Solid-state battery production system and method based on artificial intelligence

Abstract

The invention belongs to the technical field of solid-state battery production, and provides a solid-state battery production system and method based on artificial intelligence. The method comprises the following steps: acquiring physical contact state data and interface chemical state data of the solid-state battery in production in real time, and predicting whether an electrolyte layer meets a preset condition of incomplete lamination according to the real-time data; if so, processing the physical contact state data, the interface chemical state data and the historical process parameters by using a depth deterministic strategy gradient algorithm to obtain a high-voltage compensation instruction and a plasma processing extension instruction; and synchronously adjusting a hot pressing process and an interface treatment process based on the high-voltage compensation instruction and the plasma treatment extension instruction. According to the method, intervention can be completed before defects occur, and the problems of physical attachment (mechanical parameters) and chemical stability (interface modification) are synchronously solved through linkage adjustment of high-voltage compensation and plasma treatment extension, so that the interface impedance is remarkably reduced, and the yield is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of solid-state battery production, and in particular to an artificial intelligence-based solid-state battery production system and method thereof. Background Art

[0002] Solid-state batteries are considered the core development direction of next-generation energy storage technology due to their high energy density and safety. However, their manufacturing process, especially the hot pressing and interface treatment, still faces the following key issues: (1) Hot pressing process defects: The solid electrolyte layer (such as LLZO, LGPS) and the electrode material are prone to pores, delamination or cracks due to uneven temperature and pressure distribution during the hot pressing process, which significantly increases the interface impedance and reduces battery performance. (2) Interface chemical instability: Insulating layers such as Li2CO3 and LiOH are easily generated at the electrode-electrolyte interface, hindering lithium ion transmission. Traditional plasma treatment process parameters are fixed and cannot dynamically adapt to changes in the material state.

[0003] To address the above key issues, the current mainstream industry approach uses PID control or rule-based strategies to regulate hot pressing and interface treatment processes, but these methods have significant shortcomings: (1) Single-parameter closed-loop control: Adjustments are made based solely on real-time deviations in temperature or pressure, making it impossible to synergistically optimize the coupling effects of hot pressing (mechanical densification) and interface treatment (chemical modification). For example, simply increasing the pressure may cause electrolyte fragmentation while ignoring the need for simultaneous suppression of interface side reactions. (2) Strong reliance on experience: Process parameters rely on trial-and-error settings by engineers, making it difficult to adapt to differences in material batches or complex working conditions (such as multi-layer stacked batteries). (3) Response hysteresis: Traditional control only adjusts after defects (such as non-bonding) occur, lacking predictive intervention capabilities.

[0004] In recent years, some studies have attempted to apply machine learning to battery manufacturing, but there are still limitations: (1) Supervised learning: It relies on a large amount of labeled data to train static models and cannot adapt to dynamic changes in the production line. (2) Traditional reinforcement learning (such as Q-Learning): It is only applicable to discrete action spaces (such as on-off control) and cannot meet the continuous high-precision adjustment requirements of the hot pressing process (such as ±0.5°C and 10Hz pressure control).

[0005] In summary, how to utilize the synergistic effect generated by adjusting the hot pressing process and the interface treatment process to significantly improve the interface performance and yield of solid-state batteries is a technical problem that needs to be solved at present. Summary of the Invention

[0006] In this regard, the present invention provides an artificial intelligence-based solid-state battery production method, system, electronic device, computer storage medium and computer program product to solve at least one of the above-mentioned technical problems.

[0007] In a first aspect, the present invention provides a solid-state battery production method based on artificial intelligence, comprising the following steps:

[0008] Real-time collection of physical contact state data and interface chemical state data of solid-state batteries in production, the physical contact state data including pressure sensor data, electrolyte layer deformation monitoring data, and surface contact angle measurement data; the interface chemical state data including plasma emission spectroscopy data;

[0009] Predicting whether the electrolyte layer meets a preset condition of incomplete bonding based on the real-time data, and if so, using a deep deterministic policy gradient algorithm to process the physical contact state data, the interface chemical state data, and historical process parameters to obtain a high-voltage compensation instruction and a plasma treatment extension instruction;

[0010] A hot pressing process and an interface treatment process are synchronously adjusted based on the high voltage compensation instruction and the plasma treatment extension instruction.

