An electrolyte high-throughput screening method, system, device and medium based on closed-loop feedback

Abstract

The present application relates to the cross-application technical field of new energy electrolyte and artificial intelligence, in particular to an electrolyte high-throughput screening method, system, device and medium based on closed-loop feedback, which comprises the following steps: S1, receiving formula parameters and converting them into a molecular structure feature matrix, inputting a pre-trained machine learning model to output a performance prediction index; S2, when the prediction index reaches an optimal standard, automatically converting the corresponding formula into a standardized experiment task work order and issuing it to a laboratory task board; S3, after receiving a physical experiment sample, performing electrochemical performance testing and exporting an original data file; S4, automatically analyzing the original data file based on an equivalent circuit fitting algorithm and calculating real physical test results; S5, comparing and analyzing the error between the prediction index and the real test results, marking the data with real results as high-confidence fact samples, giving them a higher learning proportion than theoretical simulation samples when the model is retrained, and completing closed-loop updating.

Description

Technical Field

[0001] This invention relates to the field of cross-application technology of new energy electrolytes and artificial intelligence, specifically to a high-throughput screening method, system, equipment and medium for electrolytes based on closed-loop feedback. Background Technology

[0002] In the research and development of new energy electrolytes, the traditional "expert experience combined with high-throughput exhaustive trial and error" model faces core bottlenecks such as extremely high trial and error costs, long development cycles, and a tendency to get stuck in local optima. To address this issue, existing technologies have attempted to introduce artificial intelligence or automated hardware for formula screening.

[0003] For example, the invention patent with publication number CN119943176B (A method for optimizing electrolyte formulation using artificial intelligence) discloses a technical solution that outputs an optimized formulation through data cleaning, feature engineering, and machine learning model training. However, the output of such systems is merely a formulation table; subsequent reagent weighing, mixing, and testing rely entirely on manual operation. The system's operation ends at the formulation output, failing to form an effective closed loop of R&D data, and the predictive model cannot learn and correct from real physical experimental results. Another example is the invention patent application with publication number CN121459961A (Machine Learning-Assisted Electrolyte Optimization Method, Device, Equipment, Medium, and Product), which discloses a technical solution that uses pure computational software such as molecular dynamics to generate simulated labels and then trains a machine learning model. This approach suffers from the problem that the calculated high-resolution materials often fail in actual synthesis due to the influence of complex physical environments such as impurities and interfacial side reactions. It lacks hardware-level coupling with the real physical experimental environment, and the model cannot learn from real experimental residuals.

[0004] In summary, there is an urgent need for a fully closed-loop material screening scheme that can utilize high-dimensional features for high-throughput intelligent prediction, automatically drive experimental equipment to complete physical verification, and automatically extract and feed back raw spectral data to the model. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention aims to provide a high-throughput electrolyte screening method based on closed-loop feedback, so as to achieve fully closed-loop automatic iterative optimization of intelligent prediction and physical experimental verification of electrolyte formulation.

[0006] The basic solution provided by this invention is a high-throughput screening method for electrolytes based on closed-loop feedback, comprising the following steps: S1. Feature Fusion and Prediction Steps: Receive formulation parameters, use the underlying chemical calculation tool library to convert the formulation parameters into a molecular structure feature matrix, input it into a pre-trained machine learning model, and output performance prediction indicators. S2. Physical work order triggering and issuance steps: When the performance prediction index reaches the preset optimization standard, the corresponding formula is automatically converted into a standardized experimental task work order and issued to the laboratory task board to instruct the execution end to perform physical preparation and on-machine testing of the sample according to the work order formula. S3. Physical test acquisition steps: After receiving the prepared physical experimental sample, perform electrochemical performance testing and export the raw data file including AC impedance spectroscopy. S4. Automatic test data parsing steps: After obtaining the original data file, the signal data in the original data file is processed based on the equivalent circuit fitting algorithm to automatically calculate the real physical test results, including the real ionic conductivity and charge transfer impedance. S5. Model closed-loop correction and update steps: Compare and analyze the error between the performance prediction index and the actual physical test results, and mark the data with the actual physical test results as high-confidence fact samples and store them in the database; when the model triggers retraining, assign the high-confidence fact samples a higher learning weight than the theoretical simulation samples, guide the machine learning model to correct the prediction bias, and complete the closed-loop update.

