021] The piezoelectric drive voltage compensation value is decoupled in the digital domain into an amplitude compensation component and a rising edge slope compensation component. The amplitude compensation component is used to set the correction amount of the peak voltage of the main pulse of the drive waveform, and the rising edge slope compensation component is used to set the correction amount of the transition time required for the drive waveform to climb from the baseline voltage to the peak voltage of the main pulse.

[0022] The embedded motion controller transmits the amplitude compensation component to the built-in digital-to-analog converter to generate an analog amplitude adjustment voltage. After the analog amplitude adjustment voltage is amplified by the drive stage consisting of a high-voltage operational amplifier and a push-pull output circuit, the peak voltage of the main pulse of the piezoelectric drive waveform is changed to the target amplitude in real time.

[0023] The embedded motion controller dynamically adjusts the pulse width modulation duty cycle of the rising edge segment of the drive waveform and the resistance value of the drive resistor according to the rising edge slope compensation component to obtain the adjusted rising edge slope, so as to control the length of the transition time.

[0024] The target amplitude and the adjusted rising edge slope parameter are written into the waveform register corresponding to each inkjet printhead. At the beginning of the next ejection cycle, the waveform generator is triggered, and the piezoelectric actuator is driven to execute ink droplet ejection according to the updated waveform parameters.

[0025] In one alternative approach, dynamically adjusting the spatial synchronization interpolation parameters between the ejection trigger phase of each inkjet printhead and the substrate motion trajectory based on the inkjet timing phase delay compensation value further includes:

[0026] The inkjet timing phase delay compensation value is decomposed into an ejection trigger phase offset and a spatial synchronization interpolation offset. The ejection trigger phase offset is used to correct the amount by which the ejection time of a single inkjet printhead is advanced or delayed relative to the system synchronization clock. The spatial synchronization interpolation offset is used to correct the amount by which the position of the interpolation point of the substrate motion trajectory corresponding to the ejection time is shifted forward or backward along the motion direction.

[0027] The embedded motion controller sends the jet trigger phase offset to the field programmable gate array timing generator to generate an independently phase-adjustable jet trigger pulse for each inkjet printhead. The phase delay relative to the system synchronization clock is locked as the jet trigger phase offset. The jet trigger pulse is routed to the piezoelectric drive circuit of the corresponding inkjet printhead as a jet enable signal.

[0028] When the embedded motion controller calculates the substrate motion trajectory in real time based on the spatial synchronous interpolation offset, it dynamically corrects the spatial step size and interpolation point position coordinates in each interpolation cycle.

[0029] The ejection trigger pulse and the spatial synchronization interpolation parameter are anchored on the time axis to ensure that each ejection trigger pulse instantaneously triggers ink droplet ejection at the corresponding corrected interpolation point position coordinates. This establishes a spatiotemporal coordinated control relationship between the ejection trigger phase offset and the spatial synchronization interpolation offset, thereby achieving dynamic synchronous matching between the ejection time and the substrate movement position to optimize the accuracy of the ink droplet landing point position.

[0030] In an alternative approach, decomposing the inkjet timing phase delay compensation value into an ejection trigger phase offset and a spatial synchronization interpolation offset further includes:

[0031] After receiving the inkjet timing phase delay compensation value, the embedded motion controller inputs the inkjet timing phase delay compensation value to the built-in coupling state discriminator. Based on the current substrate movement speed, the spatial layout spacing of the inkjet printhead array, and the real-time ink droplet flight time, it calculates the proportion of the temporal component caused by the jet triggering moment drift and the proportion of the spatial component caused by the landing point position system deviation in the inkjet timing phase delay compensation value.

[0032] The embedded motion controller reads the dominant partition label of the current crystallization stage in real time from the Transformer in-situ monitoring model. The dominant partition label includes the solvent evaporation dominant region, the mesophase transformation dominant region, and the halogen homogenization dominant region. Based on the dominant partition label, the adaptive decoupling factor is dynamically calculated and adjusted in the time domain bias direction to increase the allocation weight of the jet trigger phase bias.

[0033] The jet trigger phase offset is obtained based on the inkjet timing phase delay compensation value and the adaptive decoupling factor; the proportion of the time domain component and the proportion of the spatial domain component are input to the interpolation resolution quantizer, and the spatial synchronous interpolation offset expressed as an integer multiple of the interpolation resolution is output with the substrate motion trajectory interpolation resolution as the minimum quantization step size.

[0034] In one alternative approach, the multi-head self-attention encoder dynamically solves for the adaptive aggregation weights of the crystallized full-cycle multimodal signal using a spatiotemporal coupling correlation entropy formula, wherein the spatiotemporal coupling correlation entropy formula is:

[0035]

[0036] in, For the first The first attention head Adaptive feature aggregation weights at sampling time; , where is the temporal entropy of the photoluminescence signal, characterizing the temporal disorder of crystallization exciton relaxation; for Time-normalized photoluminescence intensity; , where is the spatial entropy of the diffuse reflection signal, characterizing the spatial inhomogeneity of the thin film surface morphology evolution; for Time-normalized diffuse reflectance intensity; For the first Attention head Gaussian kernel bandwidth hyperparameter; The temporal spatial modal coupling correlation coefficient is used to quantify the degree of correlation between dual-mode signals during the nucleation, grain growth, and homogenization stages. This is the adaptive gain coefficient for the crystallization stage; This represents the total number of sampling moments within the current sliding time window.

[0037] In one alternative approach, the formula for calculating the quantified value of the prediction uncertainty is:

[0038]

[0039] in, , which is the spatiotemporal joint feature entropy; for Time of the first Normalized activation percentage of crystalline-like features; The total number of predefined crystallization feature modes for the Transformer model; , where Gaussian dispersion of the Transformer characteristic manifold; Adaptive kernel width for manifold; The Gaussian conditional probability distribution of device performance indicators; The Frobenius norm of the temporal gradient matrix is ​​used to quantize the instantaneous drift intensity of the performance probability distribution. This represents the number of valid temporal feature sampling points within the current adaptive sliding time window. , These are the Transformer deep crystallization feature vectors corresponding to the i-th and j-th different temporal sampling times within the window, respectively; This is the Bayesian iterative regularization factor; This is the convergence threshold bias constant; The convergence gain coefficient for model iteration; This represents the number of convergence steps for continuous time-series sampling. It is a crystallization kinetics constraint factor; This refers to the real-time crystallization relaxation time.

[0040] According to another aspect of the present invention, a smart control device for perovskite thin film inkjet printing based on a Transformer in-situ monitoring model is provided, comprising:

[0041] The timing signal synchronous acquisition module is used to synchronously acquire the timing evolution signal of the entire crystallization cycle through a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system during the vacuum quenching and crystallization stage of perovskite thin film inkjet printing.

[0042] The deep feature extraction and performance prediction module is used to input the time-series evolution signal into the Transformer in-situ monitoring model, capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage, and extract the multi-stage non-stationary deep features caused by solvent volatilization kinetics, intermediate phase transformation path and halogen homogenization process during the perovskite crystallization process; map the multi-stage non-stationary deep features into conditional probability distribution parameters of device performance indicators, extract the expected value from the conditional probability distribution parameters as the predicted performance indicator, and calculate the joint quantization value corresponding to the predicted performance indicator as the prediction uncertainty quantization value;

[0043] The deviation calculation and adaptive compensation decision module is used to calculate the deviation between the predicted performance index and the preset target performance index threshold to obtain the performance deviation amount, and generate an inkjet printing control compensation signal based on the predicted uncertainty quantification value. Specifically, if the predicted uncertainty quantification value is lower than a preset uncertainty tolerance upper limit, the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value for each inkjet printhead are calculated based on the dynamic confidence weight determined by the performance deviation amount and the predicted uncertainty quantification value, and an inkjet printing control compensation signal is generated. If the predicted uncertainty quantification value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantification value converges to below the uncertainty tolerance upper limit.

[0044] The piezoelectric drive and motion trajectory coordinated control module is used to send the inkjet printing control compensation signal to the embedded motion controller of the inkjet printing device in real time. It adjusts the amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead according to the piezoelectric drive voltage compensation value, and dynamically adjusts the spatial synchronization interpolation parameters of the ejection trigger phase of each inkjet printhead and the substrate motion trajectory according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold.