[0011] In a second aspect, the present invention provides an artificial intelligence-based solid-state battery production system, the system comprising an acquisition unit, an instruction generation unit, and an adjustment unit;

[0012] The acquisition unit is used to collect physical contact state data and interface chemical state data of the solid-state battery in production in real time, wherein the physical contact state data includes pressure sensor data, electrolyte layer deformation monitoring data, and surface contact angle measurement data; the interface chemical state data includes plasma emission spectrum data;

[0013] The instruction generation unit is configured to predict, based on the real-time data, whether the electrolyte layer meets a preset condition of incomplete bonding, and if so, use a deep deterministic policy gradient algorithm to process the physical contact state data, the interface chemical state data, and historical process parameters to obtain a high-voltage compensation instruction and a plasma treatment extension instruction;

[0014] The adjustment unit is used to synchronously adjust the hot pressing process and the interface treatment process based on the high-voltage compensation instruction and the plasma treatment extension instruction.

[0015] In a third aspect of the present invention, an electronic device is provided, comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements any of the methods described above when executed by the processor.

[0016] According to a fourth aspect of the present invention, a computer storage medium is provided, wherein the computer storage medium stores a computer program executable by a processor to implement any of the methods described above.

[0017] According to a fifth aspect of the present invention, a computer program product is provided, which comprises a computer program executable by a processor to implement any of the methods described above.

[0018] The beneficial technical effects of the technical solution of the present invention are:

[0019] On the one hand, compared to the delayed response of traditional PID control, the solution of the present invention can complete intervention before defects occur. On the other hand, through the coordinated adjustment of high-voltage compensation and extended plasma treatment, the problems of physical adhesion (mechanical parameters) and chemical stability (interface modification) are simultaneously solved. For example, during the hot pressing stage, pressure compensation is used to eliminate electrolyte layer pores, while extending the plasma treatment time to reduce interfacial chemical side reactions caused by pressure changes, ultimately significantly reducing interfacial impedance and improving yield. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 This is a schematic flow chart of a solid-state battery production method based on artificial intelligence disclosed in an embodiment of the present invention;

[0022] Figure 2 is a schematic diagram of using an external model to improve the efficiency of obtaining optimal pressure correction values ​​and duration correction values ​​disclosed in an embodiment of the present invention;

[0023] Figure 3 This is a structural diagram of an artificial intelligence-based solid-state battery production system disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] The following specific embodiments illustrate the implementation of this application. Those familiar with the art can easily understand the other advantages and functions of this application from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of this application, but not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0025] In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0026] like Figure 1As shown, an embodiment of the present invention discloses a solid-state battery production method based on artificial intelligence, including the following method steps:

[0027] S10, real-time collection of physical contact state data and interface chemical state data of solid-state batteries in production, wherein the physical contact state data includes pressure sensor data, electrolyte layer deformation monitoring data, and surface contact angle measurement data; and the interface chemical state data includes plasma emission spectrum data.

[0028] During the hot pressing and interface treatment processes in solid-state battery production, real-time physical contact state data and interface chemical state data are collected. The details are as follows:

[0029] 1. Physical contact status data

[0030] Pressure distribution gradient: Thin-film pressure sensors (e.g., FPC pressure sensors) are embedded in the pressing mold to monitor the pressure distribution gradient in various areas within the hot pressing equipment in real time. This helps identify pressure anomalies (e.g., local overpressure or underpressure) caused by equipment wear and uneven material stacking, thus preventing cracks or delamination in the electrolyte layer due to uneven pressure.

[0031] Deformation rate: The degree of deformation of the electrolyte layer per unit time (e.g., thickness compression rate, planar extension rate). For example, by scanning the electrode surface with a laser, the thickness change rate can be monitored in real time.

[0032] Surface contact angle: A contact angle meter is used to detect the wettability of the electrode-electrolyte interface. The size of the contact angle directly reflects the degree of interfacial adhesion. A contact angle that is too large (for example, >90°) indicates the presence of air gaps or chemical repulsion at the interface, which may increase the interfacial impedance.

[0033] Interface temperature distribution: The temperature field distribution at the interface during pressing (especially hot pressing) affects material diffusion and interfacial reaction. Micro-thermocouples can be embedded in the press head or mold to measure temperature at multiple points simultaneously, and the interface temperature can be inferred by combining heat conduction models.

[0034] 2. Interface chemical state data

[0035] Plasma emission spectrum data: Spectrometers are used to analyze the characteristic spectral lines released during plasma treatment (e.g., spectral signals of Li, O, and C elements) in real time to determine whether insulating layers such as Li2CO3 and LiOH are generated at the interface, as well as the effect of plasma treatment on interface chemical modification (e.g., the extent of surface redox reactions). These data mainly include Li2CO3 peak intensity ratio, plasma electron temperature, plasma electron density, and interface redox potential, as follows:

[0036] Li2CO3 peak intensity ratio: The ratio of the spectral signal intensity of the insulating layer and the matrix Li element; during the plasma treatment process, the spectrometer collects the optical signal generated by the plasma discharge in real time and screens the emission peaks of the Li element (670.8nm characteristic spectrum line) and the C element (247.9nm characteristic spectrum line). Calculate the peak intensity ratio I li / I C The lower the ratio, the thicker the Li2CO3 insulating layer.