[0007] The principle of this invention lies in constructing a complete closed-loop link from virtual prediction to physical verification and then to model evolution. Specifically, the system first converts the electrolyte formulation parameters into a molecular structure feature matrix that can be recognized by the machine learning model, outputs performance prediction indicators, and achieves high-throughput initial screening of a large number of candidate formulations. When the prediction indicators reach the optimization criteria, the system automatically converts the formulation into a standardized experimental work order and sends it to the laboratory to drive the preparation and testing of physical experiments. Subsequently, the system automatically parses the raw data files exported by the electrochemical testing instruments to extract the actual physical test results such as ionic conductivity and charge transfer impedance. Finally, the system compares and analyzes the above-mentioned actual physical test results with the previous prediction indicators, calculates the error between the two, and marks the data with actual results as high-confidence fact samples. When the model is retrained, it is given a higher learning weight than the theoretical simulation samples, so that the model can continuously learn and correct from the residuals of the actual physical experiments.

[0008] The beneficial effects of this invention are as follows: 1. Automated extraction and standardization of test data. In existing technologies, test data such as electrochemical impedance spectroscopy must be manually fitted by experimenters using third-party software before being entered into the system. This process not only incurs significant system delays, but also suffers from inconsistent data standards due to subjective factors such as the selection of fitting intervals and equivalent circuits by different experimenters, severely restricting the quality and consistency of model training data. This invention, through an automatic test data parsing module, automatically processes the raw data files exported from the electrochemical workstation based on an equivalent circuit fitting algorithm. This replaces the cumbersome process of manual fitting and entry using third-party software, eliminating subjective biases arising from manual spectrum processing and ensuring the objectivity and accuracy of converting electrochemical test signals into structured physical characteristic results.

[0009] 2. Significantly reduces the trial-and-error costs and resource consumption of high-throughput material screening. Existing technologies for high-throughput experimental screening can only perform indiscriminate exhaustive testing according to a preset formulation table, failing to resolve the contradiction between the upper limit of experimental throughput and the infinite chemical space. Open-loop AI optimization schemes only output the formulation table, requiring manual verification one by one, resulting in low efficiency. This invention deeply couples front-end algorithmic intelligent prediction with back-end physical testing, using machine learning models to perform high-throughput initial screening of a vast initial formulation space. Physical experiment work orders are issued only to high-potential formulations that meet the selection criteria, overcoming the blindness of traditional indiscriminate exhaustive experiments, reducing the ineffective occupation of expensive chemical reagents and testing hardware channels by low-potential formulations, and improving the utilization efficiency of R&D resources.

[0010] 3. Constructing a closed-loop correction mechanism for models based on real physical verification. Existing technologies, which use purely computational methods such as molecular dynamics to generate simulated labels for training machine learning models, often result in high-resolution materials that fail in actual synthesis due to complex physical environments such as impurities and interfacial side reactions. These models lack hardware-level coupling with the real physical environment and cannot learn from real experimental residuals. Systems using open-loop or semi-closed-loop approaches also cannot dynamically correct their models based on actual experimental results. This invention precisely traces the real physical test results produced by automated equipment using a globally unique identifier and feeds them back to the data scheduling center as high-confidence factual samples. During model iteration, these samples are given a higher learning weight than theoretical simulated samples, enabling the model to learn and correct specifically from the real residuals of physical experiments. With the continuous accumulation of actual test data, the system can continuously adjust the learning preferences of the prediction algorithm, steadily improving the prediction accuracy and generalization ability of the machine learning model within a specific chemical space.

[0011] Furthermore, in S1, the formulation parameters are transformed into a molecular structure feature matrix using a low-level chemical calculation tool library, specifically including: One-dimensional chemical formula characters are parsed into a multi-dimensional topological graph structure; core molecular physicochemical descriptors, including molecular weight, lipid-water partition coefficient, polar surface area, number of rotatable bonds, and number of hydrogen bond donors and acceptors, are extracted and spliced ​​to form the molecular structure feature matrix.

