A computational optimization method for exascale OLED macro-inkjet printing

Abstract

The present invention relates to a million-level OLED macro-inkjet printing optimization calculation method, including: S1, initialization; S2, constructing a Markov decision process model for nozzle optimization, a Markov decision process model for million-level ink droplet jetting optimization, and a Markov decision process model for nozzle step scheduling; S3, solving the Markov decision process model for nozzle optimization using a P-TLQ algorithm to obtain a nozzle deflection angle, an initial position of the nozzle in the X direction, and a nozzle movement speed; S4, solving the Markov decision process model for million-level ink droplet jetting optimization using the P-TLQ algorithm to obtain a nozzle combination selection for a pixel slot; S5, if printing of all pixel slots is not completed, obtaining a nozzle X-direction step scheduling, returning to step S3, if printing of all pixel slots is completed, otherwise printing ends. The present invention realizes multi-objective optimization of scheduling tasks and accelerates the algorithm convergence speed.

Description

Technical Field

[0001] The present invention relates to the field of inkjet printing optimization technology, and in particular to a method for optimizing the calculation of exascale OLED macro-inkjet printing. Background Art

[0002] The application of inkjet printing technology in the manufacture of organic light-emitting diode (OLED) displays involves depositing organic light-emitting materials directly onto a substrate through inkjet printing technology to form the required pixels and patterns, thereby achieving efficient production of displays. The core challenge of this technology lies in optimizing the scheduling of tens of thousands of nozzles to spray ink droplets into the trillions of pixels on a moving substrate. During the production process, the inkjet printing scheduling task needs to complete the process of a group of nozzles moving across the display panel, controlling tens of thousands of nozzles with different performances to spray ink droplets at each spraying opportunity. Through the spraying of multiple nozzles, the trillions of pixel slots are filled with a certain volume, and finally the entire substrate can be printed in the shortest time. In addition, the difference in ink droplet volume between the trillions of pixel slots, the uniform distribution of ink droplet landing points within the pixel slots, and the balanced use of nozzles can also be optimized.

[0003] Printing optimization challenges arise from the numerous and often changing pixel slot shapes and distribution schemes on the panel. For example, the shapes, sizes, and arrangements of the basic red, green, and blue sub-pixel slots all vary. Furthermore, in the display panel production process, to improve substrate cutting efficiency, a partitioned printing method (Multi-Model Glass) is often used. Multiple OLED screens with different pixel slot sizes and arrangements are arranged on the same display panel for printing the display ink. Therefore, printing optimization must be able to adapt to changes in the pixel slot position and size on the display panel. Furthermore, factors such as the uncertainty of whether each nozzle will fail, the varying nozzle spray volumes, and the varying and easily changing spray trajectories create a highly dynamic environment.

[0004] Computational optimization methods for exascale OLED macro-inkjet printing aim to effectively control and optimize nozzle jetting tasks to improve production efficiency and quality. Currently, optimization algorithms for inkjet OLED printing scheduling are immature, and related research is relatively limited. However, by drawing on other scheduling problems and scheduling optimization algorithms, some preliminary solutions can be provided for the scheduling problem of inkjet OLED technology. One research direction is to model the OLED printing scheduling optimization problem as a classic scheduling optimization problem, such as job shop scheduling, box packing, and backpacking, and then solve it using established algorithms. However, this research direction is difficult to directly apply due to the highly dynamic nature of inkjet OLED scheduling and the large scale of the problem. Similarly, borrowing and introducing intelligent computing methods such as genetic algorithms, simulated annealing algorithms, and tabu search algorithms to minimize overall printing time or maximize resource utilization through task sequencing, resource allocation, and time scheduling is also difficult to directly apply due to the highly dynamic nature, large problem scale, and multi-objective requirements of inkjet OLED scheduling. The more commonly used scheduling method is a rule-based scheduling method, which establishes scheduling rules based on experience to achieve the scheduling and control of the jetting tasks in the printing process. Its scheduling effect is completely dependent on experience and is difficult to guarantee. There are also some multi-objective scheduling methods that are oriented towards specific fields and specific problems, such as multi-objective deep reinforcement learning methods for energy storage scheduling optimization. However, these methods usually model and construct scheduling algorithms for specific problems, are not universal, and are difficult to apply to the scheduling of inkjet printed OLEDs. Therefore, it is necessary to propose corresponding computational methods for inkjet printing task scheduling and optimization to address the specific problems faced by large-scale, high-dynamic, and multi-objective inkjet printing of exascale OLEDs. Summary of the Invention

[0005] The purpose of this invention is to provide a method for optimizing the macro-inkjet printing of exascale OLEDs. This method improves the DQN algorithm based on threshold dictionary sorting (TLO), combines the idea of ​​potential-based reward shaping (PBRS), designs an effective printing potential function, realizes multi-objective optimization of scheduling tasks, and reduces the algorithm oscillation time, thereby accelerating the convergence of the algorithm.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A method for optimizing the inkjet printing of exascale OLEDs in large quantities, comprising:

[0008] S1. Initialize the nozzle parameters and pixel slot parameters, and set the optimization target;

[0009] S2. Constructing Markov decision process models for printhead optimization, exascale droplet ejection optimization, and printhead step scheduling;

[0010] S3. Solve the Markov decision process model for the nozzle optimization using the P-TLQ algorithm to obtain the nozzle deflection angle, the initial X-axis position of the nozzle, the nozzle movement speed, and the set of droplets that can land in the pixel slot. The P-TLQ algorithm is a DQN algorithm that combines a global situation energy function and a threshold lexicon sorting algorithm.