[0037] Plasma electron temperature: the average kinetic energy of electrons in plasma. Through microwave diagnosis or Langmuir probe, the interaction between electrons in plasma and electromagnetic waves / probe is used to infer the average kinetic energy (temperature) of electrons.

[0038] Plasma electron density: the number of plasma electrons per unit volume (unit: cm -3 ) Based on microwave interferometry or Langmuir probe, the number of electrons per unit volume is calculated through the phase shift of electromagnetic waves or the probe current.

[0039] Interfacial redox potential: the electrochemical potential at the interface (unit: V, can be measured by Kelvin probe or electrochemical workstation).

[0040] S20, predicting whether the electrolyte layer will meet the preset condition of incomplete bonding based on the real-time data, and if so, using a deep deterministic policy gradient algorithm to process the physical contact state data, the interface chemical state data and historical process parameters to obtain high-voltage compensation instructions and plasma treatment extension instructions.

[0041] Based on the physical contact state data and interface chemical state data obtained above, it can be predicted whether the solid-state battery will be at risk of incomplete bonding according to the current pressing process and interface treatment process. For example, if the predicted pressure change rate deviates from the historical value by more than 10% (pressure anomaly), the real-time compression amount differs from the theoretical model by ≥5% (deformation anomaly), or the contact angle increases by ≥10% (wettability deterioration), it is determined that there is a risk of incomplete bonding.

[0042] When the risk of incomplete alignment is predicted, a deep deterministic policy gradient algorithm is used to predict high-voltage compensation instructions and plasma treatment extension instructions:

[0043] The above-mentioned physical contact state data, interface chemical state data and historical process parameters (average temperature, average pressure, average deformation rate, and dwell time integral over the past N seconds, for example, N = 10s) are integrated to construct a state space.

[0044] By learning dynamic mapping relationships through neural networks, optimization instructions for continuous action space (such as pressure compensation value and plasma treatment extension time) are output, solving the problem that traditional reinforcement learning (such as Q-Learning) cannot handle continuous parameter adjustment.

[0045] The coupling effects of hot pressing densification (mechanical parameters) and interfacial chemical modification (plasma treatment) are considered simultaneously to avoid the limitations of single parameter adjustment (for example, simply increasing the pressure may cause electrolyte fragmentation, and the plasma treatment needs to be extended simultaneously to suppress interfacial side reactions).

[0046] Generate instructions: (1) High-voltage compensation instruction: Based on the output of the DDPG algorithm, dynamically adjust the hot pressing pressure (for example, when delamination risk is predicted, superimpose a 5-10% pressure compensation pulse at a frequency of 10Hz) to improve interface adhesion. (2) Plasma treatment extension instruction: To address interface chemical instability, extend the plasma treatment time (for example, extend the original setting of 30s to 5-10s), suppress the formation of the insulating layer through high-energy particle bombardment, and enhance the lithium ion transmission capacity.

[0047] S30 , synchronously adjusting a hot pressing process and an interface treatment process based on the high-voltage compensation instruction and the plasma treatment extension instruction.

[0048] After the deep deterministic policy gradient algorithm predicts the high-voltage compensation instruction and the plasma treatment extension instruction, the high-voltage compensation instruction is directly connected to the hot pressing equipment control system to achieve high-frequency dynamic adjustment of temperature (control accuracy ±0.5°C) and pressure (adjustment frequency 10Hz); and the plasma treatment extension instruction is directly connected to the plasma control system to extend the plasma treatment time in real time.

[0049] On the one hand, compared to the delayed response of traditional PID control, the solution of the present invention can complete intervention before defects occur. On the other hand, through the coordinated adjustment of high-voltage compensation and extended plasma treatment, the problems of physical adhesion (mechanical parameters) and chemical stability (interface modification) are simultaneously solved. For example, during the hot pressing stage, pressure compensation is used to eliminate electrolyte layer pores, while extending the plasma treatment time to reduce interfacial chemical side reactions caused by pressure changes, ultimately significantly reducing interfacial impedance and improving yield.

