A backplane arrangement management method based on a QD-OLED panel

Abstract

The application discloses a backboard arrangement management method based on a QD-OLED panel and belongs to the technical field of display panel design. The method first acquires panel base material, TFT device and quantum dot parameters, generates an initial arrangement through computational fluid dynamics topology optimization with current density uniformity and heat dissipation consistency as constraints; then constructs a correlation graph model with electrical coupling, optical crosstalk and heat conduction as weights, and completes iterative optimization through space-time graph convolution and multi-head self-attention; feeds back luminous efficiency and color coordinate offset data to constrain the network, and realizes multi-objective balance by combining reinforcement learning and Nash equilibrium; finally generates a backboard arrangement scheme, and distributes driving timing based on time division multiple access and queuing theory to smooth transient current. The application realizes joint optimization of electric, thermal and optical multi-physical fields, suppresses sub-pixel color deviation and uneven decay, and improves backboard reliability and display effect.

Description

Technical Field

[0001] This invention discloses a backplane layout management method based on a QD-OLED panel, belonging to the field of display panel design technology. Background Technology

[0002] The conventional design of backplane layout for quantum dot organic light-emitting diode (OLED) panels primarily employs a single-dimensional optimization strategy. In practice, designers typically separate the routing of traces on the backplane substrate and the arrangement of thin-film transistor (TFT) units into independent electrical and thermal performance verification processes. For electrical performance, conventional solutions aim to meet preset current density distribution targets by altering trace widths or adjusting TFT node positions. For thermal performance, conventional solutions plan heat dissipation paths within the layout gaps to reduce localized temperatures. This single-dimensional design typically uses empirical formulas combined with finite element simulation software for iterative processing. Electrical and thermal simulations are independent of each other, with no data interaction or joint constraints between them.

[0003] Based on the aforementioned single-dimensional optimization design approach, the Joule heat generated by the trace current during actual panel operation, the electrical coupling between thin-film transistor nodes, and the optical crosstalk between quantum dot sub-pixels are all simultaneous and superimposed physical fields. Therefore, the independent simulations that separate electrical and thermal aspects cannot solve the technical problems of inconsistent light emission attenuation and color coordinate shift of quantum dot sub-pixels in different regions of the backplane under the coupling effect of multiple physical fields. While adjusting the trace topology can improve the local current distribution, it alters the Joule heat distribution, leading to new optical crosstalk. Consequently, the superimposed interference from multiple physical fields cannot be eliminated during the layout design stage. Summary of the Invention

[0004] The purpose of this invention is to provide a backplane arrangement management method based on QD-OLED panels, which can effectively solve the problems in the background art mentioned above.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for backplane layout management based on QD-OLED panels, comprising: Obtain the physical parameters of the backplane substrate, the electrical parameters of the TFT devices, and the spectral and luminous decay characteristics of the quantum dot sub-pixels of the target QD-OLED panel; With the uniformity of current density and the consistency of heat dissipation flow field distribution of the backplane traces as constraints, the initial backplane trace and TFT cell layout topology is generated by computational fluid dynamics topology optimization. Each TFT unit and sub-pixel driving node are treated as graph nodes. The electrical coupling, optical crosstalk, and thermal conduction correlation between nodes are used as edge weights to construct a backplane layout correlation graph model. The node positions and connection relationships of the initial layout topology are iteratively optimized. Data on the luminous efficiency and color coordinate offset of sub-pixels under different arrangement positions are collected, and the collected data is fed back as a constraint term into the loss function and boundary conditions of the graph convolutional neural network for topology optimization. Generate a backplane layout scheme adapted to the target QD-OLED panel, and output the corresponding driving timing control parameters of the layout scheme to the display driving control unit.

[0006] Preferably, the initial backplane routing and TFT cell layout topology is generated through computational fluid dynamics topology optimization, including: The backplane routing network is mapped as a virtual fluid micropipe network, the current density distribution is equivalent to the fluid pressure field inside the pipe, and the heat dissipation distribution is equivalent to the fluid velocity field inside the pipe. Based on the microgrid power flow scheduling strategy, virtual fluid particles are injected into the virtual fluid micropipe network, and a pipe resistance evaluation function is constructed with the uniformity of current density and the consistency of heat dissipation flow field distribution. By utilizing the particle swarm optimization strategy in multi-agent reinforcement learning, virtual fluid particles are guided to navigate between pipeline nodes. After convergence, the initial backplane routing and TFT cell layout topology are generated.

[0007] Preferably, a backplane layout association diagram model is constructed and iteratively optimized, including: The electrical coupling, optical crosstalk, and thermal conduction correlation between nodes are extracted to construct an initial adjacency matrix that integrates spatial proximity and physical field strength. A dynamic congestion propagation model is introduced in the traffic network, and the superimposed interference of multi-physics fields is regarded as the congestion load in the graph network. The congestion state vector of each graph node at different time steps is calculated. Spatiotemporal graph convolutional networks are used to extract the spatiotemporal features of congestion state vectors. By alternating mapping between graph convolutional layers and temporal convolutional layers, the edge weights in the adjacency matrix are dynamically updated to complete the iterative optimization of node positions and connectivity.