[0045] 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;

[0046] The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model.

[0047] According to the solution provided by the present invention, during the vacuum quenching and crystallization stage of perovskite thin film inkjet printing, the time-series evolution signal of the entire crystallization cycle is simultaneously acquired by a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system; the time-series evolution signal is input into the Transformer in-situ monitoring model to capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage, and to extract the multi-stage non-stationary deep features caused by solvent evaporation kinetics, intermediate phase transformation path and halogen homogenization process during perovskite crystallization; the multi-stage non-stationary deep features are mapped to conditional probability distribution parameters of device performance indicators, and the expected value is extracted from the conditional probability distribution parameters as the predicted performance indicator, and the joint quantization value corresponding to the predicted performance indicator is calculated as the prediction uncertainty quantization value; the deviation between the predicted performance indicator and the preset target performance indicator threshold is calculated to obtain the performance deviation amount, and an inkjet printing control compensation signal is generated according to the prediction uncertainty quantization value; wherein, if the prediction uncertainty quantization value is lower than the preset uncertainty tolerance upper limit, according to The dynamic reliability weight determined by the performance deviation and the predicted uncertainty quantification value is used to calculate the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value for each inkjet printhead, generating an inkjet printing control compensation signal. If the predicted uncertainty quantification value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantification value converges to below the uncertainty tolerance upper limit. The inkjet printing control compensation signal is sent to the embedded motion controller of the inkjet printing device in real time. The amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead are adjusted according to the piezoelectric drive voltage compensation value, and the spatial synchronization interpolation parameters of the ejection trigger phase and substrate motion trajectory of each inkjet printhead are dynamically adjusted according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold. This invention provides adaptive control for the crystallization process driven by inkjet printing of perovskite thin films, which significantly improves the consistency of film quality and the stability of the production line.

[0048] 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

[0049] 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:

[0050] Figure 1 A flowchart illustrating the perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model according to an embodiment of the present invention is shown.

[0051] Figure 2a A schematic diagram illustrating the internal inference and execution mapping of the Transformer model according to an embodiment of the present invention is shown;

[0052] Figure 2b The timing diagram of deviation calculation and real-time control compensation according to an embodiment of the present invention is shown;

[0053] Figure 3 A topological diagram of perovskite thin film inkjet printing according to an embodiment of the present invention is shown;

[0054] Figure 4 This invention presents a schematic diagram of the framework of an intelligent control device for perovskite thin film inkjet printing based on the Transformer in-situ monitoring model according to an embodiment of the present invention.

[0055] Figure 5 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. Detailed Implementation

[0056] 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.

[0057] Figure 1 A schematic flowchart of a perovskite thin film inkjet printing control method based on a Transformer in-situ monitoring model, 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:

[0058] In step S101, during the vacuum quenching and crystallization stage of perovskite thin film inkjet printing, the time-series evolution signal of the entire crystallization cycle is simultaneously acquired by a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system.

[0059] In this embodiment, pulsed laser excitation and broadband light source illumination are used, both acting on the thin film surface in a non-contact and non-destructive manner, without affecting the natural crystallization process. Meanwhile, the spectrometer and imaging system are both industrial-grade standardized optical modules that can be integrated outside the observation window of the vacuum quenching chamber without compromising the chamber's sealing or modifying the core process structure, offering excellent production line deployability and maintainability.

[0060] Specifically, at the transient zero moment of the start of the vacuum quenching crystallization stage, the embedded industrial control computer generates a rising-edge hard trigger signal through its built-in timer / counter board. After passing through a TTL level conversion circuit, this signal is simultaneously distributed to the external trigger input port of the time-resolved photoluminescence spectrometer and the external trigger input port of the diffuse reflection imaging system via a coaxial cable. The sampling frequency is pre-set based on the expected total crystallization time and the time resolution requirements of the subsequent model (typically 10 Hz to 50 Hz, balancing dynamic capture accuracy and data storage burden). Before formal production, a preliminary experiment is conducted using the same batch of perovskite precursor ink and the same vacuum chamber pumping speed conditions. The total time from vacuum start-up to the stabilization of the diffuse reflection intensity on the film surface (change rate between adjacent frames <1%) is recorded. This time is multiplied by a safety factor of 1.2 and then written into the window length field of the trigger command to ensure coverage of the entire crystallization cycle and allowance for redundancy. Upon receiving an external trigger command, the spectrometer immediately applies a pulsed laser (center wavelength 405 nm, repetition frequency consistent with the sampling frequency) vertically to the perovskite film surface for excitation. The single-pulse energy is controlled within the range of 0.5–2 nJ to avoid photodegradation or heating effects. The transient photoluminescence signal generated by the excitation is collected via optical fiber and transmitted to the single-photon counting module (TCSPC module), which operates in a time-correlated single-photon counting mode, accumulating a histogram of photon arrival times within each laser pulse cycle. The output at each sampling moment is a complete fluorescence decay curve. To form a time sequence, the decay curve is further integrated or the peak intensity is extracted to generate the photoluminescence intensity value at that moment. Upon receiving the same external trigger command, the diffuse reflection imaging system synchronously activates its CMOS area array camera (frame rate consistent with the sampling frequency). Under uniform illumination from a broadband white LED ring light source, the camera continuously captures frames of the same detection area on the film surface. The exposure time is automatically adjusted to ensure that the grayscale value is not saturated (typical target grayscale median value is 150–200 / 255). After acquiring each image frame, median filtering (3×3 window) is applied to reduce noise. Valid pixels are extracted according to the preset region of interest (ROI, covering the actual film-forming area of ​​the inkjet print). The grayscale value is accumulated pixel by pixel and divided by the total number of valid pixels to obtain the average diffuse reflectance intensity of that frame. This average value is aligned with the timestamp of the photoluminescence intensity time series and then arranged chronologically to form a diffuse reflectance intensity time series.

[0061] For example, the precursor solvent was set to DMF:DMSO = 4:1 (volume ratio), the saturated vapor pressure at 25°C was approximately 0.5 kPa, and the vacuum chamber pumping speed was 10 L / s. Preliminary experiments determined the total time required from vacuum start-up to the stabilization of the solid film's diffuse reflection to be 110 seconds. Multiplying this by a safety factor of 1.2, the sampling time window length was set to 132 seconds. The sampling frequency was set to 20 Hz, resulting in a total of 2640 sampling points. At the instant of vacuum quenching start-up (t=0 s), the embedded industrial control computer issued an external trigger command. The spectrometer emitted 405 nm laser pulses at a repetition frequency of 20 Hz. The TCSPC module collected photon counts within a 200 ns window after each pulse, integrating the PL intensity at that moment. Simultaneously, a CMOS camera continuously captured images of the thin film surface at 20 fps (ROI of 10 mm × 10 mm, corresponding to the center area of ​​the inkjet-printed pattern). Acquisition ended at t=132 s, forming two time-series sequences, each 2640 seconds long. At t=5.2 s, the PL sequence shows a sharp drop in intensity (corresponding to enhanced nonradiative exciton recombination caused by extensive nucleation), while the diffuse reflectance sequence shows an increase in grayscale value (increased surface roughness leading to enhanced scattering). This moment is recorded in both sequences and, after interpolation and synchronization, is paired as (PL=2.1×10). 4 The counts (DR=152 gray) are used as the 104th row of the signal matrix (because 5.2 s / 0.05 s = 104), which fully records the spatiotemporal coordination information of the early nucleation events.

[0062] In an alternative approach, the simultaneous acquisition of time-series evolution signals throughout the crystallization cycle using a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system further includes:

[0063] At the transient zero moment of the start of the vacuum quenching and crystallization stage, the embedded industrial control computer simultaneously issues an external trigger synchronous acquisition command to the time-resolved photoluminescence spectrometer and the diffuse reflectance imaging system. The external trigger synchronous acquisition command contains a preset sampling frequency and a sampling time window length. The sampling time window length is pre-calibrated based on the saturated vapor pressure of the perovskite precursor solvent and the pumping speed of the vacuum chamber to ensure coverage of the entire crystallization cycle from supersaturated nucleation to solid film formation.