[0050] It should be noted that the hot pressing and interface treatment steps of the present invention are carried out simultaneously. Specifically, the closed high-pressure environment of hot pressing (such as a vacuum chamber or an inert gas atmosphere) and the gas discharge conditions of plasma treatment can be compatible through integrated equipment. For example, a plasma nozzle is embedded in the hot pressing cavity, and the plasma discharge is maintained under high pressure by using partial pressure control technology (for example, a high-frequency pulse power supply is used to overcome the influence of gas pressure). At the same time, the hot pressing temperature (100-300°C) may affect the life of the plasma active particles, but this can be solved by optimizing the plasma gas source (for example, using an Ar / H2 mixed gas to improve the dissociation efficiency at high temperature) or adjusting the nozzle position (1-5mm away from the interface to balance energy injection and heat loss).

[0051] As an example, the predicting, based on the real-time data, whether the electrolyte layer meets the preset condition of incomplete bonding includes:

[0052] Normalizing the physical contact state data and the interface chemical state data to generate a first eigenvector and a second eigenvector, respectively;

[0053] Use a pre-built time series prediction model to predict the third eigenvector and the fourth eigenvector in the next M seconds based on the first eigenvector and the second eigenvector in the past N seconds;

[0054] The probability of incomplete alignment occurring in the next M seconds is predicted based on the spliced ​​third eigenvector and the fourth eigenvector.

[0055] Existing technologies often rely on static threshold judgments (e.g., an alarm is triggered when pressure exceeds a certain value), responding only after a defect occurs and failing to prevent material damage. Furthermore, they are unable to address the compound risks caused by multi-parameter coupling (e.g., the synergistic failure of pressure anomalies and interface chemical deterioration). To address this, the present invention uses a time-series prediction model to predict feature evolution over the next M seconds, shifting from a post-processing mode to a pre-processing mode. For example, a compensation mechanism is triggered 2 seconds before a sudden pressure change, improving defect suppression efficiency.

[0056] First, the heterogeneous multi-source data, i.e., physical contact state data and interface chemical state data, are converted into feature vectors of unified dimensions to eliminate dimensional differences and improve the generalization ability of the model. For example, the Z-score normalization method can be used for continuous variables such as pressure sensor data and electrolyte layer deformation monitoring data, using the formula Mapped to the standard normal distribution; using the Min-Max scaling method, bounded variables such as surface contact angle measurement data (0°-180°) and spectral peak intensity ratio (0-1) were scaled to the interval [0,1].

[0057] Next, generate the first feature vector and the second feature vector based on the real-time physical contact state data and the interfacial chemical state data respectively. Then, collect the historical feature vectors generated in the past N seconds, and use the time series prediction model to predict the third feature vector and the fourth feature vector at M seconds in the future. The time series prediction model can select the LSTM network, which can handle long sequence dependencies and is suitable for capturing features such as pressure fluctuations and slow changes in interfacial chemistry.

[0058] For example, predict the future porosity evolution based on the pressure change trend. For example, if the current pressure fluctuation period is 0.5 seconds, predict whether there will be a pressure mutation within the next 2 seconds to analyze the delamination risk; predict the Li2CO3 growth kinetics through the spectral feature change rate. For example, when the growth rate of the peak intensity ratio > 0.05 / s, predict the thickness of the insulating layer 2 seconds later to determine whether it exceeds the critical value (5nm).

[0059] Next, concatenate the third feature vector and the fourth feature vector, and use the probability prediction model based on the neural network to make a prediction to obtain the probability P of incomplete fitting within the next M seconds.

[0060] Among them, if the probability is within the first preset probability interval, for example, P ≤ 0.3, it indicates that the probability of incomplete fitting is very low, and the current process parameters can be maintained at this time. If the probability is within the second preset probability interval, for example, 0.3 < P ≤ 0.7, then execute the subsequent deep deterministic policy gradient algorithm to pre-adjust the process parameters. If the probability is within the third preset probability interval, for example, P > 0.7, then immediately execute high-pressure compensation and plasma extension.

[0061] Take the situation of "during the hot pressing process, the pressure suddenly drops by 1 MPa (physical anomaly), and at the same time, the peak intensity ratio of Li2CO3 increases by 0.08 (chemical anomaly)" as an example, and the explanation is as follows:

[0062] Generate feature vectors after standardization:

[0063] The first feature vector:

[0064] The first feature vector:

[0065] LSTM predicts the features in the next 2 seconds:

[0066] The third feature vector: It is predicted that the pressure will continue to drop by 2 MPa, and the deformation rate will increase to 0.25% / s. The fourth feature vector: It is predicted that the peak intensity ratio of Li2CO3 will reach 0.35, and the electron density will drop to 8×10 9 cm -3 .