[0008] Preferably, the collected data is fed back as a constraint term into the loss function of the graph convolutional neural network, including: Map the luminous efficiency and color coordinate offset data of the sub-pixels at the current arrangement position to the environment state space in the reinforcement learning framework; Construct a multi-objective reward function that includes a reward term for maximizing luminous efficiency, a penalty term for minimizing color coordinate offset, and a term for maintaining backplane trace current carrying capacity margin. The gradient of the multi-objective reward function with respect to the weights of the graph convolutional neural network is calculated using the policy gradient algorithm. The weight coefficients of each constraint in the loss function are adaptively adjusted to complete the backpropagation update of the network parameters.

[0009] Preferably, the driving timing control parameters corresponding to the layout scheme are output to the display driving control unit, including: Identify the current carrying capacity distribution of traces in each area of ​​the backplane layout scheme. By combining the time division multiple access resource scheduling mechanism in the communication network, the charging and discharging operations of sub-pixels are encapsulated into independent data frames; Based on the current carrying margin distribution map, non-overlapping write time slots and staggered discharge timing are allocated to the sub-pixel data frames of different regions. The generated time slot allocation table is output as the driving timing control parameter to smooth the transient current peaks in the backplane routing network.

[0010] Preferably, the virtual fluid particle path navigation is guided by a particle swarm optimization strategy in multi-agent reinforcement learning, including: Encode the path node sequence in the virtual fluid microtube network into a chromosome gene chain in a genetic algorithm; The reciprocal of the pipeline resistance evaluation function is used as the chromosome fitness, and a roulette wheel selection mechanism is used to select high-fitness particle paths. Multi-point crossover and Gaussian mutation operations are performed on the chromosome gene chains corresponding to the selected particle paths to generate a new generation of particle walking paths, breaking the local optimum unbinding to converge to the global optimum initial arrangement topology.

[0011] Preferably, the spatiotemporal features of the congestion state vector are extracted using a spatiotemporal graph convolutional network, including: Embed a multi-head self-attention mechanism from natural language processing in the feature mapping stage of a spatiotemporal graph convolutional network; The electrical coupling, optical crosstalk, and thermal conduction correlation are projected onto different attention heads, and the self-attention weight matrix of each physical field feature between graph nodes is calculated independently. The multi-head self-attention weight matrix is ​​concatenated and linearly transformed to generate an adaptive feature vector containing multi-physics-dominated interference factors, which replaces the fixed-weight adjacency matrix in spatiotemporal feature extraction.

[0012] Preferably, a multi-objective reward function is constructed, including: Maximizing luminous efficiency, minimizing color coordinate offset, and maintaining current carrying capacity are defined as three competing game participants; Based on the non-cooperative game model in economics, calculate the marginal revenue function of each game participant under the current topology. With the goal of finding the Nash equilibrium point of the marginal revenue function, the Nash equilibrium constraint is embedded as a regularization term into the multi-objective reward function to suppress the predatory suppression of the performance improvement of a single physical field on the performance of other physical fields.

[0013] Preferably, allocating write time slots for sub-pixel data frames based on the current carrying capacity margin distribution map includes: The write request of sub-pixel data frame is regarded as the customer arrival flow in the queuing theory model of operations research, and the current carrying capacity of the backplane wiring is regarded as the service desk processing rate. Based on the M / G / 1 queuing model, the variance of the arrival time interval of sub-pixel data frames in each region and the variance of the service processing time are statistically analyzed. The optimal queuing threshold for each region is calculated using the steady-state distribution analytical formula of the M / G / 1 queuing model. Based on the optimal queuing threshold, the write time slots of low-priority sub-pixel data frames are dynamically truncated or delayed.

[0014] Preferably, the self-attention weight matrix of each physical field feature among the graph nodes is calculated independently, including: Obtain the difference between the real-time color coordinate offset error and the set target error in the current display screen; The difference is used as the input signal of the proportional-integral-derivative controller in classical control theory, and dynamic adjustment parameters are generated through a linear combination of proportional, integral, and derivative functions. By dynamically adjusting the parameters of the key-value matrix in the multi-head self-attention mechanism, closed-loop self-tuning of the graph network edge weights is achieved as the display content changes, thus eliminating the lag in the adaptation of static attention weights to dynamic content.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a graph model of thin-film transistor units and sub-pixel driving nodes, and using electrical coupling, optical crosstalk, and thermal conductivity correlation as edge weights for graph convolutional iterative optimization, this method unifies the physical field interferences of the electrical, thermal, and optical dimensions into a single network structure for joint calculation. This multi-physics coupling-based optimization breaks down the barriers of single-dimensional design, eliminating the superimposed effects of electrical coupling, thermal aggregation, and optical crosstalk in the backplane arrangement on the uniformity of light emission in the quantum dot organic light-emitting diode panel, and solving the problems of inconsistent sub-pixel light emission attenuation and color coordinate shift caused by multi-physics coupling.