[0011] S4. Based on the set of droplets that can land in the pixel slot, the Markov decision process model for optimizing the exascale ink droplet ejection is solved using the P-TLQ algorithm to obtain a nozzle combination selection for the pixel slot;

[0012] S5. If printing of all pixel slots is not completed, the Markov decision process model of the nozzle step scheduling is solved using the P-TLQ algorithm to obtain the nozzle X-direction step scheduling, and return to step S3. If printing of all pixel slots is completed, otherwise printing ends.

[0013] Optionally, the nozzle parameters include: nozzle arrangement, nozzle spraying frequency and nozzle movable speed range;

[0014] The pixel slot parameters include: substrate pixel slot size / arrangement and ink droplet landing area within the pixel slot;

[0015] Setting the optimization target includes: setting the target volume of ink droplets in a single pixel slot, the difference in total ink volume between all pixel slots, the unevenness of droplet spacing in the pixel slots, the printing time, and the imbalance of the nozzles.

[0016] Optionally, the Markov decision process model for nozzle optimization constructed in S2 includes:

[0017] Establish the state space of the nozzle optimization stage:

[0018]

[0019] in, Represents the state space of the nozzle optimization stage, Represents the volume of droplets to be ejected in the pixel slot (i, j) that needs to be ejected in the current production stage, Represents the distribution state of the ejected droplets in the pixel slot (i, j) in the current production stage, V′ i，j Represents the volume of liquid ejected by the nozzle in a single shot, h a，b Represents the number of times the nozzle has sprayed, Δt a，b Represents the time interval between the nozzle and the last injection, N pass represents the number of available nozzles in the pass, a and b represent the nozzles in row a and column b, and sub represents the set of pixel slots that need to be sprayed;

[0020] Establish the action space for the nozzle optimization stage:

[0021]

[0022] Among them, A sub Represents the action space of the nozzle optimization stage, is the nozzle combination selected for pixel slot (i, j) in the current production stage, P(n) is the set of pixel slots scanned by all nozzles, and n is the number of all nozzles;

[0023] The system reward function in the nozzle optimization stage is:

[0024]

[0025] Among them, r1, r2, and r3 are the volume difference reward, the droplet landing point uniform distribution reward, and the nozzle balanced use reward respectively. V i，j is the volume of droplets to be ejected in all pixel slots (i, j), To meet the standard volume, represents the number of droplets within the effective area of ​​the kth landing point in the pixel slot (i, j), represents the average number of droplets in the four effective areas of the pixel slot (i, j), max(l i，j ) represents the maximum value among the distances between each landing point, l ε represents the optimal placement distance, τ a，b is the number of times the nozzle numbered (a, b) in the pixel slot (i, j) is used, is the average number of times all nozzles in the pixel slot (i, j) are used, K i，j is the number of locations where ink droplets can be placed in the pixel slot (i, j), p a，b is the subset of nozzles that are actually ejected from the nozzles that can eject pixel slot (i, j), p i，j is the set of nozzles that can spray pixel slot (i, j).

[0026] Optionally, the Markov decision process model for exascale droplet ejection optimization in S2 includes:

[0027] Constructing the state space for the exascale droplet ejection optimization phase:

[0028] S 2，t ={V t ,D t ,V ts ,h t ,Δt,N t}

[0029] Among them, S 2，t Optimizing the state space for exascale droplet ejection, V trepresents the volume of ink droplets to be sprayed in the pixel slot at time t, D t Represents the distribution state of the ink droplets sprayed in the pixel slot at time t, V ts represents the volume of liquid sprayed by the nozzle at time t, h t represents the number of times the nozzle has sprayed at time t, Δt represents the interval from the nozzle to the last spraying, N t Represents the number of pixel slot nozzles that have passed through this printing round;

[0030] Establishing the action space for the exascale droplet ejection optimization phase:

[0031] A 2，t ={P c (n)∈P(n),n∈(0,a*b-1)}

[0032] Among them, A 2，t The action space for the optimization phase of exascale ink droplet ejection, P c (n) is the nozzle combination selected for the cth pixel slot, a and b represent the nozzles in the ath row and bth column on the nozzle, P(n) is the set of pixel slots swept by all nozzles, and n is the number of all nozzles;

[0033] The system reward function in the exascale ink droplet ejection optimization phase is:

[0034]

[0035]

[0036] in, is the number of nozzles within the effective area of ​​the kth landing point in the (i, j)th pixel slot, is the average number of nozzles of the four dots in the (i, j)th pixel slot;

[0037]

[0038] Among them, r′1, r′2, and r′3 are the three reward functions for exascale ink droplet ejection, is the average number of times all nozzles in the (i, j)th pixel slot are used, p a，b is the subset of nozzles that are actually ejected from the nozzles that can eject pixel slot (i, j), p i，j is the set of nozzles that can spray pixel slot (i, j).