[0067] Risk probability calculation: Based on the neural network-based probability prediction model output P = 0.65, the DDPG algorithm generates +8% pressure compensation and +7 seconds plasma extension instructions to pre-adjust process parameters.

[0068] As an example, a deep deterministic policy gradient algorithm is used to process the physical contact state data, the interface chemical state data, and historical process parameters to obtain a high-voltage compensation instruction and a plasma treatment extension instruction, including:

[0069] Inputting the physical contact state data, the interface chemical state data, and the historical process parameters into the Actor network of the DDPG algorithm to obtain an output initial instruction combination, including an initial high-voltage compensation instruction and an initial plasma treatment extension instruction;

[0070] The critic network calculates a Q value of the initial instruction combination, and when the Q value is lower than a first threshold, obtains a pressure correction value and a duration correction value based on gradient reversal, and adjusts the initial instruction combination based on the pressure correction value and the duration correction value until a target instruction combination is obtained that satisfies a Q value not lower than the first threshold;

[0071] The target instruction combination is simulated and executed using a digital twin system. If the interface porosity is lower than the second threshold and the Li2CO3 coverage is lower than the third threshold in the simulation results, the target instruction combination is confirmed, including the high-voltage compensation instruction and the plasma treatment extension instruction.

[0072] In the prior art, pressure compensation and plasma treatment extension are usually implemented separately. When high-pressure compensation is performed alone, it is difficult to further reduce the porosity after it drops to 5% (limited by the material yield strength, 8MPa is the critical pressure of aluminum-plastic film packaging); when plasma treatment is performed alone, after the Li2CO3 coverage drops to 3%, extending the time will cause surface amorphization (XRD peak half-height width increases from 0.8° to 1.5°), which in turn increases the interface impedance. To address the defects of the above-mentioned prior art of separate regulation, the present invention designs a method of synergistically optimizing mechanical regulation and chemical regulation. High pressure is used to create a physical environment suitable for plasma action. Plasma treatment can in turn optimize the pressure conduction efficiency. Finally, the optimal solution of the two in nonlinear space is found through the DDPG algorithm, achieving dual optimization of interface performance and process efficiency.

[0073] First, the real-time collected physical contact state data (pressure distribution gradient, deformation rate, surface contact angle, interface temperature distribution, etc.), interface chemical state data (Li2CO3 peak intensity ratio, plasma electron temperature, plasma electron density, interface redox potential, etc.) and historical process parameters (temperature average, pressure average, deformation rate average, holding time integral, etc. in the past 10 seconds) are integrated into a 12-dimensional state vector.

[0074] Then, forward propagation is performed through the Actor network (a two-layer 128-neuron ReLU neural network), directly outputting the initial command combination for the continuous action space, enabling rapid generation of preliminary control parameters based on the current production status. The initial command combination includes: an initial high-pressure compensation command: based on the baseline pressure, outputting a pressure compensation ratio within a range of ±10% (for example, a +7.5% pressure pulse); and an initial plasma treatment extension command: outputting a treatment time extension value within a range of 0-15 seconds (for example, +5.2 seconds).

[0075] The 12-dimensional state vector and the initial instruction combination are input into the Critic network (a three-layer 256-neuron LeakyReLU neural network) to calculate the Q value of the initial instruction combination to evaluate the expected effect of the instruction on reducing interface impedance and suppressing defects. If the Q value is lower than the first threshold (for example, 0.7), the following correction mechanism is triggered:

[0076] (1) Gradient reversal: Apply a gradient reversal layer to the Q value, reverse the gradient sign, and calculate the corrected gradient of the command parameters (pressure compensation ratio, plasma extension time);

[0077] (2) Parameter adjustment: Generate a pressure correction value (e.g., +0.8%) and a duration correction value (e.g., +1.6 seconds) based on the correction gradient, and iteratively adjust the initial instruction;

[0078] (3) Loop optimization: Repeat the evaluation and correction process until the Q value is no less than the first threshold (e.g., adjusted Q = 0.72), and output the target instruction combination.

[0079] In the above steps, quantitative evaluation and dynamic correction can ensure the optimization of instructions at the data level and avoid falling into local optimal solutions.

[0080] Next, the present invention further verifies the feasibility of the target instruction combination in an actual physical environment through virtual simulation, avoiding process failures caused by purely data-driven decisions and ensuring the synergistic effect of mechanical and chemical regulation. Specifically:

[0081] The target command combination that meets the Q-value requirements is input into the digital twin system, which is then connected to real-time status data from the solid-state battery production process. The digital twin system uses the finite element method to construct a coupled model of solid-state battery hot pressing and interface processing. It simulates the process two seconds after the command execution and extracts key physical indicators: interface porosity, which reflects the degree of physical adhesion and is compared with a second threshold (e.g., ≤5%); and Li2CO3 coverage, which reflects chemical stability and is compared with a third threshold (e.g., ≤3%).