[0012] The chemical information of electrolyte formulations is usually presented as text characters (such as the solvent name "EC"), and this one-dimensional text data cannot be directly processed by machine learning models. The aforementioned further approach parses the chemical formulation characters into a multi-dimensional topological graph structure (containing information such as atom types, chemical bonds, and ring structures), and extracts core molecular physicochemical descriptors from multiple dimensions. This transforms the abstract chemical text information into a structured numerical feature matrix, enabling machine learning models to perceive the structural and property differences between different formulation components at the molecular level, providing an informational foundation for subsequent accurate predictions. Specifically, molecular weight reflects molecular size, lipid-water partition coefficient reflects hydrophilic / hydrophobic properties, and polar surface area and hydrogen bond donor / acceptor number reflect molecular polarity and the ability to form intermolecular interactions. These descriptors collectively characterize the influence of formulation components on the macroscopic performance of the electrolyte from different dimensions, enhancing the model's ability to represent the formulation.

[0013] Furthermore, the pre-trained machine learning model is a hybrid architecture model that integrates tree model algorithm and deep neural network; S1 also includes: when outputting performance prediction indicators, extracting feature decision weights of multiple dimensions in the hybrid model in real time, and classifying them into thermodynamic condition category, macroscopic physical category and spatial topology category according to preset rules, and dynamically rendering them into a visual attribution decision chart in the display interface.

[0014] Tree-based model algorithms (such as XGBoost) possess strong capabilities in fitting tabular data and ranking feature importance, while deep neural networks excel at capturing high-order nonlinear interactions between features. The aforementioned further approach integrates both into a hybrid architecture model, balancing the model's fitting accuracy and generalization ability for electrolyte formulation data. Simultaneously, by extracting the decision weights within the model in real time and presenting them in attribution charts, researchers can intuitively understand which molecular features have a significant impact on the current prediction results, thereby enhancing the interpretability of the AI ​​model and assisting researchers in making more accurate judgments during the manual verification phase.

[0015] Furthermore, S2 also includes: after receiving the manual pre-verification and confirmation instruction from the experimenter regarding the physical compatibility of the formula, activating the physical experiment execution permission of the corresponding work order to filter formulas with abnormal physical properties.

[0016] Although the system uses machine learning models to initially screen formulation performance, some formulations with predicted high compatibility still lack practical operability at the physical level (e.g., the solute cannot be completely dissolved in the solvent, or the solution exhibits stratification or precipitation). The aforementioned further solution incorporates a pre-testing step before physical experiments. This leverages the expertise of experimenters to quickly assess the physical compatibility of the formulations. Experiment execution is only activated after successful testing, filtering out invalid formulations with physical anomalies at a lower time cost. This avoids wasting chemical reagents and testing channels in subsequent physical preparation and testing stages, further reducing overall screening costs.

[0017] Furthermore, in S4, the signal data in the original data file is processed based on the equivalent circuit fitting algorithm, specifically including: For the raw data file exported from the electrochemical workstation, regular expressions and specific instrument header file filtering rules are used to clean and extract the data. The semicircular and oblique features of the Nyquist plot are fitted from the cleaned data to output the core electrochemical parameters of the material.

[0018] The raw data files exported from different electrochemical workstations vary in format, header information, delimiters, and column order. The aforementioned further approach uses regular expressions and instrument-specific header file filtering rules to uniformly clean and extract raw files of different formats, ensuring the accuracy and reliability of the data segments used in subsequent fitting processing. Based on this, the semicircular and oblique line features of the Nyquist plot are automatically fitted from the cleaned data, where the semicircular diameter corresponds to charge transfer impedance and the high-frequency intercept corresponds to solution resistance. This allows for the automatic output of core electrochemical parameters without requiring manual operation of third-party fitting software, ensuring the objectivity and real-time nature of data analysis.

[0019] Furthermore, the model closed-loop correction and update steps in S5 specifically include: The real physical test results in the original data file are automatically aligned and back-archived with the recipe entries corresponding to the performance prediction index generated in S1 using a globally unique identifier; the absolute error between the performance prediction index and the real physical test results is calculated, and the loss function weight of the high-confidence fact sample is adjusted in the next model iteration training based on the value of the absolute error.