[0064] The time-resolved photoluminescence spectrometer, in response to the external trigger synchronous acquisition command, applies pulsed laser excitation to the surface of the perovskite thin film at a preset sampling frequency. A time-correlated single-photon counting process is performed by the single-photon counting module, outputting a photoluminescence intensity time-series sequence continuously recorded at nanosecond-level time resolution. Simultaneously, the diffuse reflection imaging system, in response to the same external trigger synchronous acquisition command, activates a CMOS camera to continuously capture frames of the thin film surface under broadband illumination, acquiring a diffuse reflection image sequence. An image preprocessing module calculates the grayscale mean for each pixel region of each frame image, outputting a diffuse reflection intensity time-series sequence aligned with the timestamp of the photoluminescence intensity time-series sequence.

[0065] The photoluminescence intensity time series and the diffuse reflection intensity time series are synchronized by bilinear interpolation and spatial coordinate matching using a unified timestamp index, forming a multimodal synchronization time series signal matrix as the time series evolution signal for the entire crystallization cycle; wherein, each row of the multimodal synchronization time series signal matrix corresponds to a discrete sampling time, and each column carries the photoluminescence intensity value and diffuse reflection intensity value of the discrete sampling time, respectively.

[0066] In this embodiment, as Figure 3As shown, time-correlated single-photon counting (TCSPC) technology enables photoluminescence spectrometers to continuously record exciton recombination dynamics signals with nanosecond-level time resolution. Transient nucleation events in the early stages of perovskite crystallization are often accompanied by abrupt changes in exciton lifetimes on the picosecond to nanosecond scale (such as a surge in nonradiative recombination centers leading to fluorescence quenching at the moment of nucleation). Conventional spectrometers with millisecond-level sampling cannot distinguish such transient characteristics. By using histogram accumulation in TCSPC mode and high-repetition-rate pulsed laser excitation, each nucleation pulse event is precisely locked to the corresponding timestamp in the time sequence, providing high-fidelity data. The embedded industrial control computer simultaneously issues external trigger commands to both devices at the transient zero moment of vacuum quenching startup, instead of using software handshake or asynchronous startup methods. The transmission delay of this hard trigger signal is controlled at the nanosecond level (<100 ns), far less than the sampling interval (milliseconds to tens of milliseconds), fundamentally eliminating the stage misalignment problem caused by the startup time difference in the traditional "spectrometer first, camera later" scheme. The dual-modal signals are strictly aligned on the same physical time axis, ensuring the accurate preservation of the causal relationship between "exciton behavior changes" and "surface morphology evolution" during crystallization. This avoids misjudging the temporal sequence of nucleation and grain growth due to inconsistent time bases. The evaporation rates of different solvent systems (such as DMF, DMSO, NMP, and their mixtures) can vary by several times, and the vacuum chamber pumping speed directly determines the rate of decrease in environmental supersaturation. A property-driven calibration method automatically matches the window length to the actual crystallization cycle under different ink formulations and process conditions. This avoids both excessively short windows that would truncate the homogenization signal later and excessively long windows that introduce invalid baseline data and increase computational burden. The diffuse reflection intensity acquired by the diffuse reflection imaging system is calculated pixel by pixel by the image preprocessing module and matched with the local average value of the field of view of the photoluminescence spectrometer. This ensures that the photoluminescence signal (from an area with a field of view diameter of about 2 mm in the fiber optic cable) and the diffuse reflection signal (from a set of pixels corresponding to the same physical region in the CMOS image) originate from the same micro-region of the thin film, thus avoiding misjudging nucleation differences at different locations as temporal fluctuations due to spatial misalignment.

[0067] Step S102: Input the time-series evolution signal into the Transformer in-situ monitoring model to capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage. Extract the multi-stage non-stationary deep features caused by solvent volatilization kinetics, intermediate phase transformation path and halogen homogenization process during the perovskite crystallization process. Map the multi-stage non-stationary deep features to the conditional probability distribution parameters of the device performance index. Extract the expected value from the conditional probability distribution parameters as the predicted performance index. Calculate the joint quantization value corresponding to the predicted performance index as the prediction uncertainty quantization value.

[0068] In this embodiment, by processing the time-series evolution signal through the Transformer in-situ monitoring model, it is possible to simultaneously identify transient nucleation signals in the early stage of crystallization and grain growth transformation signals in the middle stage. It can extract non-stationary deep features generated by the coupling of multiple physicochemical processes such as solvent evaporation, intermediate phase transformation, and halogen homogenization, enabling monitoring to cover the entire crystallization lifecycle and avoiding the loss of key intermediate state information. The deep features are mapped to conditional probability distribution parameters (mean and variance) of device performance indicators, providing both expected values ​​as predicted performance indicators and calculating a joint quantized value as predicted uncertainty. This allows operators to simultaneously obtain the "optimal estimate" and the "degree of confidence in that estimate," avoiding blindly executing controls when model predictions are unreliable. The quantified value of predicted uncertainty can be actively adjusted when the prediction is reliable and adaptively extended when the prediction is unreliable, improving the safety of the control strategy, especially suitable for vacuum quenching crystallization scenarios with high noise and strong dynamics. The performance deviation and uncertainty weights can be dynamically adjusted according to the current prediction confidence level to control the strength of the compensation signal, avoiding excessive or unstable adjustments when the model is fuzzy. This effectively improves the convergence speed of device performance while suppressing process fluctuations caused by incorrect adjustments.

[0069] In one alternative approach, the Transformer in-situ monitoring model includes:

[0070] The input embedding layer is used to perform local temporal feature extraction and dimensional projection on the multimodal signal vector at each sampling time in the multimodal synchronous temporal signal matrix through a one-dimensional convolutional layer to generate an embedded feature vector; a learnable positional encoding vector is superimposed on the embedded feature vector to inject temporal order information, and the output is a feature embedding sequence with positional encoding that preserves global temporal dependencies.

[0071] A multi-head self-attention encoder is used to calculate the attention weight matrix between the query vector at each time step and the key vector at all time steps of the feature embedding sequence; wherein, the attention weight matrix adaptively assigns a higher aggregation weight to the early transient nucleation signal of crystallization than that of the mid-term grain growth transformation signal based on the correlation distribution of the photoluminescence intensity time sequence and the diffuse reflection intensity time sequence on short-range exciton relaxation and long-range solvent diffusion, so as to amplify the key crystallization events and obtain a context-enhanced feature sequence;

[0072] The temporal feature aggregation module is used to receive the context-enhanced feature sequence, perform max pooling and average pooling operations in parallel along the time dimension, and concatenate the dual pooling results along the feature channel dimension to form a fixed-dimensional feature vector. The fixed-dimensional feature vector encodes the multi-stage non-stationary deep features of the entire perovskite crystallization process, including early solvent volatilization feature components, mid-term feature components dominated by the intermediate phase transformation path, and late-term feature components dominated by the halogen homogenization process.

[0073] The probability prediction head module is used to receive the fixed-dimensional feature vector, and after dimensionality reduction and activation layer by layer through a multi-layer fully connected network, output two parallel branches. One branch generates the mean parameter of the device performance index through linear mapping, and the other branch generates a positive variance parameter through Softplus activation. The conditional probability distribution parameters of the device performance index are constructed based on the mean parameter and the positive variance parameter.

[0074] In this embodiment, by using one-dimensional convolution and learnable positional encoding in the input embedding layer, local temporal patterns can be automatically extracted directly from the original bimodal temporal signal while preserving global order dependencies. This avoids information loss and subjective bias caused by relying on manually designed statistical features (such as peak time and decay rate), achieving end-to-end mapping from the original signal to high-dimensional semantic features. The multi-head self-attention encoder does not employ a general attention distribution. Instead, based on the physical correlation between photoluminescence and diffuse reflection signals in short-range exciton relaxation and long-range solvent diffusion, it adaptively assigns higher aggregation weights to transient nucleation signals in the early stages of crystallization. This proactively focuses on key crystallization events that truly determine the final device performance, suppressing interference from noise in non-critical periods and significantly improving the fidelity of key transient feature extraction. The probabilistic prediction head module outputs the mean and variance of the conditional probability distribution, extending deterministic prediction to probabilistic inference. It not only predicts the result but also provides the degree of certainty of the prediction, offering rigorous statistical data for controlling the confidence weighting of the compensation signal and adaptive window adjustment.