[0039] Optionally, a Markov decision process model for nozzle step scheduling is constructed including:

[0040] Establish the state space of the multi-round scheduling phase:

[0041]

[0042] Among them, S mul Represents the state space of the multi-round scheduling phase, pass represents the pass-th printing round, Indicates the number of pixel slots filled in the pass printing round, Indicates the number of pixel slots covered by the nozzle track during the pass printing round. represents the variance of the number of available nozzles in different pixel slots during the printing pass;

[0043] Establish the action space for the multi-round scheduling phase:

[0044] A mul ={L pass ±ΔL}

[0045] Among them, A mul is the action space in the multi-round scheduling phase, ΔL is the minimum distance the nozzle moves laterally, and L pass is the lateral movement of the sprinkler head between adjacent rounds;

[0046] The reward function in the multi-round scheduling phase is:

[0047]

[0048] Among them, N s ∈[0, M·N] represents the number of pixel slots that the inkjet head can fill in the pass-th printing round, M and N are the total number of rows and columns of pixel slots;

[0049]

[0050] in, Indicates the number of nozzles available in the kth landing area of ​​pixel slot (i, j) in the pass-th printing round, represents the average number of available nozzles in all landing areas of pixel slot (i, j);

[0051] r″3=Ψ(s)

[0052] Among them, Ψ(s) is the printing completion index, r″1, r″2 and r"3 are the three reward functions of the system in the multi-round scheduling stage.

[0053] Optionally, the global potential energy function Φ=Φ1+Φ2+Φ3;

[0054] Φ1=(1-var v )*ω

[0055] Φ2=(1-var d )*η

[0056]

[0057] Among them, Φ1 is the variance of droplet volume difference, var v represents the variance of the droplet volume difference in F pixel slots, Φ2 is the variance of the droplet landing point distance, var d represents the variance of the droplet landing distance of F pixel slots, K i，j is the number of locations where ink droplets can land in the pixel slot (i, j), var represents the variance of the product of the average number of times the nozzles of the F pixel slots are used and the number of usable nozzles, and ω, η, and δ are weight parameters.

[0058] Optionally, before solving the Markov decision process model for the nozzle optimization using the P-TLQ algorithm, training the P-TLQ algorithm includes:

[0059] S31, initialization parameters, using random strategy to determine the multi-target initial threshold T u ;

[0060] S32, calculating the current state value s according to the Markov decision process model of the nozzle optimization;

[0061] S33. Calculate the threshold T through the reward function in the Markov decision process model u Update the initial threshold and obtain the available action set A, T according to the threshold u represents the reward threshold of the u-th target;

[0062] S34: Select an action combination from the action set A according to the TLO algorithm at the state value s, calculate the reward of the action combination, and obtain the next state s′ and the reward value TR s，a，u , TR s，a，u =min(R s，a，u , T u ), where R s，a，u represents the u target reward function component obtained by taking action a in state s;

[0063] S35. Calculate the global situation energy function Φ and add it to the reward value, TR′ s，a，u =TR s，a，u +Φ;

[0064] S36, will<s,a,TR′,s′> Store in the experience collection;

[0065] S37, extracting a set of data from the experience set, inputting the data into the DQN neural network, and updating the DQN neural network parameters using the gradient;

[0066] S38. Update TR″ based on the additional reward function G threshold s，a，u =TR′ s，a，u+G, G(s, s′) = γΦ(s′) - Φ(s), where γ is a manually set discount factor;

[0067] S39, s = s', update T u =TR″ s，a，u , the number of training times increases by one;

[0068] S310: If the maximum number of training times is reached, the optimal action combination, each optimization target value, and DQN network parameters are recorded; otherwise, the process returns to step S33.