[0020] In high-throughput screening scenarios, the system simultaneously processes a large number of formulations and their corresponding prediction and physical experiment tasks. Accurately tracing each test data file to its corresponding predicted formulation entry is a crucial prerequisite for achieving closed-loop feedback. This solution assigns a globally unique identifier to each task during work order generation. This identifier accompanies the formulation throughout the entire process from prediction to preparation, testing, and analysis, enabling precise automatic alignment between physical test results and predicted entries, avoiding data confusion and erroneous associations. Furthermore, the loss function weights of samples are adjusted based on error values; the larger the error, the higher the weight of that sample in the loss function. This allows for targeted correction of highly biased samples during model retraining, guiding the model to quickly correct prediction biases with optimal sample efficiency, accelerating model convergence and evolution.

[0021] Correspondingly, the present invention also discloses a high-throughput electrolyte screening system based on closed-loop feedback for implementing the above method, comprising: Data scheduling module: used to receive formula parameters, use the underlying chemical calculation tool library to convert the formula parameters into a molecular structure feature matrix, and perform closed-loop correction and update; Model prediction module: used to output performance prediction indicators, and when the performance prediction indicators reach the preset optimization criteria, to issue an experimental task work order to the laboratory task dashboard; Laboratory automated testing equipment group: used to perform physical preparation and on-machine testing of samples according to the experimental task work order, and to export raw data files; Automatic test data parsing module: used to obtain the original data file and automatically calculate the actual physical test results based on the equivalent circuit fitting algorithm.

[0022] Furthermore, the test data automatic parsing module runs residently as a daemon process, and monitors file creation events in the network shared directory in real time through a system-level file monitoring mechanism to achieve real-time capture of the original data files.

[0023] Correspondingly, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0024] Correspondingly, the present invention also discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Attached Figure Description

[0025] Figure 1 This is a structural diagram of an embodiment of a high-throughput electrolyte screening system based on closed-loop feedback according to the present invention.

[0026] Figure 2 This is a flowchart illustrating an embodiment of a high-throughput electrolyte screening method based on closed-loop feedback according to the present invention.

[0027] Figure 3 This is a flowchart illustrating the feature fusion of the underlying chemical calculation tool library in an embodiment of the high-throughput screening method for electrolytes based on closed-loop feedback according to the present invention.

[0028] Figure 4 This is an equivalent circuit fitting flowchart of an embodiment of a high-throughput electrolyte screening method based on closed-loop feedback according to the present invention.

[0029] Figure 5 This is a flowchart illustrating the closed-loop triggered retraining process of a model in an embodiment of a high-throughput electrolyte screening method based on closed-loop feedback according to the present invention. Detailed Implementation

[0030] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: A high-throughput electrolyte screening system based on closed-loop feedback includes a data scheduling module, a model prediction module, a group of automated laboratory testing equipment, and an automatic test data parsing module; The data scheduling module is used to receive formula parameters, use the underlying chemical calculation tool library to convert the formula parameters into a molecular structure feature matrix, and perform closed-loop correction and update. The model prediction module is used to output performance prediction indicators, and when the performance prediction indicators reach the preset optimization criteria, it sends an experimental task work order to the laboratory task dashboard. The automated laboratory testing equipment group is used to perform physical preparation and on-machine testing of samples according to the experimental task work order, and to export raw data files; The test data automatic parsing module is used to obtain the original data file and automatically calculate the real physical test results based on the equivalent circuit fitting algorithm. This module runs residently as a daemon process and listens for file creation events in the network shared directory in real time through a system-level file monitoring mechanism to achieve real-time capture of the original data file.

[0031] In terms of physical topology, the data scheduling module, model prediction module, and automatic test data parsing module are deployed within a core server. This core server communicates with the laboratory automated testing equipment group and the laboratory task dashboard via a local area network (LAN). Specifically, the core server is deployed within the laboratory LAN, and internally, the data scheduling module, model prediction module, and automatic test data parsing module are isolated using containerization technology. The laboratory automated testing equipment group includes at least an electrochemical workstation and a battery charge / discharge test cabinet. The host computer for these testing devices is connected to the core server within the same LAN segment via a LAN switch. The laboratory task dashboard receives and displays experimental task work orders issued by the model prediction module and allows experimental personnel to perform manual pre-verification and confirmation operations.

[0032] The core server is configured with a network shared directory mounting service. The host computer software for devices such as electrochemical workstations is configured to automatically save or synchronize the raw spectral data after each test to this designated network shared directory. The automatic test data parsing module runs resident as a daemon process, monitoring file creation events in this shared directory in real time based on a system-level file monitoring mechanism, enabling real-time capture of test files.