[0075] For example, the input embedding layer includes a one-dimensional convolutional layer and a learnable positional encoding layer. The one-dimensional convolutional kernel (e.g., kernel widths of 3, 5, 7, etc.) performs local temporal feature extraction and dimensional projection on the multimodal signal vector (photoluminescence intensity and diffuse reflection intensity) at each sampling time along the time dimension. This effectively captures local change patterns between adjacent time steps (e.g., the rate of increase in luminescence intensity, the curvature of reflectance change, etc.) while mapping the original dual-channel input to a high-dimensional hidden space. The learnable positional encoding superimposes a learnable positional encoding vector onto the embedded feature vector. Since the Transformer's self-attention mechanism itself does not contain temporal order information, positional encoding is used to inject absolute or relative positional information of time steps, enabling the differentiation of early, middle, and late crystallization stages. Multi-head attention employs multiple parallel attention heads (e.g., 8 heads), each independently executing Scaled Dot-Product Attention. It maps the input feature embedding sequence into query, key, and value vectors, respectively, and calculates the attention weight matrix. This allows for simultaneous attention to multiple temporal patterns (e.g., short-term exciton relaxation and long-term solvent diffusion) from different subspaces, modeling both long-term and short-term dependencies. The attention weight matrix adaptively assigns higher aggregation weights to transient nucleation signals in the early stages of crystallization based on the correlation distribution of photoluminescence and diffuse reflection signals along short-range exciton relaxation and long-range solvent diffusion. This proactively focuses on key crystallization events that determine the final device performance. In the Feed-Forward Network (FFN), each attention head output is followed by a fully connected feed-forward network consisting of two linear transformation layers and a nonlinear activation function (e.g., ReLU), performing an independent nonlinear transformation on the attention output at each time step. The temporal feature aggregation module receives context-enhanced feature sequences and executes them along the time dimension. Max pooling extracts the maximum value of each feature channel within the entire window, capturing peak responses during the crystallization process (such as instantaneous signal changes during nucleation bursts). Average pooling calculates the arithmetic mean of each feature channel within the entire window, reflecting overall trend information (such as the baseline of overall luminescence and reflection levels). In the probabilistic prediction head module, a multilayer fully connected dimensionality reduction network (MLP) receives fixed-dimensional feature vectors and sequentially passes them through a multilayer fully connected network (including batch normalization and ReLU activation) to extract highly abstract performance-related features layer by layer. The last layer of the mean branch outputs the mean parameter of the conditional probability distribution of device performance indicators (such as photoelectric conversion efficiency) through a linear mapping without an activation function. The variance branch, after a linear mapping, is followed by a Softplus activation function to serve as the variance parameter of the conditional probability distribution, explicitly quantifying prediction uncertainty.

[0076] In one alternative approach, the multi-head self-attention encoder dynamically solves for the adaptive aggregation weights of the crystallized full-cycle multimodal signal using a spatiotemporal coupling correlation entropy formula, wherein the spatiotemporal coupling correlation entropy formula is:

[0077]

[0078] in, For the first The first attention head Adaptive feature aggregation weights at sampling time; , where is the temporal entropy of the photoluminescence signal, characterizing the temporal disorder of crystallization exciton relaxation; for Time-normalized photoluminescence intensity; , where is the spatial entropy of the diffuse reflection signal, characterizing the spatial inhomogeneity of the thin film surface morphology evolution; for Time-normalized diffuse reflectance intensity; For the first Attention head Gaussian kernel bandwidth hyperparameter; The temporal spatial modal coupling correlation coefficient is used to quantify the degree of correlation between dual-mode signals during the nucleation, grain growth, and homogenization stages. This is the adaptive gain coefficient for the crystallization stage; This represents the total number of sampling moments within the current sliding time window.

[0079] In this embodiment, the temporal entropy of the photoluminescence signal (characterizing exciton relaxation disorder) and the spatial entropy of the diffuse reflection signal (characterizing morphological evolution inhomogeneity) are directly embedded into the attention weight calculation through the spatiotemporal coupling correlation entropy formula. This gives the weight allocation a clear crystallization physics connotation. High-entropy regions correspond to critical moments of drastic changes in crystallization dynamics, naturally obtaining higher attention focus, making the model decision process interpretable and traceable. Different attention heads use different Gaussian kernel bandwidth hyperparameters, making different heads have different sensitivities to time scales. Small bandwidth heads are more sensitive to short-term transient events (such as nucleation spikes), while large bandwidth heads are more sensitive to long-range trend events (such as slow solvent evaporation). Multi-head parallel computing can simultaneously capture multi-scale crystallization dynamic events from millisecond-level exciton relaxation to second-level solvent diffusion without the need for trade-offs between different scales. The adaptive gain coefficient of the crystallization stage can dynamically adjust the amplitude of the attention weight according to the current crystallization stage (early nucleation, mid-stage growth, late-stage homogenization). For example, in the early nucleation stage, the gain coefficient can improve the overall attention level, forcing the model to be more sensitive to capture weak nucleation signals; in the later homogenization stage, the gain coefficient is reduced to prevent the model from overreacting to normal fluctuations, and significantly enhances the model's adaptability throughout the entire crystallization cycle.

[0080] Step S103: The deviation between the predicted performance index and the preset target performance index threshold is calculated to obtain the performance deviation amount. An inkjet printing control compensation signal is generated based on the predicted uncertainty quantization value. If the predicted uncertainty quantization value is lower than the preset uncertainty tolerance upper limit, the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value of each inkjet printhead are calculated based on the dynamic confidence weight determined by the performance deviation amount and the predicted uncertainty quantization value, and an inkjet printing control compensation signal is generated. If the predicted uncertainty quantization value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantization value converges to below the uncertainty tolerance upper limit.

[0081] In this embodiment, the quantified value of prediction uncertainty is used as the "entry threshold" for control execution. Control compensation signals are only output when the model is sufficiently confident in its predictions (uncertainty is below the tolerance limit); otherwise, more data is actively awaited. This fundamentally avoids risky adjustments in a fuzzy model state, significantly improving the safety of the control system. Under the premise of reliable prediction, the strength of the control compensation signal is not fixed but modulated by a dynamic reliability weight jointly determined by the performance deviation and the quantified value of prediction uncertainty. When prediction uncertainty is low, the reliability weight approaches 1, and control compensation is output close to full, achieving rapid correction. When prediction uncertainty is not excessive but at a high level, the reliability weight automatically decreases, and control compensation is moderately attenuated to prevent over-adjustment and avoid oscillations or insufficient response caused by "all-or-nothing" hard threshold control. When prediction uncertainty exceeds the limit, the sliding time window is adaptively extended and subsequent time-series signals are reacquired, actively increasing the observation context. Extending the window allows for the acquisition of more crystallization evolution data, thereby reducing characteristic entropy and manifold dispersion, leading to convergence of the prediction distribution. Compensation signals are generated only after the uncertainty reaches the target, and are decomposed into piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value, which correspond to droplet volume control and drop position control, respectively, so that the control decision based on uncertainty perception can be connected to the actual execution level of the inkjet printing device.

[0082] In one alternative approach, the formula for calculating the quantified value of the prediction uncertainty is:

[0083]

[0084] in, , which is the spatiotemporal joint feature entropy; for Time of the first Normalized activation percentage of crystalline-like features; The total number of predefined crystallization feature modes for the Transformer model; , where Gaussian dispersion of the Transformer characteristic manifold; Adaptive kernel width for manifold; The Gaussian conditional probability distribution of device performance indicators; The Frobenius norm of the temporal gradient matrix is ​​used to quantize the instantaneous drift intensity of the performance probability distribution. This represents the number of valid temporal feature sampling points within the current adaptive sliding time window. , These are the Transformer deep crystallization feature vectors corresponding to the i-th and j-th different temporal sampling times within the window, respectively; This is the Bayesian iterative regularization factor; This is the convergence threshold bias constant; The convergence gain coefficient for model iteration; This represents the number of convergence steps for continuous time-series sampling. It is a crystallization kinetics constraint factor; This refers to the real-time crystallization relaxation time.