[0069] The beneficial effects of the present invention are as follows: the present invention decomposes the optimization problem of the trillion-level OLED macro-inkjet printing task into multiple rounds of scheduling optimization according to the printing process (one round is defined as one time when the nozzle scans the substrate). In each round, the multi-objective optimization problem is decomposed into two scheduling optimization problems: the nozzle optimization problem and the trillion-level ink droplet injection optimization problem. In the trillion-level ink droplet injection optimization problem, the nozzle is not used as the main body to select which nozzles will spray which landing areas at a certain moment t. Instead, the pixel slot selects which nozzles will spray ink droplets at which moments, so as to reduce the dimension of optimization. At the same time, in order to solve the reward sparseness problem that is difficult to evaluate the overall optimization effect when optimizing only the pixel slot as the main body, a reward shaping method is adopted to improve the overall optimization effect. In view of the action characteristics of the OLED printing task, a multi-objective reinforcement learning method is designed that can effectively solve high spatial dimensions and global reward sparsity. This method improves the DQN algorithm based on threshold lexicographic sorting (TLO), combines it with the idea of ​​potential-based reward shaping (PBRS), designs an effective printing potential function, realizes multi-objective optimization of scheduling tasks, and reduces the algorithm oscillation time, thereby accelerating the convergence of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0071] Figure 1 This is a flow chart of an optimization calculation method for exascale OLED macro-inkjet printing according to an embodiment of the present invention;

[0072] Figure 2 Flowchart of the P-TLQ algorithm according to an embodiment of the present invention. DETAILED DESCRIPTION

[0073] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0074] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0075] The optimization problem of exascale OLED macro-inkjet printing task can be described as:

[0076] First, based on the pixel slot arrangement and pixel slot size data on the display panel and the process requirements, a set of areas {Area} within the pixel slot where the pixel can land is established. Then, based on the nozzle arrangement of the selected different nozzle models, as well as the spray volume and trajectory of each nozzle, given different panel pixel sizes and structures, how to optimize the number of times the nozzle scans the substrate, the nozzle angle α during each scan, and the nozzle movement speed v s , the X-axis distance between the nozzle and the substrate And determine which nozzles will spray at each sprayable moment under the premise of a given nozzle spraying frequency f during each scanning process, and optimize the following multiple goals with preferences (sorted by importance) under the premise that the sprayed ink droplets can fall into {Area}: 1) reduce the total ink volume difference between all pixel slots; 2) reduce the uneven droplet spacing in the pixel slot; 3) shortest printing time; 4) reduce the unbalanced use of nozzles.

[0077] This example decomposes the optimization problem of exascale OLED macro-inkjet printing tasks into multiple rounds of scheduling optimization (one round is a single pass of the nozzle over the substrate). In each round, the multi-objective optimization problem is decomposed into two scheduling optimization problems: 1) The nozzle optimization problem: determining the nozzle deflection angle, the nozzle's initial X-axis position, and the nozzle's movement speed; 2) The exascale ink droplet ejection optimization problem: Based on the first-stage scheduling results, which determine the set of droplets that can land in each pixel slot, as the nozzle scans the substrate in this round, the hundreds of millions of pixel slots select the exascale ink droplets ejected by the tens of thousands of nozzles at different ejection times. If all pixel slots have not been printed in the Lth round, the nozzle is scheduled to step X distance in the X direction to start printing in the L+1 round. In this problem, the nozzle is not the main body that selects which nozzles will eject which droplet areas at a certain time t. Instead, the pixel slots select the ink droplets ejected by the nozzles at which times, reducing the optimization dimensionality. At the same time, in order to solve the problem of sparse rewards caused by optimizing only pixel slots and making it difficult to evaluate the overall optimization effect, a reward shaping method is adopted to improve the overall optimization effect.

[0078] This example models the printhead optimization problem, the exascale ink droplet ejection optimization problem, and the printhead step scheduling problem as Markov decision models, and designs corresponding P-TLQ algorithms to solve these three scheduling problems. The specific steps are as follows:

[0079] like Figure 1 As shown, this embodiment provides an optimization calculation method for exascale OLED macro-inkjet printing, including:

[0080] S1. Initialize the nozzle parameters and pixel slot parameters, and set the optimization target;

[0081] S2. Constructing Markov decision process models for printhead optimization, exascale droplet ejection optimization, and printhead step scheduling;

[0082] S3. Use the P-TLQ algorithm to solve the Markov decision process model for printhead optimization, obtain the printhead deflection angle, the initial X-axis position of the printhead, the printhead movement speed, and the set of droplets that can land in the pixel slot. The P-TLQ algorithm is a DQN algorithm that combines the global situation energy function and threshold dictionary sorting.

[0083] S4. Based on the set of droplets that can land in the pixel slot, the Markov decision process model for optimizing the ejection of exascale droplets is solved using the P-TLQ algorithm to obtain the ejection nozzle combination selection for the pixel slot;

[0084] S5. If the printing of all pixel slots is not completed, the Markov decision process model of the nozzle step scheduling is solved using the P-TLQ algorithm to obtain the nozzle X-direction step scheduling, and return to step S3. If the printing of all pixel slots is completed, otherwise the printing is terminated.

[0085] Furthermore, the nozzle parameters include: nozzle arrangement, nozzle spraying frequency and nozzle movable speed range;

[0086] Pixel slot parameters include: substrate pixel slot size / arrangement and ink droplet landing area within the pixel slot;

[0087] Setting optimization goals includes: setting the target volume of ink droplets in a single pixel slot, the difference in total ink volume between all pixel slots, the unevenness of droplet spacing in the pixel slot, printing time and nozzle imbalance.