[0082] If the simulation results simultaneously satisfy the conditions that porosity is less than the second threshold and Li2CO3 coverage is less than the third threshold (for example, porosity is 4.7% and coverage is 2.9%), the target instruction combination is confirmed as the final high-voltage compensation instruction and plasma treatment extension instruction; otherwise, the Actor network is returned to re-optimize.

[0083] As an example, Figure 2 As shown, the pressure correction value and the duration correction value are obtained based on the gradient reversal, including:

[0084] Processing the physical contact state data and the interface chemical state data using a pre-built external model to obtain a first correction step coefficient and a second correction step coefficient for use in a gradient reversal process;

[0085] The first correction step coefficient and the second correction step coefficient are used to adjust the initial pressure correction value and the initial duration correction value calculated during the gradient reversal process, respectively, to obtain the pressure correction value and the duration correction value.

[0086] Traditional gradient reversal relies on the Q-value feedback of the critic network for sequential fine-tuning, which has the following major drawbacks: Dimensional curse: In the 12-dimensional state space, there are over 200 possible combinations of pressure and plasma parameters, and blind search requires 5-8 iterations to converge. Physical defocus: Pure data optimization may generate overpressure instructions (e.g., 8.5 MPa), which, although meeting the Q-value requirement, violates the material mechanics constraints. Dynamic hysteresis: A fixed step size cannot adapt to process transients (e.g., when the temperature fluctuates by ±5°C, the plasma activity changes and the correction strategy needs to be dynamically adjusted). To address this, the present invention further constructs an external model for DDPG. This external model embeds process priors (e.g., the Pareto optimal surface of pressure-plasma) through pre-training, reducing the search dimension from 200+ to 10 key combinations. Combined with the real-time state (e.g., current pressure 7 MPa, Li2CO3 coverage 2.8%), it directly outputs the correction direction (e.g., pressure compensation + 0.5% + plasma treatment time + 2 seconds), reducing the number of iterations to 2-3, improving the efficiency of obtaining the optimal pressure and time correction values. The details are as follows:

[0087] The inputs to the pre-established external model are the previously described real-time monitored physical contact state data and interface chemical state data. The outputs are the correction step size coefficients k1 and k2 during the gradient reversal process. k1 is used to adjust the pressure correction step size, with a range of [0.5, 1.5]; k2 is used to adjust the duration correction step size, with a range of [0.3, 1.2]. Furthermore, the output of the external model can include a maximum number of iterations, N, ranging from 3 to 5, to avoid excessive iterations.

[0088] When the Q value of the initial instruction combination is lower than the first threshold, the initial pressure correction value and initial duration correction value are calculated by gradient reversal. The initial pressure correction value and initial duration correction value are adjusted using the correction step coefficients k1 and k2 obtained above to obtain the pressure correction value and duration correction value, respectively. The adjustment formula is as follows:

[0089]

[0090] Where ΔP adj is the pressure correction value, k1 is the first correction step coefficient, ΔP grad is the initial pressure correction value; Δt adj is the time correction value, k2 is the second correction step coefficient, Δt grad This is the initial duration correction value.

[0091] Combined with the N output by the external model, the following dual conditions for iteration termination are set: (1) the Q value reaches or exceeds the first threshold; (2) the number of iterations reaches N and the Q value improvement rate is less than 0.05 / time.

[0092] As an example, the external model includes a graph neural network module and an LSTM module. The graph neural network is used to process the coupling relationship between parameters, and the LSTM is used to capture time series features.

[0093] like Figure 2 As shown in the figure, the external model can adopt the architecture of graph neural network (GNN) and LSTM structure. The graph neural network (GNN) processes the coupling relationship between parameters and combines with LSTM to capture time series features.

[0094] The external model is trained in the following way:

[0095] Collect historical production data (temperature average, pressure average, deformation rate average, holding time integral, etc.) and digital twin simulation data to build a data set containing a large number of samples.

[0096] 28 key features such as pressure fluctuation frequency and Li2CO3 growth rate were extracted, and the 12-dimensional feature vector that was most relevant to the efficiency of obtaining the optimal pressure correction value and time correction value was screened out through mutual information calculation.

[0097] Design the following multi-objective loss function: L = αL step +βL iter +γL constraint

[0098] L step To minimize the modified step-size prediction error, L iter In order to optimize the balance between the number of iterations and the convergence speed, L constraint To penalize predictions that violate physical constraints (e.g., k1 > 1.5), α, β, and γ are weight coefficients.