[0085] In this embodiment, By adjusting the activation percentage of the predefined C-type crystalline eigenmodes in the Transformer model Entropy is calculated to quantify the uniformity of feature distribution across categories at each time step. When the feature distribution is highly concentrated in a certain category (e.g., clearly in the nucleation stage), the entropy value is low, indicating that the model has a very clear understanding of the current crystallization state. When the feature distribution is dispersed across multiple categories (e.g., in the ambiguous transition zone between nucleation and growth), the entropy value is high, indicating that the model has cognitive ambiguity regarding the current state, thus enabling the uncertainty measure to have semantic interpretability for the crystallization stage. Calculating the mean Gaussian kernel similarity among all temporal sampling point pairs on the deep feature manifold of the Transformer essentially measures the degree of feature clustering within the window in the manifold space. When feature vectors at different times within the window are close to each other (i.e., the crystallization process is in a stable evolution stage, and features at different times are highly similar), the dispersion is low and the uncertainty is small; when feature vectors are dispersed (e.g., including drastic nucleation transients and subsequent stable growth, with features spanning a large area on the manifold), the dispersion is high and the uncertainty is large, supplementing the information on the degree of dispersion between features that feature entropy cannot reflect. This is achieved by calculating the conditional probability distribution parameters. The gradient Frobenius norm along the time direction quantifies the degree of fluctuation in the performance prediction result with continuous sampling time. When the prediction mean and variance of adjacent time times change drastically (such as the prediction value fluctuating due to signal noise), the gradient norm increases, reflecting that the current prediction result is highly sensitive to time-series noise and is regarded as a high uncertainty state. It can identify the dynamic instability mode of rapid fluctuation in prediction value.

[0086] Step S104: The inkjet printing control compensation signal is sent to the embedded motion controller of the inkjet printing device in real time. The amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead are adjusted according to the piezoelectric drive voltage compensation value. The spatial synchronization interpolation parameters of the ejection trigger phase of each inkjet printhead and the substrate motion trajectory are dynamically adjusted according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold.

[0087] In this embodiment, through the joint adjustment of the embedded motion controller, the piezoelectric drive waveform adjustment and the ejection timing phase adjustment are executed collaboratively within the same control cycle, achieving synchronous coordination of volume control and position control. This avoids unintentional changes in droplet volume due to timing phase adjustment or ejection timing offsets due to voltage adjustment. The rise edge slope is used as a control variable. By adjusting the pulse width modulation duty cycle and the drive resistor value, the transition time of the piezoelectric actuator from the baseline voltage to the peak voltage is independently controlled, affecting the deformation rate of the piezoelectric ceramic and thus changing the droplet ejection speed and breakage length, significantly improving the accuracy and range of droplet volume control. The spatial synchronous interpolation offset corrects the position coordinates of the interpolation point on the substrate motion trajectory, ensuring precise synchronization between the ejection trigger moment and the actual position reached by the substrate. This compensates for systematic deviations in the landing position caused by substrate speed fluctuations, changes in droplet flight time, etc., making it particularly suitable for high-speed scanning printing scenarios and significantly improving patterning accuracy.

[0088] In one alternative approach, adjusting the amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead according to the piezoelectric drive voltage compensation value further includes:

[0089] The piezoelectric drive voltage compensation value is decoupled in the digital domain into an amplitude compensation component and a rising edge slope compensation component. The amplitude compensation component is used to set the correction amount of the peak voltage of the main pulse of the drive waveform, and the rising edge slope compensation component is used to set the correction amount of the transition time required for the drive waveform to climb from the baseline voltage to the peak voltage of the main pulse.

[0090] The embedded motion controller transmits the amplitude compensation component to the built-in digital-to-analog converter to generate an analog amplitude adjustment voltage. After the analog amplitude adjustment voltage is amplified by the drive stage consisting of a high-voltage operational amplifier and a push-pull output circuit, the peak voltage of the main pulse of the piezoelectric drive waveform is changed to the target amplitude in real time.

[0091] The embedded motion controller dynamically adjusts the pulse width modulation duty cycle of the rising edge segment of the drive waveform and the resistance value of the drive resistor according to the rising edge slope compensation component to obtain the adjusted rising edge slope, so as to control the length of the transition time.

[0092] The target amplitude and the adjusted rising edge slope parameter are written into the waveform register corresponding to each inkjet printhead. At the beginning of the next ejection cycle, the waveform generator is triggered, and the piezoelectric actuator is driven to execute ink droplet ejection according to the updated waveform parameters.

[0093] In this embodiment, the piezoelectric driving voltage compensation value is decoupled into amplitude compensation component and rise-edge slope compensation component in the digital domain. This allows for simultaneous control of droplet ejection speed (dominated by amplitude) and droplet breakage behavior (dominated by rise-edge slope), significantly expanding the adjustment range and precision of droplet volume control. This is particularly suitable for perovskite precursor inks, which are non-Newtonian fluids sensitive to shear rates. The decoupling of the compensation value is completed in the digital domain, avoiding errors such as temperature drift distortion introduced by analog domain signal processing. This ensures high precision (up to millivolt and microsecond levels) and batch-to-batch repeatability of the compensation amount, meeting the stringent requirements for process consistency in perovskite thin film manufacturing. The analog amplitude adjustment voltage (typically 0~10V) output by the digital-to-analog converter cannot directly drive the piezoelectric ceramic (requiring tens to hundreds of volts). Power amplification is achieved through a drive stage composed of a high-voltage operational amplifier and a push-pull output circuit, realizing high-fidelity conversion from low-voltage control signal to high-voltage drive signal. The push-pull output circuit has low output impedance and high current driving capability, ensuring that the piezoelectric ceramic still obtains sufficient transient power at high-speed ejection frequencies, avoiding waveform distortion. The target amplitude and adjusted rise-edge slope parameters are written into the waveform register corresponding to each inkjet printhead. At the start of the next ejection cycle, the waveform generator is triggered to execute according to the new parameters, allowing each ejection cycle to use independent and different waveform parameters. This enables real-time adjustment per ejection, with response times reaching milliseconds or even sub-milliseconds, far faster than the time scale (seconds) of the crystallization process, providing ample time margin for closed-loop control. Each inkjet printhead has an independent waveform register set, which can configure the optimal amplitude and rise-edge slope parameters for the individual characteristics of different printheads (such as piezoelectric ceramic tolerance, nozzle orifice diameter differences, and ink channel flow resistance differences). This solves the common "poor consistency" problem in multi-head arrays, ensuring a high degree of consistency in droplet volume and ejection velocity across all printheads in the array.

[0094] In one alternative approach, dynamically adjusting the spatial synchronization interpolation parameters between the ejection trigger phase of each inkjet printhead and the substrate motion trajectory based on the inkjet timing phase delay compensation value further includes:

[0095] The inkjet timing phase delay compensation value is decomposed into an ejection trigger phase offset and a spatial synchronization interpolation offset. The ejection trigger phase offset is used to correct the amount by which the ejection time of a single inkjet printhead is advanced or delayed relative to the system synchronization clock. The spatial synchronization interpolation offset is used to correct the amount by which the position of the interpolation point of the substrate motion trajectory corresponding to the ejection time is shifted forward or backward along the motion direction.

[0096] The embedded motion controller sends the jet trigger phase offset to the field programmable gate array timing generator to generate an independently phase-adjustable jet trigger pulse for each inkjet printhead. The phase delay relative to the system synchronization clock is locked as the jet trigger phase offset. The jet trigger pulse is routed to the piezoelectric drive circuit of the corresponding inkjet printhead as a jet enable signal.

[0097] When the embedded motion controller calculates the substrate motion trajectory in real time based on the spatial synchronous interpolation offset, it dynamically corrects the spatial step size and interpolation point position coordinates in each interpolation cycle.

[0098] The ejection trigger pulse and the spatial synchronization interpolation parameter are anchored on the time axis to ensure that each ejection trigger pulse instantaneously triggers ink droplet ejection at the corresponding corrected interpolation point position coordinates. This establishes a spatiotemporal coordinated control relationship between the ejection trigger phase offset and the spatial synchronization interpolation offset, thereby achieving dynamic synchronous matching between the ejection time and the substrate movement position to optimize the accuracy of the ink droplet landing point position.