[0088] Furthermore, the Markov decision process model for nozzle optimization constructed in S2 includes:

[0089] 1) Establish the state space of the nozzle optimization stage:

[0090] When establishing the state space, it is necessary to consider the impact of different OLED substrate printing requirements on the feature space. This paper adds initial conditions under different requirements to the state. With hundreds of millions of pixel slots as the decision-making entities, the state space of the system is defined as:

[0091]

[0092] in, Represents the state space of the nozzle optimization stage, Represents the volume of droplets to be sprayed in the pixel slot (i, j) that needs to be sprayed in the current production stage, Represents the distribution state of the ejected droplets in the pixel slot (i, j) in the current production stage, V′ i，j Represents the volume of liquid ejected by the nozzle in a single shot, h a，b Represents the number of times the nozzle has sprayed, Δt a，b Represents the time interval between the nozzle and the last injection, N pass represents the number of available nozzles in the pass-th printing round, a and b represent the nozzles in the a-th row and b-th column, and sub represents the set of pixel slots that need to be sprayed.

[0093] 2) Establish the action space of the nozzle optimization stage:

[0094] Based on the billion-level pixel slot And the nozzle movement path relative to the substrate, the nozzle set P swept on the pixel slot can be obtained sub (n), in order to select the appropriate nozzle combination, the action space is established as follows:

[0095]

[0096] Among them, A sub Represents the action space of the nozzle optimization stage, is the nozzle combination selected for pixel slot (i, j) in the current production stage, then Nozzle All constraints should be met. P(n) is the set of pixel slots swept by all nozzles, and n is the number of all nozzles;

[0097] 3) The reward function in the nozzle optimization stage is:

[0098] The reward consists of three parts:

[0099]

[0100] Among them, r1, r2, and r3 are the volume difference reward, the droplet landing point uniform distribution reward, and the nozzle balanced use reward respectively. V i，j is the volume of droplets to be ejected in all pixel slots (i, j), To meet the standard volume, Indicates the number of available nozzles within the effective area of ​​the kth landing point in the pixel slot (i, j), represents the average number of droplets in the four effective areas of the pixel slot (i, j), max(l i，j ) represents the maximum value among the distances between each landing point, l ε represents the optimal placement distance, τ a，b is the number of times the nozzle numbered (a, b) in the pixel slot (i, j) is used, is the average number of times all nozzles in the pixel slot (i, j) are used, K i，j is the number of locations where ink droplets can be placed in the pixel slot (i, j), p i，j is the nozzle set that can spray pixel slot (i, j), p a，b It is the subset of nozzles that are actually ejected, selected from the nozzles that can eject pixel slot (i, j).

[0101] Furthermore, the Markov decision process model for exascale droplet ejection optimization in S2 includes:

[0102] 1) Constructing the state space for the exascale ink droplet ejection optimization phase:

[0103] Nozzle scheduling still uses pixel slots as the decision-making body, and the state space of the system is defined as:

[0104] S 2，t ={V t ,D t ,V ts ,h t,Δt,N t}

[0105] Among them, S 2，t Optimizing the state space for exascale droplet ejection, V t represents the volume of ink droplets to be sprayed in the pixel slot at time t, D t Represents the distribution state of the ink droplets sprayed in the pixel slot at time t, V ts represents the volume of liquid sprayed by the nozzle at time t, h t represents the number of times the nozzle has sprayed at time t, Δt represents the interval from the nozzle to the last spraying, N t represents the number of nozzles passing through the pixel slot in this printing round. The subscript t represents the time at which the nozzle can spray when the spray frequency is F, t = 0, 1 / F, 2 / F, ...;

[0106] 2) Establishing the action space for the exascale ink droplet ejection optimization phase:

[0107] Based on the pixel slot, 2，t As well as the nozzle motion path relative to the substrate, we can obtain the nozzle set P(n) that sweeps over the pixel slot. To select the appropriate nozzle combination, we establish the action space as follows:

[0108] A 2，t ={P c (n)∈P(n),n∈(0,a*b-1)}

[0109] Among them, A 2，t The action space for the optimization phase of exascale ink droplet ejection, P c (n) is the nozzle combination selected for the cth pixel slot, a and b represent the nozzles in the ath row and bth column on the nozzle, P(n) is the set of pixel slots swept by all nozzles, and n is the number of all nozzles;

[0110] 3) The reward function for the exascale ink droplet ejection optimization phase is:

[0111] The reward function consists of three parts:

[0112]

[0113] Among them, V i，j is the volume of droplets to be ejected in all pixel slots (i, j), Indicates the qualified volume;

[0114]

[0115] in, is the number of nozzles within the effective area of ​​the kth landing point in the (i, j)th pixel slot, is the average number of nozzles of the four dots in the (i, j)th pixel slot;

[0116]

[0117] Among them, r′1, r′2, and r′3 are the three reward functions for exascale ink droplet ejection, is the average number of times all nozzles in the (i, j)th pixel slot are used, and K is the number of positions in the pixel slot where ink drops can land.