[0033] Based on the above system, this invention provides a high-throughput electrolyte screening method based on closed-loop feedback, as shown in the attached figure. Figure 2 As shown, it includes the following steps: S1. Feature Fusion and Prediction Steps: Receive the formulation parameters, use the underlying chemical calculation tool library to convert the formulation parameters into a molecular structure feature matrix, and input it into a pre-trained machine learning model to output performance prediction indicators.

[0034] Specifically, the received formulation parameters include basic formulation information such as solvent ratio and lithium salt concentration. These parameters are then converted into a molecular structure feature matrix using a low-level chemical calculation tool library, as shown in the attached figure. Figure 3As shown, the specific components include: (1) Formulation parameter parsing and molecular topology graph construction. The underlying chemical calculation tool library first uses the built-in solvent name-SMILES mapping dictionary to automatically convert the one-dimensional chemical formulation characters (such as "EC", "DMC", "G1" and other solvent abbreviations) entered by the user into the corresponding standard SMILES (Simplified Molecular Input Line Entry System) strings. Then, it calls the molecular parsing engine of the open-source cheminformatics tool library RDKit to parse the SMILES string into a multi-dimensional molecular topology graph structure containing a set of atoms and a set of chemical bonds. Each atom node uses the atomic number as the initial feature, and each edge is encoded by the chemical bond connection relationship between atoms. For multi-solvent mixture formulations (such as "EC:DMC=1:1"), the tool library decomposes each component and its ratio according to the separator, constructs a topology graph for each solvent molecule, and records the molar or volume ratio weight of each component. (2) Core molecular physicochemical descriptor extraction tool library uses the molecular descriptor calculation engine of RDKit to extract the following physicochemical descriptors for each solvent molecule in batches. (3) Graph Neural Network Topology Feature Extraction: The tool library embeds a pre-trained Graph Neural Network (GNN) feature extractor. This extractor uses a two-layer Graph Convolutional Network (GCNConv), with the atomic number of the molecular topology graph constructed in (1) as the node input feature and the chemical bond connection relationship as the edge index. After two layers of graph convolution and ReLU activation, global mean pooling is performed, and then it is mapped to a fixed-length 16-dimensional topology feature embedding vector through a fully connected layer. This embedding vector encodes the three-dimensional topological spatial information of the solvent molecule (such as branched structure, cyclic skeleton, functional group connection mode, etc.), which makes up for the spatial structural features that traditional two-dimensional descriptors cannot represent. (4) Multidimensional Feature Matrix Concatenation: The tool library performs weighted averaging of the RDKit physicochemical descriptor submatrices of each solvent and the GNN topology embedding vector according to the ratio weights, to obtain a unified topology feature vector at the formulation level. Subsequently, the process parameters (temperature, lithium salt concentration, lithium salt type ID) are horizontally concatenated with the two types of molecular features mentioned above to form the final molecular structure feature matrix. This feature matrix integrates information from three levels: macroscopic process conditions, microscopic molecular physicochemical properties, and molecular three-dimensional topological space features, providing a complete multi-dimensional input representation for downstream prediction models. One-dimensional chemical formula characters are parsed into multi-dimensional topological graph structures. Core molecular physicochemical descriptors, including molecular weight, lipid-water partition coefficient, polar surface area, number of rotatable bonds, and number of hydrogen bond donors and acceptors, are extracted and concatenated to form the molecular structure feature matrix.The underlying chemical calculation tool library first parses one-dimensional chemical formula characters (such as "EC") into multi-dimensional topological graph structures. Then, it extracts core molecular physicochemical descriptors such as molecular weight, lipid-water partition coefficient, polar surface area, number of rotatable bonds, and number of hydrogen bond donors and acceptors for each molecule, and splices them to form a multi-dimensional molecular structure feature matrix.

[0035] The pre-trained machine learning model is a hybrid architecture model that combines a tree model algorithm and a deep neural network. After receiving the feature matrix, the model outputs performance prediction indicators such as the predicted ionic conductivity and the predicted electrochemical stability window of the formulation. When outputting performance prediction indicators, the feature decision weights of multiple dimensions in the hybrid model are extracted in real time and divided into thermodynamic condition categories, macroscopic physical categories, and spatial topological categories according to preset rules. These are then dynamically rendered into a visual attribution decision chart (such as a SHAP force graph) on the display interface to help researchers understand the model's prediction logic.