[0099] In this embodiment, a field-programmable gate array (FPGA) timing generator is used to independently generate phase-adjustable ejection trigger pulses for each inkjet printhead, with the phase delay precisely locked to the nanosecond level. The hardware parallelism of the FPGA ensures that the trigger pulses of each printhead do not interfere with each other and are output synchronously, realizing differentiated independent timing control of the multi-printhead array and compensating for inherent phase deviations caused by differences in printhead installation position, ink path length, and piezoelectric response. The spatial synchronous interpolation offset is directly involved in the real-time calculation of the substrate motion trajectory, dynamically correcting the spatial step size and interpolation point position coordinates within each interpolation cycle. The compensation amount for the landing point deviation is pre-positioned in the trajectory planning stage to achieve real-time closed-loop correction of the landing point position, which is particularly suitable for dynamic accuracy control in high-speed scanning printing scenarios. The ejection trigger pulse and the spatial synchronous interpolation parameters are precisely anchored on the time axis to ensure that each ejection trigger pulse triggers ink droplet ejection instantaneously at the corresponding corrected interpolation point position coordinates, so that the synchronization accuracy between the ejection time and the actual arrival position of the substrate can reach the sub-microsecond and sub-micrometer level, fundamentally guaranteeing the landing point position accuracy. The jet trigger phase offset changes the timing of the jetting, and the spatial synchronization interpolation offset changes the location of the jetting. The two work together under the spatiotemporal anchoring mechanism to achieve dynamic synchronization between the jetting timing and the substrate movement position. This can simultaneously compensate for the combined landing point deviation caused by various dynamic factors such as substrate speed fluctuations, droplet flight time changes, and nozzle height changes.

[0100] In an alternative approach, decomposing the inkjet timing phase delay compensation value into an ejection trigger phase offset and a spatial synchronization interpolation offset further includes:

[0101] After receiving the inkjet timing phase delay compensation value, the embedded motion controller inputs the inkjet timing phase delay compensation value to the built-in coupling state discriminator. Based on the current substrate movement speed, the spatial layout spacing of the inkjet printhead array, and the real-time ink droplet flight time, it calculates the proportion of the temporal component caused by the jet triggering moment drift and the proportion of the spatial component caused by the landing point position system deviation in the inkjet timing phase delay compensation value.

[0102] The embedded motion controller reads the dominant partition label of the current crystallization stage in real time from the Transformer in-situ monitoring model. The dominant partition label includes the solvent evaporation dominant region, the mesophase transformation dominant region, and the halogen homogenization dominant region. Based on the dominant partition label, the adaptive decoupling factor is dynamically calculated and adjusted in the time domain bias direction to increase the allocation weight of the jet trigger phase bias.

[0103] The jet trigger phase offset is obtained based on the inkjet timing phase delay compensation value and the adaptive decoupling factor; the proportion of the time domain component and the proportion of the spatial domain component are input to the interpolation resolution quantizer, and the spatial synchronous interpolation offset expressed as an integer multiple of the interpolation resolution is output with the substrate motion trajectory interpolation resolution as the minimum quantization step size.

[0104] In this embodiment, the adaptive decoupling factor is dynamically calculated based on the dominant partition label of the crystallization stage output in real time by the Transformer in-situ monitoring model. During the solvent evaporation-dominant period, the drastic changes in ink properties lead to large fluctuations in droplet flight time. The decoupling factor automatically adjusts towards the time-domain bias direction, giving greater weight to the jet trigger phase bias. During the halogen homogenization-dominant period, the film morphology tends to stabilize, but the landing point accuracy has a significant impact on the final uniformity. The decoupling factor automatically reverts towards the spatial domain, ensuring that the compensation strategy precisely matches the crystallization physical state, achieving adaptive optimal adjustment throughout the entire cycle. After inputting the proportions of the time-domain and spatial components into the interpolation resolution quantizer, the minimum quantization step size is set to the interpolation resolution of the substrate motion trajectory. The spatial synchronous interpolation offset, expressed as an integer multiple of the interpolation resolution, is output, ensuring that the spatial offset falls entirely within the physically feasible range of the motion control system. This avoids the accumulation of interpolator rounding errors due to non-integer multiple offsets, while ensuring that the spatial offset command is fully compatible with the interpolator's native data format, requiring no additional scaling transformation. The contribution ratios of the time and spatial domains are quantified by a coupled state discriminator, and then adjusted towards the time domain by an adaptive decoupling factor. The jet trigger phase offset (time domain) and spatial synchronization interpolation offset (spatial domain) are calculated independently. The two outputs are executed independently by the FPGA timing generator (time domain) and trajectory interpolator (spatial domain), respectively, without interference between them. This avoids incorrectly correcting time domain deviations through spatial domain correction (and vice versa), significantly improving the convergence speed and final accuracy of the compensation.

[0105] According to the solution provided by the present invention, during the vacuum quenching and crystallization stage of perovskite thin film inkjet printing, the time-series evolution signal of the entire crystallization cycle is simultaneously acquired by a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system; the time-series evolution signal is input into the Transformer in-situ monitoring model to capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage, and to extract the multi-stage non-stationary deep features caused by solvent evaporation kinetics, intermediate phase transformation path and halogen homogenization process during perovskite crystallization; the multi-stage non-stationary deep features are mapped to conditional probability distribution parameters of device performance indicators, and the expected value is extracted from the conditional probability distribution parameters as the predicted performance indicator, and the joint quantization value corresponding to the predicted performance indicator is calculated as the prediction uncertainty quantization value; the deviation between the predicted performance indicator and the preset target performance indicator threshold is calculated to obtain the performance deviation amount, and an inkjet printing control compensation signal is generated according to the prediction uncertainty quantization value; wherein, if the prediction uncertainty quantization value is lower than the preset uncertainty tolerance upper limit, according to The dynamic reliability weight determined by the performance deviation and the predicted uncertainty quantification value is used to calculate the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value for each inkjet printhead, generating an inkjet printing control compensation signal. If the predicted uncertainty quantification value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantification value converges to below the uncertainty tolerance upper limit. The inkjet printing control compensation signal is sent to the embedded motion controller of the inkjet printing device in real time. The amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead are adjusted according to the piezoelectric drive voltage compensation value, and the spatial synchronization interpolation parameters of the ejection trigger phase and substrate motion trajectory of each inkjet printhead are dynamically adjusted according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold. This invention provides adaptive control for the crystallization process driven by inkjet printing of perovskite thin films, which significantly improves the consistency of film quality and the stability of the production line.

[0106] Figure 4 A schematic diagram of the framework of an intelligent control device for perovskite thin film inkjet printing based on a Transformer in-situ monitoring model according to an embodiment of the present invention is shown. The intelligent control device for perovskite thin film inkjet printing based on a Transformer in-situ monitoring model includes:

[0107] The timing signal synchronous acquisition module 410 is used to synchronously acquire the timing evolution signal of the entire crystallization cycle through a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system during the vacuum quenching and crystallization stage of perovskite thin film inkjet printing.

[0108] The deep feature extraction and performance prediction module 420 is used to input the time-series evolution signal into the Transformer in-situ monitoring model, capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage, and extract the multi-stage non-stationary deep features caused by solvent volatilization kinetics, intermediate phase transformation path and halogen homogenization process during the perovskite crystallization process; map the multi-stage non-stationary deep features into conditional probability distribution parameters of device performance indicators, extract the expected value from the conditional probability distribution parameters as the predicted performance indicator, and calculate the joint quantization value corresponding to the predicted performance indicator as the prediction uncertainty quantization value;

[0109] The deviation calculation and adaptive compensation decision module 430 is used to calculate the deviation between the predicted performance index and the preset target performance index threshold to obtain the performance deviation amount, and generate an inkjet printing control compensation signal based on the predicted uncertainty quantification value. Specifically, if the predicted uncertainty quantification value is lower than a preset uncertainty tolerance upper limit, the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value for each inkjet printhead are calculated based on the dynamic confidence weight determined by the performance deviation amount and the predicted uncertainty quantification value, and an inkjet printing control compensation signal is generated. If the predicted uncertainty quantification value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantification value converges to below the uncertainty tolerance upper limit.

[0110] The piezoelectric drive and motion trajectory coordinated control module 440 is used to send the inkjet printing control compensation signal to the embedded motion controller of the inkjet printing device in real time. It adjusts the amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead according to the piezoelectric drive voltage compensation value, and dynamically adjusts the spatial synchronization interpolation parameters of the ejection trigger phase of each inkjet printhead and the substrate motion trajectory according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and the landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold.