[0016] 2. Initial layout is generated using computational fluid dynamics topology optimization. Sub-pixel luminous efficiency and color coordinate offset data are then fed into the loss function of the graph convolutional neural network as constraints. This ensures that the layout scheme balances the current carrying stability of the backplane traces and the calibration requirements of the quantum dot sub-pixels during the iteration process. The layout scheme is then converted into driving timing control parameters and output to the display driver control unit. By allocating non-overlapping write time slots and staggered discharge timings, transient current peaks in the backplane trace network are smoothed out, shortening the design cycle of the backplane layout scheme. Attached Figure Description

[0017] Figure 1 This is the overall flowchart for QD-OLED backplane layout management of the present invention; Figure 2 A flowchart for generating the initial layout for CFD topology optimization in this invention; Figure 3 This is a flowchart of the iterative optimization process for the backplane association diagram model of the present invention; Figure 4 This is a flowchart of the multi-objective reinforcement learning feedback constraint process of the present invention; Figure 5 This is a flowchart illustrating the generation of drive timing control parameters for the present invention. Figure 6 This is a flowchart of the multi-head self-attention and PID closed-loop control process of the present invention. Detailed Implementation

[0018] Please refer to the attached document. Figure 1 This embodiment provides a backplane layout management method based on a QD-OLED panel, running on a QD-OLED panel design simulation system. The system includes a panel parameter acquisition unit, a computational fluid dynamics (CFD) topology optimization unit, a graphical model construction and iterative optimization unit, a multiphysics simulation and data acquisition unit, and a timing control parameter generation unit. The system is coupled to the display driver control unit of the target QD-OLED panel. The display driver control unit includes a timing controller (TCON), a source driver chip, and a gate driver chip, used to receive and execute the driving timing control parameters output by the timing control parameter generation unit.

[0019] In a preferred embodiment, the panel parameter acquisition unit first acquires the physical parameters of the backplane substrate, the electrical parameters of the TFT devices, and the spectral and luminous decay characteristics of the quantum dot sub-pixels of the target QD-OLED panel.

[0020] The physical parameters of the backplate substrate include the substrate dielectric constant. Thermal conductivity λ, glass transition temperature Tg, Young's modulus E, substrate thickness The substrate surface roughness Ra; the electrical parameters of the TFT device include TFT channel carrier mobility μ, threshold voltage Vth, on-state current Ion, off-state current Ioff, channel width-to-length ratio W / L, gate-source parasitic capacitance Cgs, gate-drain parasitic capacitance Cgd, drain-source parasitic capacitance Cds, and breakdown voltage Vbr; the spectral characteristic parameters of the quantum dot sub-pixels include the peak emission wavelengths of red, green, and blue light sub-pixels. , , emission spectrum half width at half maximum (FWHM) , , Standard color coordinates under standard driving conditions , , external quantum efficiency , , The luminescence decay characteristic parameters of the quantum dot sub-pixels include the brightness decay half-life at rated current density. Brightness attenuation coefficient k, Joule thermally induced attenuation acceleration factor The efficiency roll-off factor γ induced by changes in current density.

[0021] Secondly, the CFD topology optimization unit uses the uniformity of current density and the consistency of heat dissipation flow field distribution of the backplane wiring as constraints, and generates the initial backplane wiring and TFT cell layout topology through computational fluid dynamics topology optimization.

[0022] The mathematical expression for the current density uniformity constraint is: in, Let N be the standard deviation of the global current density of the backplane traces, and N be the total number of trace grids. Let be the current density of the i-th trace grid. This represents the average current density over the entire trace area. The preset current density uniformity threshold; The mathematical expression for the consistency constraint of the heat dissipation flow field distribution is: in, Let M be the standard deviation of the global temperature field of the backplate, and M be the total number of grid cells in the temperature field of the backplate. Let j be the steady-state temperature of the j-th temperature field grid. The average steady-state temperature over the entire backplate area. This is a preset threshold for temperature distribution consistency.

[0023] Based on the above constraints, the CFD topology optimization element uses a variable density method to discretize the backplane design domain, dividing it into several finite element meshes of uniform size. The density variable of each mesh... ,in The corresponding mesh represents the solid trace area. The corresponding grid is a non-trace blank area; with trace volume fraction as constraint and current density uniformity and heat dissipation flow field distribution as optimization objectives, a topology optimization mathematical model is constructed. After iterative solution, an initial backplane trace and TFT cell layout topology is generated. The initial layout topology includes the initial coordinate set of TFT cells, the initial topology path of the backplane trace, the initial trace width and stack-up structure.

[0024] Furthermore, the graph model construction and iterative optimization unit takes each TFT unit and sub-pixel driving node as graph nodes, and uses the electrical coupling, optical crosstalk, and thermal conduction correlation between nodes as edge weights to construct a backplane layout correlation graph model, and iteratively optimizes the node positions and connection relationships of the initial layout topology.