[0036] S2. Physical work order triggering and distribution steps: When the performance prediction index reaches the preset optimization standard, the corresponding formula is automatically converted into a standardized experimental task work order and distributed to the laboratory task board to instruct the execution end to perform physical preparation and on-machine testing of the sample according to the work order formula.

[0037] Specifically, when the predicted ionic conductivity exceeds the system's preset optimal threshold, the model prediction module automatically generates a standardized experimental task order containing a globally unique identifier (UUID). This task order details the material ratios and testing environment parameters, including solvent ratio, lithium salt concentration, mixing temperature, and stirring time. The model prediction module then pushes this task order to the laboratory task dashboard. Researchers review the task order on the dashboard and manually pre-verify the physical compatibility of the formulation (e.g., whether it separates into layers or precipitates). After successful verification, researchers confirm the instructions on the dashboard to activate the physical experiment execution permission for the corresponding task order, filtering out formulations with abnormal physical properties. Subsequently, researchers prepare the physical sample according to the task order ratio, or the system automatically activates the laboratory's automated testing equipment (e.g., an automatic liquid dispenser) to perform automated preparation.

[0038] S3. Physical test acquisition steps: After receiving the prepared physical experimental sample, perform electrochemical performance testing and export the raw data file including AC impedance spectroscopy.

[0039] Specifically, the researchers loaded the prepared sample onto the electrochemical workstation for electrochemical impedance spectroscopy (EIS) scanning. After the test was completed, the host computer of the electrochemical workstation exported the raw data file named in the format "UUID_test time.txt" to the network shared directory configured in the core server according to the preset path.

[0040] S4. Automatic test data parsing steps: After obtaining the original data file, the signal data in the original data file is processed based on the equivalent circuit fitting algorithm to automatically calculate the real physical test results, including the true ionic conductivity and charge transfer impedance.

[0041] Specifically, as shown in the attached document Figure 4 As shown, the equivalent circuit fitting algorithm is implemented through the following steps: (1) Data preprocessing The test data automatic parsing module runs residently as a daemon process, monitoring the network shared directory in real time through a system-level file monitoring mechanism. Upon detecting a new file creation event, it immediately retrieves the raw data file. For the raw data files exported from the electrochemical workstation, it uses regular expressions and instrument-specific header file filtering rules for cleaning and extraction, stripping away data segments corresponding to high-frequency inductive reactance (imaginary part Z''>0) and low-frequency diffusion impedance, retaining the mid-frequency main data segment containing charge transfer semicircular characteristics.

[0042] (2) Equivalent circuit model For the liquid electrolyte system of lithium-ion batteries, a Randle-type equivalent circuit model is adopted, with its topology being an ohmic resistor. Series of charge transfer resistors The interface impedance unit is formed by connecting it in parallel with the constant phase angle element CPE, and then connecting it in series with the semi-infinite diffusion Warburg impedance. W ,Right now:

[0043] Frequency domain impedance of each component as follows: Ohm resistance: .

[0044] Constant phase angle element: ,in For CPE coefficient, CPE index ( ).

[0045] The charge transfer resistor is connected in parallel with the CPE: .

[0046] Warburg diffusion impedance: ,in is the diffusion coefficient.

[0047] The total impedance of the equivalent circuit is: .

[0048] (3) Fitting method Without requiring manual operation of third-party data fitting software, the equivalent circuit model described above is automatically fitted to the measured impedance spectrum data using the complex nonlinear least squares (CNLS) method combined with the Levenberg-Marquardt iterative algorithm, and the parameter vector is solved. The fitting aims to minimize the weighted sum of squared residuals between the measured values ​​and the model values. The weights are normalized using a modulus normalization strategy to balance the contributions of data points with different impedance amplitudes.

[0049] (4) Derivation of electrochemical parameters After the fitting converges, the core physical quantities are extracted: True ionic conductivity (mS / cm): ,in For electrolyte thickness, The effective area of ​​the electrode; True charge transfer impedance: directly take the fitted result value.