[0118] Furthermore, the Markov decision process model for nozzle step scheduling is constructed including:

[0119] 1) Establishing the state space of the multi-round scheduling phase:

[0120]

[0121] Among them, S mul Represents the state space of the multi-round scheduling phase, pass represents the pass-th printing round, Indicates the number of pixel slots filled in the pass printing round, Indicates the number of pixel slots covered by the nozzle track during the pass printing round. represents the variance of the number of available nozzles in different pixel slots during the printing pass;

[0122] 2) Establishing the system action space for the multi-round scheduling phase:

[0123] A mul ={L pass ±ΔL}

[0124] Among them, A mul is the system action space in the multi-round scheduling phase, ΔL is the minimum distance the nozzle moves laterally, and L pass It is the lateral movement of the nozzle in the Lth round relative to the L-1th round, i.e. the adjacent round;

[0125] 3) The reward function of the system in the multi-round scheduling phase is:

[0126] The reward consists of three parts:

[0127]

[0128] Among them, N s ∈[0, M·N] represents the number of pixel slots that the inkjet head can fill in the pass-th printing round, M and N are the total number of rows and columns of pixel slots;

[0129] Represents the difference in the number of available nozzles for different pixel slots, where Indicates the number of nozzles available in the kth landing area of ​​pixel slot (i, j) in the pass-th printing round, Indicates the number of positions within pixel slot (i, j) where an ink droplet can land;

[0130] r″3=Ψ(s), which is the reward given when the printing completion index is met;

[0131] Among them, r1″, r2″ and r3″ are the three reward functions of the system in the multi-round scheduling stage, and Ψ(s) is the printing completion index, which is equal to the number of pixel slots that can complete printing / the total number of pixel slots.

[0132] Furthermore, based on the DQN algorithm based on threshold lexicographic ordering (TLQ), based on the idea of ​​reward shaping, an OLED printing scheduling potential function and a DQN algorithm based on PBRS (Potential-based reward shaping) (P-TLQ) were designed.

[0133] 1) Design of OLED printing scheduling potential energy function Φ

[0134] Aiming at the three optimization goals of the OLED scheduling problem, a global situation energy function Φ = Φ1 + Φ2 + Φ3 is designed to ensure that the global indicators meet expectations:

[0135] a) Variance of droplet volume difference: Φ1=(1-var v )*ω,var v It represents the variance of droplet volume differences in F pixel slots, where F represents the number of pixels within the coverage range of the nozzle at a certain moment as a typical value.

[0136] b) Variance of droplet landing distance: Φ2=(1-var d )*η,var d represents the variance of the droplet landing distance of F pixel slots.

[0137] c) Variance of nozzle usage times: is the average number of times all nozzles in pixel slot (i, j) are used, and var represents the variance of the product of the average number of times the nozzles in F pixel slots are used and the number of available nozzles.

[0138] Among them, ω, η, and δ are weight parameters set based on experience.

[0139] 2) P-TLQ algorithm:

[0140] Due to the dynamic nature of the printing optimization problem and the difficulty of determining static Q-values, this embodiment employs a dynamic threshold design approach for a quantified reward function, making it easier for the algorithm to converge to the desired goal. Furthermore, the thresholds in the TLO algorithm are dynamically adjusted using PBRS, and the additional reward G from PBRS is used as a guide to improve the updating of the secondary goal reward threshold. This allows users to simply declare their preferences without having to manually change the thresholds. This method is called the PBRS-based threshold lexicographic ranking (DQN) algorithm.

[0141] According to the state space, action space, and reward function of the three Markov decision models established above, three corresponding P-TLQs are constructed respectively. Each corresponding P-TLQ is trained according to the following steps; Figure 2 As shown, the steps of training the P-TLQ algorithm include:

[0142] 1) Initialize the relevant parameters of the nozzle and substrate such as Figure 1 The first and second steps in the above are to use random strategy to determine the initial threshold T of multiple targets. u ;

[0143] 2) starting task scheduling according to the established Markov decision process model and calculating the current state value s, wherein the Markov decision process model is a Markov decision process model for printhead optimization, a Markov decision process model for exascale ink droplet ejection optimization, or a Markov decision process model for printhead step scheduling;

[0144] 3) Calculate the threshold T through the reward function in the Markov decision process model u Update the initial threshold and obtain the available action set A, T according to the threshold u represents the reward threshold (minimum acceptable amount) for the u-th goal;

[0145] 4) Select action combination a in A according to the TLO algorithm in state value s k , calculate the reward of the selected action combination, and get the next state s′ and reward value TR s，a，u .TR s，a，u =min(R s，a，u , T u ), where R s，a，u represents the target reward function component u obtained by taking action a in state s.