[0025] The backplate arrangement relationship graph model is defined as an undirected weighted graph. ,in: V represents the set of graph nodes, where the total number of nodes equals the total number of TFT units in the target QD-OLED panel. Each node... Corresponding to a TFT unit and its driven sub-pixel driving node, the node attribute vector ,in , For nodes The planar coordinates of the node, and the remaining parameters are the feature parameters of the TFT and sub-pixel corresponding to the node; Let E be the set of edges, for any two nodes , If and only if the Euclidean distance between the two is... At that time, there exists an edge Connecting nodes and , The preset neighborhood distance threshold; W is the adjacency matrix, and the elements of the adjacency matrix are... For the edge The edge weights, which are determined by the nodes and Electrical coupling between Optical crosstalk Thermal conductivity correlation The weighted fusion generation, the fusion formula is: in, , , The preset weighting coefficients, and ; The electrical coupling The coupling strength coefficient for the parasitic capacitance between nodes is calculated using the following formula: in, For nodes and The traces and parasitic capacitances of the components. , , For nodes The parasitic capacitance corresponding to TFT devices; The optical crosstalk For nodes Light emission pairs of corresponding sub-pixels The formula for calculating the optical crosstalk contribution of the corresponding sub-pixel is: in, For nodes Photons emitted by the corresponding sub-pixel are incident on the node The luminous flux of the corresponding sub-pixel quantum dot emitting layer For nodes The rated emitted light flux corresponding to the sub-pixel itself; The thermal conductivity correlation For nodes Joule heat loss at nodes The contribution of temperature increase to the overall temperature is calculated using the following formula: in, For nodes The rated power consumption heats up the node The steady-state temperature change For nodes The change in steady-state temperature caused by the heat generated by rated power consumption.

[0026] Based on the aforementioned backplane layout association graph model, the graph model construction and iterative optimization unit iteratively optimizes the initial layout topology. Each iteration updates the coordinates of nodes and edge connections, synchronously updating the element values ​​of the adjacency matrix W. The iteration terminates when the change in the Frobenius norm of the adjacency matrix is ​​determined. Or the number of iterations reaches the preset maximum number of iterations. ,in This is the preset convergence threshold.

[0027] Subsequently, the multiphysics simulation and data acquisition unit collects the luminous efficiency and color coordinate offset data of sub-pixels under different arrangement positions, and feeds the collected data as a constraint term into the loss function and topology optimization boundary conditions of the graph convolutional neural network.

[0028] The multiphysics simulation and data acquisition unit constructs an electrical-thermal-optical joint simulation model based on the current iteration's topology. Under standard driving voltage and rated refresh frequency conditions, it simulates and acquires the actual luminous efficiency of each sub-pixel. Compared with actual color coordinates Calculate the color coordinate offset ,in , These are the standard color coordinates of the sub-pixel.

[0029] The collected luminous efficiency With color coordinate offset As constraints, these are fed back into the loss function of the graph convolutional neural network, adding a luminous efficiency constraint and a color coordinate shift penalty term to the loss function; simultaneously, luminous efficiency below a preset efficiency threshold is considered a constraint. The color coordinate offset is greater than the preset offset threshold. The region where the node is located is set as a no-routing boundary condition for CFD topology optimization, which constrains the design domain range of subsequent topology optimization.

[0030] Finally, after iterative convergence, a backplane layout scheme adapted to the target QD-OLED panel is generated, and the timing control parameter generation unit outputs the driving timing control parameters corresponding to the layout scheme to the display driving control unit.

[0031] The backplane layout scheme includes the final coordinate position set of the TFT cells, the final topology of the backplane traces, the linewidth and stack-up structure of each trace segment, and the via position and size parameters. The driving timing control parameters include the sub-pixel charging voltage amplitude, gate scan timing, and source data writing timing parameters. The timing control parameter generation unit encapsulates the above parameters into a timing configuration file that can be recognized by the display driving control unit and outputs it to the timing controller of the display driving control unit. The timing controller then controls the source driving chip and the gate driving chip to perform corresponding driving operations.

[0032] As a preferred embodiment, refer to the appendix. Figure 2 First, the CFD topology optimization unit maps the backplane routing network to a virtual fluid micropipe network, equates the current density distribution to the fluid pressure field inside the pipe, and equates the heat dissipation distribution to the fluid velocity field inside the pipe.

[0033] The design domain of the backplane is discretized into several microtube elements, each corresponding to a grid node of the backplane trace. The on / off state of the microtube element corresponds to whether the grid node is a trace entity region. The current transmission within the trace is equivalent to the flow of virtual fluid within the microtube, and the distribution of the current density J is mapped to the fluid pressure field within the microtube. The current continuity equation is equivalent to the fluid mass conservation equation; the heat dissipation and conduction of Joule heat in the wiring are equivalent to the kinetic energy transfer of a virtual fluid, and the distribution of the backplate temperature field T is mapped to the fluid velocity field inside the microtube. The heat conduction equation is equivalent to the fluid momentum conservation equation.

[0034] Secondly, based on the microgrid power flow scheduling strategy, virtual fluid particles are injected into the virtual fluid micropipe network to construct a pipeline resistance evaluation function with the uniformity of current density and the consistency of heat dissipation flow field distribution.

[0035] The microgrid power flow scheduling strategy aims to balance the node pressure of the micropipe network. It sets virtual fluid injection and outflow rates for each micropipe node, where the node corresponding to the TFT unit is the fluid outflow node, and the node corresponding to the power input is the fluid injection node. The expression for the pipe resistance evaluation function is: in, The total resistance of the pipeline, , The preset weighting coefficients, The standard deviation of the current density distribution. The standard deviation of the temperature field distribution is represented by the value of the pipeline resistance evaluation function, which is negatively correlated with the uniformity of current density and the consistency of heat dissipation flow field distribution.

[0036] Subsequently, the particle swarm optimization strategy in multi-agent reinforcement learning is used to guide the virtual fluid particles to travel between pipeline nodes. After convergence, the initial backplane routing and TFT cell layout topology are generated.