[0050] Specifically, the automatic test data parsing module runs resident as a daemon process, monitoring the shared directory in real time through a system-level file monitoring mechanism. Upon detecting a new file creation event, it immediately retrieves the original data file. For the original data file exported from the electrochemical workstation, it is cleaned and extracted using regular expressions and specific instrument header file filtering rules, stripping away data segments corresponding to high-frequency inductive reactance and low-frequency diffusion impedance. Subsequently, without requiring manual operation of third-party data fitting software, scientific computing logic is directly invoked from the cleaned data to automatically locate and fit the semicircular and oblique features of the Nyquist plot, calculating the sample's true charge transfer impedance and true ionic conductivity.

[0051] S5. Model closed-loop correction and update steps: Compare and analyze the error between the performance prediction index and the actual physical test results, and mark the data with the actual physical test results as high-confidence fact samples and store them in the database; when the model triggers retraining, assign the high-confidence fact samples a higher learning weight than the theoretical simulation samples, guide the machine learning model to correct the prediction bias, and complete the closed-loop update.

[0052] Specifically, the model retraining is triggered by a composite triggering strategy through the data scheduling module, as shown in the attached figure. Figure 5 As shown, the model retraining process will be automatically initiated when any of the following conditions are met: (1) Periodic Timed Trigger: The system automatically triggers model retraining at a preset fixed period (default is every 7 days) in the form of a Cron timed daemon task. When the trigger time is reached, the data scheduling module first checks whether any valid high-confidence fact samples have been added since the last retraining; if no new samples have been added, the trigger is skipped to avoid meaningless empty runs. The trigger period can be manually adjusted in the system configuration interface according to the frequency of experimental progress feedback (the range can be set from 3 to 30 days).

[0053] (2) Sample Accumulation Trigger: The system monitors the number of data entries marked as "high-confidence fact samples" in real time. When this number accumulates to a preset threshold N (default N=50), model retraining is triggered immediately. This mechanism ensures that during periods of intensive experimental feedback, such as the high-throughput batch screening phase, the model can absorb new data in a timely manner without waiting for the periodic trigger point, achieving a tightly coupled closed-loop rhythm of "learning as soon as data arrives". The threshold N can be adjusted according to the model maturity: N=30 is recommended in the initial stage (v1 stage), and N=50~100 is recommended in the mid-term stage (v2 stage).

[0054] (3) Prediction Error Drift Trigger: The data scheduling module maintains a sliding window to track the prediction error (|AI predicted value - actual test value|) of the most recent M samples (M=20 by default) that have completed closed-loop feedback. When the mean absolute error (MAE) within the sliding window exceeds the preset warning threshold (1.5mS / cm by default or twice the MAE of the test set during model training), it is determined that the prediction distribution has drifted significantly (Concept Drift), and model retraining is immediately triggered. This mechanism is used to capture model performance degradation caused by changes in external factors such as experimental conditions, material batches, and environmental temperature and humidity.

[0055] (4) Manual triggering: In addition, the system retains an entry button for manual retraining in the management panel, allowing experimenters to initiate model updates when they discover systematic prediction biases.

[0056] Specifically, the data scheduling module uses the globally unique identifier (UUID) in the filename to precisely align the extracted true ionic conductivity with the original predicted formulation entry in step S1, completing the automatic data write-back and archiving. The data scheduling module calculates the absolute error between the predicted and true values. If the error is within an acceptable range, the data is stored in the database as a normal fact sample; if the error is large, the data scheduling module marks the complete data chain with the true result as a high-confidence fact sample. During periodically triggered model iteration training, the data scheduling module dynamically adjusts the loss function weight of this high-confidence fact sample based on the error value: the larger the error, the higher the weight of the sample in the loss function. By assigning a higher learning weight to high-confidence fact samples than to theoretically simulated samples, the model prediction module can more effectively correct prediction biases caused by purely theoretical derivations during retraining, achieving adaptive closed-loop upgrades of the core algorithm.

[0057] By cyclically executing the above steps S1 to S5, the prediction accuracy of this system continuously improves with the accumulation of physical experimental data.