[0146] 5) Calculate the global energy function Φ and add it to the reward value, TR′ s，a，u =TR s，a，u +Φ;

[0147] 6)<s,a,TR′,s′> Store in the experience collection;

[0148] 7) Take a set of data from the experience set, input it into the DQN neural network, and use the gradient to update the DQN network parameters;

[0149] 8) Adjust TR″ according to the additional reward function G s，a，u =TR′ s，a，u +G, G(s, s′) = γΦ(s′) - Φ(s), where γ is a manually set discount factor;

[0150] 9) s = s′, update T u =TR″ s，a，u , the number of training times increases by one;

[0151] 10) If the maximum number of training times is reached, record the optimal action combination, each optimization target value, and DQN network parameters; otherwise, return to step 3).

[0152] The DQN neural network in this embodiment typically consists of one input layer, three hidden layers, and one output layer. The number of neurons in the input layer is equal to the dimension of the state S; the hidden layer is a fully connected network with a maximum of 64 neurons; the output layer contains the Q value of each possible action, with the number of neurons equal to the dimension of the action a.

[0153] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A method for optimizing the calculation of exascale OLED macro-inkjet printing, characterized in that: include: S1. Initialize the nozzle parameters and pixel slot parameters, and set the optimization target; S2. Constructing Markov decision process models for printhead optimization, exascale droplet ejection optimization, and printhead step scheduling; S3. Solve the Markov decision process model for the nozzle optimization using the P-TLQ algorithm to obtain the nozzle deflection angle, the initial X-axis position of the nozzle, the nozzle movement speed, and the set of droplets that can land in the pixel slot. The P-TLQ algorithm is a DQN algorithm that combines a global situation energy function and a threshold lexicon sorting algorithm. S4. Based on the set of droplets that can land in the pixel slot, the Markov decision process model for optimizing the exascale ink droplet ejection is solved using the P-TLQ algorithm to obtain a nozzle combination selection for the pixel slot; S5. If printing of all pixel slots is not completed, the Markov decision process model of the nozzle step scheduling is solved using the P-TLQ algorithm to obtain the nozzle X-direction step scheduling, and return to step S3. If printing of all pixel slots is completed, otherwise printing ends.

2. The exascale OLED macro-inkjet printing optimization calculation method according to claim 1, characterized in that: The nozzle parameters include: nozzle arrangement, nozzle spraying frequency and nozzle movable speed range; The pixel slot parameters include: substrate pixel slot size / arrangement and ink droplet landing area within the pixel slot; Setting the optimization target includes: setting the target volume of ink droplets in a single pixel slot, the difference in total ink volume between all pixel slots, the unevenness of droplet spacing in the pixel slots, the printing time, and the imbalance of the nozzles.

3. The exascale OLED macro-inkjet printing optimization calculation method according to claim 1, characterized in that: The Markov decision process model for nozzle optimization in S2 includes: Establish the state space of the nozzle optimization stage: in, Represents the state space of the nozzle optimization stage, Represents the volume of droplets to be sprayed in the pixel slot (i, j) that needs to be sprayed in the current production stage, Represents the distribution state of the ejected droplets in the pixel slot (i, j) in the current production stage, V′ i，j Represents the volume of liquid ejected by the nozzle in a single shot, h a，b Represents the number of times the nozzle has sprayed, Δt a，b Represents the time interval between the nozzle and the last injection, N pass represents the number of available nozzles in the pass, a and b represent the nozzles in row a and column b, and sub represents the set of pixel slots that need to be sprayed; Establish the action space for the nozzle optimization stage: Among them, A sub Represents the action space of the nozzle optimization stage, is the nozzle combination selected for pixel slot (i, j) in the current production stage, P(n) is the set of pixel slots swept by all nozzles, and n is the number of all nozzles; The system reward function in the nozzle optimization stage is: Among them, r1, r2, and r3 are the volume difference reward, the droplet landing point uniform distribution reward, and the nozzle balanced use reward respectively. V i，j is the volume of droplets to be ejected in all pixel slots (i, j), To meet the standard volume, represents the number of droplets within the effective area of ​​the kth landing point in the pixel slot (i, j), represents the average number of droplets in the four effective areas of the pixel slot (i, j), max(l i，j ) represents the maximum value among the distances between each landing point, l ε represents the optimal placement distance, τ a，b is the number of times the nozzle numbered (a, b) in the pixel slot (i, j) is used, is the average number of times all nozzles in the pixel slot (i, j) are used, K i，j is the number of locations where ink droplets can be placed in the pixel slot (i, j), p a，b is the subset of nozzles that are actually ejected from the nozzles that can eject pixel slot (i, j), p i，j is the set of nozzles that can spray pixel slot (i, j).