[0037] Each virtual fluid particle is considered an independent intelligent agent. The agent's action space is the set of neighboring nodes of the current node in the micro-pipe network. The agent's state space consists of the pressure value, velocity value, and neighborhood pipe resistance value of the current node. The agent's reward function is the negative value of the pipe resistance evaluation function. All agents share the global optimal path information. Through iterative updates of individual extreme values ​​and global extreme values, virtual fluid particles are guided to walk towards the path with the least total pipeline resistance. After iterative convergence, the walking path of the virtual fluid particles corresponds to the topology of the backplane wiring, and the position of the fluid outflow node corresponds to the arrangement position of the TFT unit, thus generating the initial backplane wiring and TFT unit arrangement topology.

[0038] As a further preferred embodiment, in the process of guiding the virtual fluid particle path using the particle swarm cooperative optimization strategy in multi-agent reinforcement learning, a genetic algorithm is also introduced for global optimization, specifically including: The path node sequence in the virtual fluid microtube network is encoded as a chromosome gene chain in the genetic algorithm. Each gene position corresponds to the on / off state of a microtube node. A gene position value of 1 corresponds to a path node, and a gene position value of 0 corresponds to a non-path node. The reciprocal of the pipeline resistance evaluation function is used as the chromosome fitness, i.e. The fitness value is positively correlated with the path performance of the corresponding chromosome; a roulette wheel selection mechanism is used to select high-fitness particle paths, and the higher the fitness of the chromosome, the greater the probability of it being selected. Multi-point crossover and Gaussian mutation operations are performed on the chromosome gene chains corresponding to the selected particle paths. The multi-point crossover operation randomly selects multiple gene positions of two chromosomes for crossover and exchange. The Gaussian mutation operation superimposes random perturbations that conform to a Gaussian distribution on the randomly selected gene positions to generate a new generation of particle walking paths. Repeat the selection, crossover, and mutation operations until the change in chromosome fitness is less than a preset threshold, converge to the globally optimal initial arrangement topology, and break the local optimum unbinding.

[0039] As a preferred embodiment, refer to the appendix. Figure 3 and attached Figure 4 First, the electrical coupling, optical crosstalk, and thermal conduction correlation between nodes are extracted to construct an initial adjacency matrix that integrates spatial proximity and physical field strength.

[0040] The initial adjacency matrix elements The calculation formula is: in, For nodes and The Euclidean distance is used, and the other parameters are defined in the same way as in the above embodiment. The initial adjacency matrix simultaneously integrates the spatial proximity of nodes and the multiphysics coupling strength.

[0041] Secondly, a dynamic congestion propagation model in the traffic network is introduced, which treats the superimposed interference of multi-physics fields as the congestion load in the graph network, and calculates the congestion state vector of each graph node at different time steps.

[0042] The electrical coupling interference, optical crosstalk interference, and thermal conduction interference of a node within each time step are superimposed to form the congestion load of the node. The state update formula of the dynamic congestion propagation model is as follows: in, For nodes The congestion state vector at time step t , , , They are nodes The cumulative electrical coupling, cumulative optical crosstalk, and cumulative thermal conduction at time step t; The pre-defined congestion dissipation coefficient characterizes the node's ability to dissipate interference.

[0043] Subsequently, a spatiotemporal graph convolutional network is used to extract the spatiotemporal features of the congestion state vector. By alternating the mapping between graph convolutional layers and temporal convolutional layers, the edge weights in the adjacency matrix are dynamically updated to complete the iterative optimization of node positions and connection relationships.

[0044] The spatiotemporal graph convolutional network comprises multiple cascaded spatiotemporal convolutional blocks. Each spatiotemporal convolutional block consists of two gated temporal convolutional layers and one spatial graph convolutional layer, with the spatial graph convolutional layer embedded between the two gated temporal convolutional layers. The spatial graph convolutional layer extracts spatial dimension features from the congestion state vector based on the current adjacency matrix, extracting spatial association features between graph nodes. The gated temporal convolutional layer extracts temporal dimension features from the spatial features, extracting the temporal evolution features of the congestion state. The output of the spatiotemporal graph convolutional network is the updated adjacency matrix. Based on the updated adjacency matrix, the positions of nodes and edge connections are adjusted to complete one iteration of optimization. The above process is repeated until the model converges.

[0045] As a further preferred embodiment, in the process of extracting the spatiotemporal features of the congestion state vector using a spatiotemporal graph convolutional network, a multi-head self-attention mechanism from natural language processing is embedded in the feature mapping stage of the spatiotemporal graph convolutional network, specifically including: A multi-head self-attention mechanism is embedded in the spatial graph convolutional layer of the spatiotemporal graph convolutional network. The electrical coupling, optical crosstalk and thermal conduction correlation are projected to different attention heads respectively, and the self-attention weight matrix of each physical field feature between graph nodes is calculated independently. For the k-th attention head, its self-attention weight matrix The calculation formula is: in, , X is the input node feature matrix. , Let be the trainable parameters of the query matrix and key matrix of the k-th attention head. Let k be the feature dimension of the attention head; The multi-head self-attention weight matrices are concatenated and linearly transformed to generate an adaptive feature vector containing multi-physics-dominated interference factors. The fusion formula is as follows: Where h is the total number of attention heads, Let A be the trainable parameter matrix of the linear transformation, and let A be the fused adaptive attention weight matrix, which replaces the adjacency matrix with fixed weights in the spatiotemporal feature extraction.