[0111] 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.

[0112] like Figure 5As shown, the computing device may include: a processor 502, a communications interface 504, a memory 506, and a communications bus 508.

[0113] 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 perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model.

[0114] Specifically, program 510 may include program code that includes computer operation instructions.

[0115] 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.

[0116] 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.

[0117] According to the solution provided by the present invention, during the vacuum quenching and crystallization stage of perovskite thin film inkjet printing, the time-series evolution signal of the entire crystallization cycle is simultaneously acquired by a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system; the time-series evolution signal is input into the Transformer in-situ monitoring model to capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage, and to extract the multi-stage non-stationary deep features caused by solvent evaporation kinetics, intermediate phase transformation path and halogen homogenization process during perovskite crystallization; the multi-stage non-stationary deep features are mapped to conditional probability distribution parameters of device performance indicators, and the expected value is extracted from the conditional probability distribution parameters as the predicted performance indicator, and the joint quantization value corresponding to the predicted performance indicator is calculated as the prediction uncertainty quantization value; the deviation between the predicted performance indicator and the preset target performance indicator threshold is calculated to obtain the performance deviation amount, and an inkjet printing control compensation signal is generated according to the prediction uncertainty quantization value; wherein, if the prediction uncertainty quantization value is lower than the preset uncertainty tolerance upper limit, according to The dynamic reliability weight determined by the performance deviation and the predicted uncertainty quantification value is used to calculate the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value for each inkjet printhead, generating an inkjet printing control compensation signal. If the predicted uncertainty quantification value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantification value converges to below the uncertainty tolerance upper limit. The inkjet printing control compensation signal is sent to the embedded motion controller of the inkjet printing device in real time. The amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead are adjusted according to the piezoelectric drive voltage compensation value, and the spatial synchronization interpolation parameters of the ejection trigger phase and substrate motion trajectory of each inkjet printhead are dynamically adjusted according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold. This invention provides adaptive control for the crystallization process driven by inkjet printing of perovskite thin films, which significantly improves the consistency of film quality and the stability of the production line.

[0118] 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 control method for inkjet printing of perovskite thin films based on a Transformer in-situ monitoring model, characterized in that, include: During the vacuum quenching and crystallization stage of perovskite thin film inkjet printing, the time-series evolution signal of the entire crystallization cycle was simultaneously acquired by a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system. The time-series evolution signal is input into the Transformer in-situ monitoring model to capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage. Multi-stage non-stationary deep features caused by solvent volatilization kinetics, intermediate phase transformation path and halogen homogenization process during perovskite crystallization are extracted. The multi-stage non-stationary deep features are mapped to the conditional probability distribution parameters of device performance indicators. The expected value is extracted from the conditional probability distribution parameters as the predicted performance indicator. The joint quantization value corresponding to the predicted performance indicator is calculated as the prediction uncertainty quantization value. The predicted performance index is deviated from the preset target performance index threshold to obtain the performance deviation amount. An inkjet printing control compensation signal is generated based on the predicted uncertainty quantification value. If the predicted uncertainty quantification value is lower than the preset uncertainty tolerance upper limit, the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value of each inkjet printhead are calculated based on the dynamic confidence weight determined by the performance deviation amount and the predicted uncertainty quantification value, and an inkjet printing control compensation signal is generated. If the predicted uncertainty quantification value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantification value converges to below the uncertainty tolerance upper limit. The inkjet printing control compensation signal is sent to the embedded motion controller of the inkjet printing equipment in real time. The amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead are adjusted according to the piezoelectric drive voltage compensation value. The spatial synchronization interpolation parameters of the ejection trigger phase of each inkjet printhead and the substrate motion trajectory are dynamically adjusted according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold.

2. The perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model according to claim 1, characterized in that, The time-series evolution signal of the entire crystallization cycle was acquired simultaneously using a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system, and further included: At the transient zero moment of the start of the vacuum quenching and crystallization stage, the embedded industrial control computer simultaneously issues an external trigger synchronous acquisition command to the time-resolved photoluminescence spectrometer and the diffuse reflectance imaging system. The external trigger synchronous acquisition command contains a preset sampling frequency and a sampling time window length. The sampling time window length is pre-calibrated based on the saturated vapor pressure of the perovskite precursor solvent and the pumping speed of the vacuum chamber to ensure coverage of the entire crystallization cycle from supersaturated nucleation to solid film formation. The time-resolved photoluminescence spectrometer, in response to the external trigger synchronous acquisition command, applies pulsed laser excitation to the surface of the perovskite thin film at a preset sampling frequency. A time-correlated single-photon counting process is performed by the single-photon counting module, outputting a photoluminescence intensity time-series sequence continuously recorded at nanosecond-level time resolution. Simultaneously, the diffuse reflection imaging system, in response to the same external trigger synchronous acquisition command, activates a CMOS camera to continuously capture frames of the thin film surface under broadband illumination, acquiring a diffuse reflection image sequence. An image preprocessing module calculates the grayscale mean for each pixel region of each frame image, outputting a diffuse reflection intensity time-series sequence aligned with the timestamp of the photoluminescence intensity time-series sequence. The photoluminescence intensity time series and the diffuse reflection intensity time series are synchronized by bilinear interpolation and spatial coordinate matching using a unified timestamp index, forming a multimodal synchronization time series signal matrix as the time series evolution signal for the entire crystallization cycle; wherein, each row of the multimodal synchronization time series signal matrix corresponds to a discrete sampling time, and each column carries the photoluminescence intensity value and diffuse reflection intensity value of the discrete sampling time, respectively.

3. The perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model according to claim 2, characterized in that, The Transformer in-situ monitoring model includes: The input embedding layer is used to perform local temporal feature extraction and dimensional projection on the multimodal signal vector at each sampling time in the multimodal synchronous temporal signal matrix through a one-dimensional convolutional layer to generate an embedded feature vector; a learnable positional encoding vector is superimposed on the embedded feature vector to inject temporal order information, and the output is a feature embedding sequence with positional encoding that preserves global temporal dependencies. A multi-head self-attention encoder is used to calculate the attention weight matrix between the query vector at each time step and the key vector at all time steps of the feature embedding sequence; wherein, the attention weight matrix adaptively assigns a higher aggregation weight to the early transient nucleation signal of crystallization than that of the mid-term grain growth transformation signal based on the correlation distribution of the photoluminescence intensity time sequence and the diffuse reflection intensity time sequence on short-range exciton relaxation and long-range solvent diffusion, so as to amplify the key crystallization events and obtain a context-enhanced feature sequence; The temporal feature aggregation module is used to receive the context-enhanced feature sequence, perform max pooling and average pooling operations in parallel along the time dimension, and concatenate the dual pooling results along the feature channel dimension to form a fixed-dimensional feature vector. The fixed-dimensional feature vector encodes the multi-stage non-stationary deep features of the entire perovskite crystallization process, including early solvent volatilization feature components, mid-term feature components dominated by the intermediate phase transformation path, and late-term feature components dominated by the halogen homogenization process. The probability prediction head module is used to receive the fixed-dimensional feature vector, and after dimensionality reduction and activation layer by layer through a multi-layer fully connected network, output two parallel branches. One branch generates the mean parameter of the device performance index through linear mapping, and the other branch generates a positive variance parameter through Softplus activation. The conditional probability distribution parameters of the device performance index are constructed based on the mean parameter and the positive variance parameter.

4. The perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model according to claim 1, characterized in that, Adjusting the amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead according to the piezoelectric drive voltage compensation value further includes: The piezoelectric drive voltage compensation value is decoupled in the digital domain into an amplitude compensation component and a rising edge slope compensation component. The amplitude compensation component is used to set the correction amount of the peak voltage of the main pulse of the drive waveform, and the rising edge slope compensation component is used to set the correction amount of the transition time required for the drive waveform to climb from the baseline voltage to the peak voltage of the main pulse. The embedded motion controller transmits the amplitude compensation component to the built-in digital-to-analog converter to generate an analog amplitude adjustment voltage. After the analog amplitude adjustment voltage is amplified by the drive stage consisting of a high-voltage operational amplifier and a push-pull output circuit, the peak voltage of the main pulse of the piezoelectric drive waveform is changed to the target amplitude in real time. The embedded motion controller dynamically adjusts the pulse width modulation duty cycle of the rising edge segment of the drive waveform and the resistance value of the drive resistor according to the rising edge slope compensation component to obtain the adjusted rising edge slope, so as to control the length of the transition time. The target amplitude and the adjusted rising edge slope parameter are written into the waveform register corresponding to each inkjet printhead. At the beginning of the next ejection cycle, the waveform generator is triggered, and the piezoelectric actuator is driven to execute ink droplet ejection according to the updated waveform parameters.