[0099] The PPO algorithm is used to further optimize the parameters of the external model, prompting the external model to prioritize outputting a fast-converging and high-quality correction strategy. The reward function is designed as:

[0100]

[0101] Among them, R is the reward value; t converge is the convergence time, that is, the time it takes for the gradient reversal correction process to reach stability (Q value meets the requirements and iteration is terminated), in milliseconds (ms) or seconds (s); Q final The final Q value is the state value evaluation output by the Critic network (such as in the DDPG algorithm) after the correction is completed, reflecting the quality of the corrected process state (porosity, interface impedance, etc.).

[0102] like Figure 3 As shown, an embodiment of the present invention further provides an artificial intelligence-based solid-state battery production system 100, wherein the system 100 includes a collection unit 101, an instruction generation unit 102, and an adjustment unit 103;

[0103] The acquisition unit 101 is used to collect physical contact state data and interface chemical state data of the solid-state battery in production in real time, wherein the physical contact state data includes pressure sensor data, electrolyte layer deformation monitoring data, and surface contact angle measurement data; the interface chemical state data includes plasma emission spectrum data;

[0104] The instruction generation unit 102 is configured to predict, based on the real-time data, whether the electrolyte layer meets a preset condition of incomplete bonding, and if so, use a deep deterministic policy gradient algorithm to process the physical contact state data, the interface chemical state data, and historical process parameters to obtain a high-voltage compensation instruction and a plasma treatment extension instruction;

[0105] The adjusting unit 103 is configured to synchronously adjust the hot pressing process and the interface treatment process based on the high voltage compensation instruction and the plasma treatment extension instruction.

[0106] As an example, the instruction generation unit 102 is specifically configured to:

[0107] Normalizing the physical contact state data and the interface chemical state data to generate a first eigenvector and a second eigenvector, respectively;

[0108] Use a pre-built time series prediction model to predict the third eigenvector and the fourth eigenvector in the next M seconds based on the first eigenvector and the second eigenvector in the past N seconds;

[0109] The probability of incomplete alignment occurring in the next M seconds is predicted based on the spliced ​​third eigenvector and the fourth eigenvector.

[0110] As an example, the instruction generation unit 102 is specifically configured to:

[0111] Inputting the physical contact state data, the interface chemical state data, and the historical process parameters into the Actor network of the DDPG algorithm to obtain an output initial instruction combination, including an initial high-voltage compensation instruction and an initial plasma treatment extension instruction;

[0112] The critic network calculates a Q value of the initial instruction combination, and when the Q value is lower than a first threshold, obtains a pressure correction value and a duration correction value based on gradient reversal, and adjusts the initial instruction combination based on the pressure correction value and the duration correction value until a target instruction combination is obtained that satisfies a Q value not lower than the first threshold;

[0113] The target instruction combination is simulated and executed using a digital twin system. If the interface porosity is lower than the second threshold and the Li2CO3 coverage is lower than the third threshold in the simulation results, the target instruction combination is confirmed, including the high-voltage compensation instruction and the plasma treatment extension instruction.

[0114] As an example, the instruction generation unit 102 is specifically configured to:

[0115] Processing the physical contact state data and the interface chemical state data using a pre-built external model to obtain a first correction step coefficient and a second correction step coefficient for use in a gradient reversal process;

[0116] The first correction step coefficient and the second correction step coefficient are used to adjust the initial pressure correction value and the initial duration correction value calculated during the gradient reversal process, respectively, to obtain the pressure correction value and the duration correction value.

[0117] As an example, the external model includes a graph neural network module and an LSTM module. The graph neural network is used to process the coupling relationship between parameters, and the LSTM is used to capture time series features.

[0118] An embodiment of the present invention further provides an electronic device comprising: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements any of the aforementioned methods when executed by the processor.

[0119] An embodiment of the present invention further provides a computer storage medium storing a computer program that can be executed by a processor to implement any of the methods described above.

[0120] An embodiment of the present invention further provides a computer program product, which includes a computer program that can be executed by a processor to implement any of the methods described above.

[0121] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the composition and steps of each example according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0122] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A solid-state battery production method based on artificial intelligence, characterized by: The method comprises the following steps: Real-time collection of physical contact state data and interface chemical state data of solid-state batteries in production, the physical contact state data including pressure sensor data, electrolyte layer deformation monitoring data, and surface contact angle measurement data; the interface chemical state data including plasma emission spectroscopy data; Predicting whether the electrolyte layer meets a preset condition of incomplete bonding based on the real-time data, and if so, using a deep deterministic policy gradient algorithm to process the physical contact state data, the interface chemical state data, and historical process parameters to obtain a high-voltage compensation instruction and a plasma treatment extension instruction; A hot pressing process and an interface treatment process are synchronously adjusted based on the high voltage compensation instruction and the plasma treatment extension instruction.