[0046] As a further preferred embodiment, in the process of independently calculating the self-attention weight matrix of each physical field feature among the graph nodes, a proportional-integral-derivative (PID) controller is also introduced to achieve closed-loop self-tuning of the weights, specifically including: Obtain the difference between the real-time color coordinate offset error and the set target error in the currently displayed image. ,in This represents the average color coordinate offset of all sub-pixels across the panel at the current moment. The preset target color coordinate offset threshold; Difference As the input signal to the PID controller, the dynamic adjustment parameters are generated through a linear combination of proportional, integral, and derivative functions. The calculation formula is: in, This is the proportionality coefficient. The integral coefficient is... These are the differential coefficients; Using dynamic adjustment parameters In online scaling of the key-value matrix parameters in a multi-head self-attention mechanism, the scaled key matrix is: The value matrix is This enables closed-loop self-tuning of graph network edge weights as they change with the displayed content, eliminating the lag in the adaptation of static attention weights to dynamic content.

[0047] As a preferred embodiment, firstly, the luminous efficiency and color coordinate offset data of the sub-pixels at the current arrangement position are mapped to the environment state space in the reinforcement learning framework.

[0048] The reinforcement learning framework is a deep reinforcement learning framework, the agent is a parameter update unit of a graph convolutional neural network, and the environment is a multi-physics coupled system with a backplane arrangement; the state vector of the environment state space S. ,in This represents the average luminous efficiency of all sub-pixels across the panel at the current moment. This represents the average color coordinate offset of all sub-pixels across the panel at the current moment. The standard deviation of the current density of the backplane traces at the current moment. The standard deviation of the temperature field across the entire backplane at the current moment; the action space A of the agent is the adjustment amount of the weight parameters of the graph convolutional neural network, and the adjustment parameters of the topology optimization boundary conditions.

[0049] Secondly, a multi-objective reward function is constructed, which includes a reward term for maximizing luminous efficiency, a penalty term for minimizing color coordinate offset, and a term for maintaining backplane trace current carrying capacity margin.

[0050] The expression for the multi-objective reward function is: in, The reward item is for maximizing luminous efficiency. ,when hour, Take a positive value, otherwise take 0. This is the preset reward coefficient; The penalty term is for minimizing the color coordinate offset. ,when hour, Take a positive value, otherwise take 0. This is the preset penalty coefficient; This is a term used to maintain the current carrying capacity margin for backplane traces. ,when hour, Take a positive value, otherwise take a negative value. This represents the average current carrying capacity margin of the backplane traces at the current moment. The preset minimum load margin threshold, These are the preset margin weighting coefficients.

[0051] Subsequently, the gradient of the multi-objective reward function with respect to the weights of the graph convolutional neural network is calculated using the policy gradient algorithm. The weight coefficients of each constraint in the loss function are adaptively adjusted to complete the backpropagation update of the network parameters.

[0052] The policy gradient algorithm uses the REINFORCE algorithm, and the formula for calculating the policy gradient is: in, These are the weight parameters of a graph convolutional neural network. For the expected cumulative reward of the strategy, For the policy network in state Select action The probability distribution is calculated; based on the calculated policy gradient, the weight parameters of the graph convolutional neural network are updated using gradient descent. Simultaneously, the weight coefficients of the luminous efficiency constraint term, color coordinate offset penalty term, and current density constraint term in the loss function are adaptively adjusted to ensure that the optimization direction of the loss function is consistent with the maximization direction of the multi-objective reward function, thereby completing the backpropagation update of the network parameters.

[0053] As a further preferred embodiment, in the process of constructing the multi-objective reward function, a non-cooperative game model is introduced to optimize the balance of the multi-objectives, specifically including: Maximizing luminous efficiency, minimizing color coordinate offset, and maintaining current carrying capacity margin are defined as three competing game participants. The strategy space of each game participant is the weight adjustment range of the corresponding optimization objective, and the payoff function is the marginal payoff of the corresponding objective. Based on the non-cooperative game theory model in economics, the marginal payoff function of each player under the current topology is calculated, where the marginal payoff function of the player maximizing luminous efficiency is: The marginal benefit function for participants that minimizes the color coordinate offset is: The current carrying capacity margin maintains the marginal revenue function of the participants as follows: ,in , , These are the weighting coefficients of the three objectives in the reward function; The goal is to find the Nash equilibrium point of the marginal revenue function. At Nash equilibrium, no single participant can increase their marginal revenue by adjusting their own weight coefficients. This equilibrium satisfies the following condition: The Nash equilibrium constraint is embedded as a regularization term in the multi-objective reward function. The expression for the regularization term is: in, The preset regularization coefficient; Add a multi-objective reward function In this process, the predatory suppression of the performance improvement of a single physical field on the performance of other physical fields is suppressed, thereby achieving balanced optimization of multiple objectives.

[0054] As a preferred embodiment, refer to the appendix. Figure 5 and attached Figure 6 First, the timing control parameter generation unit identifies the current carrying capacity distribution map of the wiring in each area of ​​the backplane layout scheme.