4. The exascale OLED macro-inkjet printing optimization calculation method according to claim 3, characterized in that: The Markov decision process model for optimizing exascale droplet ejection in S2 includes: Constructing the state space for the exascale droplet ejection optimization phase: S 2，t ={V t ,D t ,V ts ,h t ,Δt,N t } Among them, S 2，t Optimizing the state space for exascale droplet ejection, V t represents the volume of ink droplets to be sprayed in the pixel slot at time t, D t Represents the distribution state of the ink droplets sprayed in the pixel slot at time t, V ts represents the volume of liquid sprayed by the nozzle at time t, h t represents the number of times the nozzle has sprayed at time t, Δt represents the interval from the nozzle to the last spraying, N t Represents the number of pixel slot nozzles that have passed through this printing round; Establishing the action space for the exascale droplet ejection optimization phase: A 2，t ={P c (n)∈P(n),n∈(0,a*b-1)} Among them, A 2，t The action space for the optimization phase of exascale ink droplet ejection, P c (n) is the nozzle combination selected for the cth pixel slot, a and b represent the nozzles in the ath row and bth column on the nozzle, P(n) is the set of pixel slots swept by all nozzles, and n is the number of all nozzles; The system reward function in the exascale ink droplet ejection optimization phase is: in, is the number of nozzles within the effective area of ​​the kth landing point in the (i, j)th pixel slot, is the average number of nozzles of the four dots in the (i, j)th pixel slot; Among them, r′1, r′2, and r′3 are the three reward functions for exascale ink droplet ejection, is the average number of times all nozzles in the (i, j)th pixel slot are used, p a，b is the subset of nozzles that are actually ejected from the nozzles that can eject pixel slot (i, j), p i，j is the set of nozzles that can spray pixel slot (i, j).

5. The exascale OLED macro-inkjet printing optimization calculation method according to claim 1, characterized in that: The Markov decision process model for nozzle step scheduling is constructed as follows: Establish the state space of the multi-round scheduling phase: Among them, S mul Represents the state space of the multi-round scheduling phase, pass represents the pass-th printing round, Indicates the number of pixel slots filled in the pass printing round, Indicates the number of pixel slots covered by the nozzle track during the pass printing round. represents the variance of the number of available nozzles in different pixel slots during the printing pass; Establish the action space for the multi-round scheduling phase: A mul ={L pass ±ΔL} Among them, A mul is the action space in the multi-round scheduling phase, ΔL is the minimum distance the nozzle moves laterally, and L pass is the lateral movement of the sprinkler head between adjacent rounds; The reward function in the multi-round scheduling phase is: Among them, N s ∈[0, M·N] represents the number of pixel slots that the inkjet head can fill in the pass-th printing round, M and N are the total number of rows and columns of pixel slots; in, Indicates the number of nozzles available in the kth landing area of ​​pixel slot (i, j) in the pass-th printing round, represents the average number of available nozzles in all landing areas of pixel slot (i, j); r″3=Ψ(s) Among them, Ψ(s) is the printing completion index, r″1, r″2 and r″3 are the three reward functions of the multi-round scheduling system.

6. The exascale OLED macro-inkjet printing optimization calculation method according to claim 3, characterized in that: The global potential energy function Φ=Φ1+Φ2+Φ3; Φ1=(1-var v )*oh Φ2=(1-var d )*or Among them, Φ1 is the variance of droplet volume difference, var v represents the variance of the droplet volume difference in F pixel slots, Φ2 is the variance of the droplet landing point distance, var d represents the variance of the droplet landing distance of F pixel slots, K i，j is the number of locations where ink droplets can land in the pixel slot (i, j), var represents the variance of the product of the average number of times the nozzles of the F pixel slots are used and the number of usable nozzles, and ω, η, and δ are weight parameters.

7. The exascale OLED macro-inkjet printing optimization calculation method according to claim 6, characterized in that: Before solving the Markov decision process model for the nozzle optimization using the P-TLQ algorithm, training the P-TLQ algorithm includes: S31, initialization parameters, using random strategy to determine the multi-target initial threshold T u ; S32, calculating the current state value s according to the Markov decision process model of the nozzle optimization; S33. Calculate the threshold T through the reward function in the Markov decision process model u Update the initial threshold and obtain the available action set A, T according to the threshold u represents the reward threshold of the u-th target; S34: Select an action combination from the action set A according to the TLO algorithm at the state value s, calculate the reward of the action combination, and obtain the next state s′ and the reward value TR s，a，u , TR s，a，u =min(R s，a，u , T u ), where R s，a，u represents the u target reward function component obtained by taking action a in state s; S35. Calculate the global situation energy function Φ and add it to the reward value, TR′ s，a，u =TR s，a，u +Φ: S36, will<s,a,TR′,s′> Store in the experience collection; S37, extracting a set of data from the experience set, inputting the data into the DQN neural network, and updating the DQN neural network parameters using the gradient; S38. Update TR″ based on the additional reward function G threshold s，a，u =TR′ s，a，u +G, G(s, s′) = γΦ(s′) - Φ(s), where γ is a manually set discount factor; S39, s = s', update T u =TR″ s，a，u , the number of training times increases by one; S310: If the maximum number of training times is reached, the optimal action combination, each optimization target value, and DQN network parameters are recorded; otherwise, the process returns to step S33.

Images