2. The artificial intelligence-based solid-state battery production method according to claim 1, characterized in that: Predicting whether the electrolyte layer will meet a preset condition of incomplete bonding based on the real-time data includes: Normalizing the physical contact state data and the interface chemical state data to generate a first eigenvector and a second eigenvector, respectively; Use a pre-built time series prediction model to predict the third eigenvector and the fourth eigenvector in the next M seconds based on the first eigenvector and the second eigenvector in the past N seconds; The probability of incomplete alignment occurring in the next M seconds is predicted based on the spliced ​​third eigenvector and the fourth eigenvector.

3. The artificial intelligence-based solid-state battery production method according to claim 1, characterized in that: The physical contact state data, the interface chemical state data, and historical process parameters are processed using a deep deterministic policy gradient algorithm to obtain a high-voltage compensation instruction and a plasma treatment extension instruction, including: Inputting the physical contact state data, the interface chemical state data, and the historical process parameters into the Actor network of the DDPG algorithm to obtain an output initial instruction combination, including an initial high-voltage compensation instruction and an initial plasma treatment extension instruction; The critic network calculates a Q value of the initial instruction combination, and when the Q value is lower than a first threshold, obtains a pressure correction value and a duration correction value based on gradient reversal, and adjusts the initial instruction combination based on the pressure correction value and the duration correction value until a target instruction combination is obtained that satisfies a Q value not lower than the first threshold; The target instruction combination is simulated and executed using a digital twin system. If the interface porosity is lower than the second threshold and the Li2CO3 coverage is lower than the third threshold in the simulation results, the target instruction combination is confirmed, including the high-voltage compensation instruction and the plasma treatment extension instruction.

4. The artificial intelligence-based solid-state battery production method according to claim 3, characterized in that: Obtain pressure and duration corrections based on gradient reversal, including: Processing the physical contact state data and the interface chemical state data using a pre-built external model to obtain a first correction step coefficient and a second correction step coefficient for use in a gradient reversal process; The first correction step coefficient and the second correction step coefficient are used to adjust the initial pressure correction value and the initial duration correction value calculated during the gradient reversal process, respectively, to obtain the pressure correction value and the duration correction value.

5. The artificial intelligence-based solid-state battery production method according to claim 4, characterized in that: The external model includes a graph neural network module and an LSTM module. The graph neural network is used to process the coupling relationship between parameters, and the LSTM is used to capture time series features.

6. An artificial intelligence-based solid-state battery production system, characterized in that: The system includes an acquisition unit, an instruction generation unit, and an adjustment unit; The acquisition unit is used to collect physical contact state data and interface chemical state data of the solid-state battery in production in real time, wherein the physical contact state data includes pressure sensor data, electrolyte layer deformation monitoring data, and surface contact angle measurement data; the interface chemical state data includes plasma emission spectrum data; The instruction generation unit is configured to predict, based on the real-time data, whether the electrolyte layer meets a preset condition of incomplete bonding, and if so, use a deep deterministic policy gradient algorithm to process the physical contact state data, the interface chemical state data, and historical process parameters to obtain a high-voltage compensation instruction and a plasma treatment extension instruction; The adjustment unit is used to synchronously adjust the hot pressing process and the interface treatment process based on the high-voltage compensation instruction and the plasma treatment extension instruction.

7. The artificial intelligence-based solid-state battery production system according to claim 6, characterized in that: The instruction generation unit is specifically used to: Normalizing the physical contact state data and the interface chemical state data to generate a first eigenvector and a second eigenvector, respectively; Use a pre-built time series prediction model to predict the third eigenvector and the fourth eigenvector in the next M seconds based on the first eigenvector and the second eigenvector in the past N seconds; The probability of incomplete alignment occurring in the next M seconds is predicted based on the spliced ​​third eigenvector and the fourth eigenvector.

8. An electronic device, characterized in that: The electronic device comprises: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, wherein the computer program implements the method according to any one of claims 1 to 5 when executed by the processor.

9. A computer storage medium, characterized in that: The computer storage medium stores a computer program that can be executed by a processor to implement the method according to any one of claims 1 to 5.

10. A computer program product, characterized in that: The computer program product comprises a computer program that can be executed by a processor to implement the method according to any one of claims 1 to 5.

Images