5. The perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model according to claim 1, characterized in that, The dynamic adjustment of the spatial synchronization interpolation parameters between the ejection trigger phase of each inkjet printhead and the substrate motion trajectory based on the inkjet timing phase delay compensation value further includes: The inkjet timing phase delay compensation value is decomposed into an ejection trigger phase offset and a spatial synchronization interpolation offset. The ejection trigger phase offset is used to correct the amount by which the ejection time of a single inkjet printhead is advanced or delayed relative to the system synchronization clock. The spatial synchronization interpolation offset is used to correct the amount by which the position of the interpolation point of the substrate motion trajectory corresponding to the ejection time is shifted forward or backward along the motion direction. The embedded motion controller sends the jet trigger phase offset to the field programmable gate array timing generator to generate an independently phase-adjustable jet trigger pulse for each inkjet printhead. The phase delay relative to the system synchronization clock is locked as the jet trigger phase offset. The jet trigger pulse is routed to the piezoelectric drive circuit of the corresponding inkjet printhead as a jet enable signal. When the embedded motion controller calculates the substrate motion trajectory in real time based on the spatial synchronous interpolation offset, it dynamically corrects the spatial step size and interpolation point position coordinates in each interpolation cycle. The ejection trigger pulse and the spatial synchronization interpolation parameter are anchored on the time axis to ensure that each ejection trigger pulse instantaneously triggers ink droplet ejection at the corresponding corrected interpolation point position coordinates. This establishes a spatiotemporal coordinated control relationship between the ejection trigger phase offset and the spatial synchronization interpolation offset, thereby achieving dynamic synchronous matching between the ejection time and the substrate movement position to optimize the accuracy of the ink droplet landing point position.

6. The perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model according to claim 5, characterized in that, The inkjet timing phase delay compensation value is further decomposed into jet trigger phase offset and spatial synchronization interpolation offset, including: After receiving the inkjet timing phase delay compensation value, the embedded motion controller inputs the inkjet timing phase delay compensation value to the built-in coupling state discriminator. Based on the current substrate movement speed, the spatial layout spacing of the inkjet printhead array, and the real-time ink droplet flight time, it calculates the proportion of the temporal component caused by the jet triggering moment drift and the proportion of the spatial component caused by the landing point position system deviation in the inkjet timing phase delay compensation value. The embedded motion controller reads the dominant partition label of the current crystallization stage in real time from the Transformer in-situ monitoring model. The dominant partition label includes the solvent evaporation dominant region, the mesophase transformation dominant region, and the halogen homogenization dominant region. Based on the dominant partition label, the adaptive decoupling factor is dynamically calculated and adjusted in the time domain bias direction to increase the allocation weight of the jet trigger phase bias. The jet trigger phase offset is obtained based on the inkjet timing phase delay compensation value and the adaptive decoupling factor; the proportion of the time domain component and the proportion of the spatial domain component are input to the interpolation resolution quantizer, and the spatial synchronous interpolation offset expressed as an integer multiple of the interpolation resolution is output with the substrate motion trajectory interpolation resolution as the minimum quantization step size.

7. The perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model according to claim 3, characterized in that, The multi-head self-attention encoder dynamically solves for the adaptive aggregation weights of the crystallization full-cycle multimodal signal using a spatiotemporal coupling correlation entropy formula, wherein the spatiotemporal coupling correlation entropy formula is: in, For the first The first attention head Adaptive feature aggregation weights at sampling time; , where is the temporal entropy of the photoluminescence signal, characterizing the temporal disorder of crystallization exciton relaxation; for Time-normalized photoluminescence intensity; , where is the spatial entropy of the diffuse reflection signal, characterizing the spatial inhomogeneity of the thin film surface morphology evolution; for Time-normalized diffuse reflectance intensity; For the first Attention head Gaussian kernel bandwidth hyperparameter; The temporal spatial modal coupling correlation coefficient is used to quantify the degree of correlation between dual-mode signals during the nucleation, grain growth, and homogenization stages. This is the adaptive gain coefficient for the crystallization stage; This represents the total number of sampling moments within the current sliding time window.

8. The perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model according to claim 1, characterized in that, The formula for calculating the quantified value of the prediction uncertainty is as follows: in, , which is the spatiotemporal joint feature entropy; for Time of the first Normalized activation percentage of crystalline-like features; The total number of predefined crystallization feature modes for the Transformer model; , where Gaussian dispersion of the Transformer characteristic manifold; Adaptive kernel width for manifold; The Gaussian conditional probability distribution of device performance indicators; The Frobenius norm of the temporal gradient matrix is ​​used to quantize the instantaneous drift intensity of the performance probability distribution. This represents the number of valid temporal feature sampling points within the current adaptive sliding time window. , These are the Transformer deep crystallization feature vectors corresponding to the i-th and j-th different temporal sampling times within the window, respectively; This is the Bayesian iterative regularization factor; This is the convergence threshold bias constant; The convergence gain coefficient for model iteration; This represents the number of convergence steps for continuous time-series sampling. It is a crystallization kinetics constraint factor; This refers to the real-time crystallization relaxation time.

9. A smart control device for perovskite thin film inkjet printing based on a Transformer in-situ monitoring model, characterized in that, include: The timing signal synchronous acquisition module is used to synchronously acquire the timing evolution signal of the entire crystallization cycle through a time-resolved photoluminescence spectrometer and a diffuse reflectance imaging system during the vacuum quenching and crystallization stage of perovskite thin film inkjet printing. The deep feature extraction and performance prediction module is used to input the time-series evolution signal into the Transformer in-situ monitoring model, capture the transient nucleation signal in the early stage of crystallization and the grain growth transformation signal in the middle stage, and extract the multi-stage non-stationary deep features caused by solvent volatilization kinetics, intermediate phase transformation path and halogen homogenization process during the perovskite crystallization process; map the multi-stage non-stationary deep features into conditional probability distribution parameters of device performance indicators, extract the expected value from the conditional probability distribution parameters as the predicted performance indicator, and calculate the joint quantization value corresponding to the predicted performance indicator as the prediction uncertainty quantization value; The deviation calculation and adaptive compensation decision module is used to calculate the deviation between the predicted performance index and the preset target performance index threshold to obtain the performance deviation amount, and generate an inkjet printing control compensation signal based on the predicted uncertainty quantification value. Specifically, if the predicted uncertainty quantification value is lower than a preset uncertainty tolerance upper limit, the piezoelectric drive voltage compensation value and inkjet timing phase delay compensation value for each inkjet printhead are calculated based on the dynamic confidence weight determined by the performance deviation amount and the predicted uncertainty quantification value, and an inkjet printing control compensation signal is generated. If the predicted uncertainty quantification value is not lower than the uncertainty tolerance upper limit, the length of the sliding time window is adaptively extended and the timing evolution signal of subsequent time steps is re-acquired until the predicted uncertainty quantification value converges to below the uncertainty tolerance upper limit. The piezoelectric drive and motion trajectory coordinated control module is used to send the inkjet printing control compensation signal to the embedded motion controller of the inkjet printing device in real time. It adjusts the amplitude and rise slope of the piezoelectric actuator drive waveform of each inkjet printhead according to the piezoelectric drive voltage compensation value, and dynamically adjusts the spatial synchronization interpolation parameters of the ejection trigger phase of each inkjet printhead and the substrate motion trajectory according to the inkjet timing phase delay compensation value, so as to coordinately optimize and control the droplet volume variation coefficient and landing position. The embedded motion controller performs joint adjustment to make the predicted performance index gradually converge to the target performance index threshold.

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 perovskite thin film inkjet printing control method based on the Transformer in-situ monitoring model.

Images