[0058] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A high-throughput screening method for electrolytes based on closed-loop feedback, characterized in that, Includes the following steps: S1. Feature Fusion and Prediction Steps: Receive formulation parameters, use the underlying chemical calculation tool library to convert the formulation parameters into a molecular structure feature matrix, input it into a pre-trained machine learning model, and output performance prediction indicators. S2. Physical work order triggering and issuance steps: When the performance prediction index reaches the preset optimization standard, the corresponding formula is automatically converted into a standardized experimental task work order and issued to the laboratory task board to instruct the execution end to perform physical preparation and on-machine testing of the sample according to the work order formula. S3. Physical test acquisition steps: After receiving the prepared physical experimental sample, perform electrochemical performance testing and export the raw data file including AC impedance spectroscopy. S4. Automatic test data parsing steps: After obtaining the original data file, the signal data in the original data file is processed based on the equivalent circuit fitting algorithm to automatically calculate the real physical test results, including the real ionic conductivity and charge transfer impedance. S5. Model closed-loop correction and update steps: Compare and analyze the error between the performance prediction index and the actual physical test results, and mark the data with the actual physical test results as high-confidence fact samples and store them in the database; when the model triggers retraining, assign the high-confidence fact samples a higher learning weight than the theoretical simulation samples, guide the machine learning model to correct the prediction bias, and complete the closed-loop update.

2. The high-throughput electrolyte screening method based on closed-loop feedback according to claim 1, characterized in that, In S1, the formulation parameters are transformed into a molecular structure feature matrix using a low-level chemical calculation tool library, specifically including: One-dimensional chemical formula characters are parsed into a multi-dimensional topological graph structure; core molecular physicochemical descriptors, including molecular weight, lipid-water partition coefficient, polar surface area, number of rotatable bonds, and number of hydrogen bond donors and acceptors, are extracted and spliced ​​to form the molecular structure feature matrix.

3. The high-throughput electrolyte screening method based on closed-loop feedback according to claim 1, characterized in that, The pre-trained machine learning model is a hybrid architecture model that integrates tree model algorithm and deep neural network; S1 also includes: when outputting performance prediction indicators, extracting feature decision weights of multiple dimensions in the hybrid model in real time, and classifying them into thermodynamic condition category, macroscopic physical category and spatial topology category according to preset rules, and dynamically rendering them into a visual attribution decision chart in the display interface.

4. The high-throughput electrolyte screening method based on closed-loop feedback according to claim 1, characterized in that, S2 also includes: after receiving a manual pre-verification instruction from the experimenter regarding the physical compatibility of the formula, activating the physical experiment execution permission of the corresponding work order to filter formulas with abnormal physical properties.

5. The high-throughput electrolyte screening method based on closed-loop feedback according to claim 1, characterized in that, S4 processes the signal data in the original data file based on the equivalent circuit fitting algorithm, specifically including: For the raw data file exported from the electrochemical workstation, regular expressions and specific instrument header file filtering rules are used to clean and extract the data. The semicircular and oblique features of the Nyquist plot are fitted from the cleaned data to output the core electrochemical parameters of the material.

6. The high-throughput electrolyte screening method based on closed-loop feedback according to claim 1, characterized in that, The specific steps for model closed-loop correction and update in S5 include: The real physical test results in the original data file are automatically aligned and back-archived with the recipe entries corresponding to the performance prediction index generated in S1 using a globally unique identifier; the absolute error between the performance prediction index and the real physical test results is calculated, and the loss function weight of the high-confidence fact sample is adjusted in the next model iteration training based on the value of the absolute error.

7. A high-throughput electrolyte screening system based on closed-loop feedback, characterized in that, For implementing the method according to any one of claims 1 to 6, comprising: Data scheduling module: used to receive formula parameters, use the underlying chemical calculation tool library to convert the formula parameters into a molecular structure feature matrix, and perform closed-loop correction and update; Model prediction module: used to output performance prediction indicators, and when the performance prediction indicators reach the preset optimization criteria, to issue an experimental task work order to the laboratory task dashboard; Laboratory automated testing equipment group: used to perform physical preparation and on-machine testing of samples according to the experimental task work order, and to export raw data files; Automatic test data parsing module: used to obtain the original data file and automatically calculate the actual physical test results based on the equivalent circuit fitting algorithm.

8. The electrolyte high-throughput screening system based on closed-loop feedback according to claim 7, characterized in that, The test data automatic parsing module runs residently as a daemon process, and monitors file creation events in the network shared directory in real time through a system-level file monitoring mechanism to achieve real-time capture of the original data files.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

Images