[0055] The entire backplane area is divided into several uniformly sized grid regions, and the current carrying capacity margin of the traces within each grid region is calculated. ,in This represents the maximum current carrying capacity of the wiring in this area. The rated operating current of the traces in this area is given; based on the current carrying margin of all area grids, a current carrying margin distribution map is generated. The map is a two-dimensional matrix that corresponds one-to-one with the backplane area, and the matrix elements are the current carrying margin values ​​of the corresponding areas.

[0056] Secondly, by combining the time-division multiple access resource scheduling mechanism in the communication network, the charging and discharging operations of sub-pixels are encapsulated into independent data frames.

[0057] Based on the time division multiple access mechanism, a refresh cycle is divided into several time slots of equal length, and each time slot corresponds to a data frame transmission and execution window; the charging and discharging operations of each sub-pixel are encapsulated into independent charging data frames and discharging data frames, and each data frame contains the address information, driving voltage information and operation duration information of the corresponding sub-pixel.

[0058] Subsequently, based on the current carrying margin distribution map, non-overlapping write time slots and staggered discharge timing are allocated to the sub-pixel data frames of different regions. The generated time slot allocation table is output as the driving timing control parameter to smooth the transient current peaks in the backplane routing network.

[0059] For low-margin regions where the current carrying capacity is lower than the preset margin threshold, non-overlapping write time slots are allocated to the sub-pixel data frames in this region, and the charging and writing operations of sub-pixels in the same row are completely staggered in time to avoid transient current spikes caused by multiple sub-pixels charging at the same time. For high-margin regions where the current carrying capacity is higher than the preset margin threshold, continuous write time slots are allocated to the sub-pixel data frames in this region to improve data writing efficiency. At the same time, the discharge timing of sub-pixels in different regions is staggered to make the time-domain distribution of discharge current smoother and avoid transient current peaks caused by concentrated discharge.

[0060] The generated time slot allocation table, charging voltage amplitude, and gate scan timing parameters are encapsulated into drive timing control parameters and output to the timing controller of the display drive control unit. The timing controller then controls the source drive chip and the gate drive chip to perform sub-pixel charging and discharging operations according to the time slot allocation table.

[0061] As a further preferred embodiment, in the process of allocating write time slots for sub-pixel data frames based on the current carrying capacity margin distribution map, a queuing theory model is introduced to optimize time slot allocation, specifically including: The write request of sub-pixel data frame is regarded as the customer arrival flow in the queuing theory model of operations research, and the current carrying capacity of the backplane wiring is regarded as the service desk processing rate. Based on the M / G / 1 queuing model, the variance of the arrival time interval of sub-pixel data frames in each region and the variance of the service processing time are statistically analyzed. Here, M indicates that the arrival process of the data frame is a Poisson process, G indicates that the service processing time is an arbitrary independent and identically distributed random process, and 1 represents a single server. The optimal queuing threshold for each region is calculated using the steady-state distribution analytical expression of the M / G / 1 queuing model, which is the Pollaczek-Khintchine (PK) formula. The formula for calculating the average queuing time is as follows: in, The average arrival rate of data frames. For the mathematical expectation of service processing time, The variance of service processing time; The optimal queuing threshold for each region is calculated based on the average queuing waiting time. The write time slots of low-priority sub-pixel data frames are dynamically truncated or delayed according to the optimal queuing waiting threshold. The low-priority sub-pixels are sub-pixels in the current display screen whose grayscale value is lower than the preset grayscale threshold. This avoids queuing blockage in high-load areas and achieves dynamic optimal allocation of write time slots.

[0062] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for backplane layout management based on QD-OLED panels, characterized in that, include: Obtain the physical parameters of the backplane substrate, the electrical parameters of the TFT devices, and the spectral and luminous decay characteristics of the quantum dot sub-pixels of the target QD-OLED panel; With the uniformity of current density and the consistency of heat dissipation flow field distribution of the backplane traces as constraints, the initial backplane trace and TFT cell layout topology is generated by computational fluid dynamics topology optimization. Each TFT unit and sub-pixel driving node are treated as graph nodes. The electrical coupling, optical crosstalk, and thermal conduction correlation between nodes are used as edge weights to construct a backplane layout correlation graph model. The node positions and connection relationships of the initial layout topology are iteratively optimized. Data on the luminous efficiency and color coordinate offset of sub-pixels under different arrangement positions are collected, and the collected data is fed back as a constraint term into the loss function and boundary conditions of the graph convolutional neural network for topology optimization. Generate a backplane layout scheme adapted to the target QD-OLED panel, and output the corresponding driving timing control parameters of the layout scheme to the display driving control unit.

2. The method for backplane layout management based on a QD-OLED panel according to claim 1, characterized in that, The initial backplane routing and TFT cell layout topology is generated through computational fluid dynamics topology optimization, including: The backplane routing network is mapped as a virtual fluid micropipe network, the current density distribution is equivalent to the fluid pressure field inside the pipe, and the heat dissipation distribution is equivalent to the fluid velocity field inside the pipe. Based on the microgrid power flow scheduling strategy, virtual fluid particles are injected into the virtual fluid micropipe network, and a pipe resistance evaluation function is constructed with the uniformity of current density and the consistency of heat dissipation flow field distribution. By utilizing the particle swarm optimization strategy in multi-agent reinforcement learning, virtual fluid particles are guided to navigate between pipeline nodes. After convergence, the initial backplane routing and TFT cell layout topology are generated.

3. The method for backplane layout management based on a QD-OLED panel according to claim 1, characterized in that, Construct a backplane layout association diagram model and perform iterative optimization, including: The electrical coupling, optical crosstalk, and thermal conduction correlation between nodes are extracted to construct an initial adjacency matrix that integrates spatial proximity and physical field strength. A dynamic congestion propagation model is introduced in the traffic network, and the superimposed interference of multi-physics fields is regarded as the congestion load in the graph network. The congestion state vector of each graph node at different time steps is calculated. Spatiotemporal graph convolutional networks are used to extract the spatiotemporal features of congestion state vectors. By alternating mapping between graph convolutional layers and temporal convolutional layers, the edge weights in the adjacency matrix are dynamically updated to complete the iterative optimization of node positions and connectivity.

4. The method for backplane layout management based on a QD-OLED panel according to claim 1, characterized in that, The collected data is fed back as a constraint term into the loss function of the graph convolutional neural network, including: Map the luminous efficiency and color coordinate offset data of the sub-pixels at the current arrangement position to the environment state space in the reinforcement learning framework; Construct a multi-objective reward function that includes a reward term for maximizing luminous efficiency, a penalty term for minimizing color coordinate offset, and a term for maintaining backplane trace current carrying capacity margin. The gradient of the multi-objective reward function with respect to the weights of the graph convolutional neural network is calculated using the policy gradient algorithm. The weight coefficients of each constraint in the loss function are adaptively adjusted to complete the backpropagation update of the network parameters.

5. The method for backplane layout management based on a QD-OLED panel according to claim 1, characterized in that, The drive timing control parameters corresponding to the layout scheme are output to the display drive control unit, including: Identify the current carrying capacity distribution of traces in each area of ​​the backplane layout scheme. By combining the time division multiple access resource scheduling mechanism in the communication network, the charging and discharging operations of sub-pixels are encapsulated into independent data frames; Based on the current carrying margin distribution map, non-overlapping write time slots and staggered discharge timing are allocated to the sub-pixel data frames of different regions. The generated time slot allocation table is output as the driving timing control parameter to smooth the transient current peaks in the backplane routing network.

6. The method for backplane layout management based on a QD-OLED panel according to claim 2, characterized in that, The particle swarm optimization strategy in multi-agent reinforcement learning is used to guide the path movement of virtual fluid particles, including: Encode the path node sequence in the virtual fluid microtube network into a chromosome gene chain in a genetic algorithm; The reciprocal of the pipeline resistance evaluation function is used as the chromosome fitness, and a roulette wheel selection mechanism is used to select high-fitness particle paths. Multi-point crossover and Gaussian mutation operations are performed on the chromosome gene chains corresponding to the selected particle paths to generate a new generation of particle walking paths.

7. The method for backplane layout management based on a QD-OLED panel according to claim 3, characterized in that, Spatiotemporal features of congestion state vectors are extracted using spatiotemporal graph convolutional networks, including: Embed a multi-head self-attention mechanism from natural language processing in the feature mapping stage of a spatiotemporal graph convolutional network; The electrical coupling, optical crosstalk, and thermal conduction correlation are projected onto different attention heads, and the self-attention weight matrix of each physical field feature between graph nodes is calculated independently. The multi-head self-attention weight matrix is ​​concatenated and linearly transformed to generate an adaptive feature vector containing multi-physics-dominated interference factors, which replaces the fixed-weight adjacency matrix in spatiotemporal feature extraction.

8. The method for backplane layout management based on a QD-OLED panel according to claim 4, characterized in that, Construct a multi-objective reward function, including: Maximizing luminous efficiency, minimizing color coordinate offset, and maintaining current carrying capacity are defined as three competing game participants; Based on the non-cooperative game model in economics, calculate the marginal revenue function of each game participant under the current topology. With the goal of finding the Nash equilibrium point of the marginal revenue function, the Nash equilibrium constraint is embedded as a regularization term into the multi-objective reward function.

9. The method for backplane layout management based on a QD-OLED panel according to claim 5, characterized in that, Write time slots are allocated to sub-pixel data frames based on the current carrying capacity margin distribution map, including: The write request of sub-pixel data frame is regarded as the customer arrival flow in the queuing theory model of operations research, and the current carrying capacity of the backplane wiring is regarded as the service desk processing rate. Based on the M / G / 1 queuing model, the variance of the arrival time interval of sub-pixel data frames in each region and the variance of the service processing time are statistically analyzed. The optimal queuing threshold for each region is calculated using the steady-state distribution analytical formula of the M / G / 1 queuing model. Based on the optimal queuing threshold, the write time slots of low-priority sub-pixel data frames are dynamically truncated or delayed.

10. The method for backplane layout management based on a QD-OLED panel according to claim 7, characterized in that, Calculate the self-attention weight matrix of each physical field feature among the graph nodes independently, including: Obtain the difference between the real-time color coordinate offset error and the set target error in the current display screen; The difference is used as the input signal of the proportional-integral-derivative controller in classical control theory, and dynamic adjustment parameters are generated through a linear combination of proportional, integral, and derivative functions. The key-value matrix parameters in the multi-head self-attention mechanism are scaled online using dynamically adjusted parameters.

Images