Nanofiber membrane production quality prediction and process optimization system based on digital twinning

Abstract

The application relates to the technical field of nanofiber membrane process optimization and intelligent control, and discloses a nanofiber membrane production quality prediction and process optimization system based on digital twinning, which comprises a multi-scale physical property gene coding unit, a physical interaction heterogeneous graph construction unit, a physical grammar graph neural network prediction unit, an active learning parameter optimization unit and a model closed-loop adaptive correction unit.The multi-scale physical property gene coding unit is used for collecting material and process parameters of a fiber membrane preparation process and constructing multi-scale physical property genes representing physical and chemical properties; the physical interaction heterogeneous graph construction unit is used for constructing a physical interaction heterogeneous graph for a specific formula and process combination; the physical grammar graph neural network prediction unit is used for generating a final node feature vector and generating a macro quality prediction index of the fiber membrane and corresponding prediction uncertainty evaluation; the active learning parameter optimization unit is used for solving a next experimental point parameter combination; and the model closed-loop adaptive correction unit is used for obtaining real experimental data corresponding to the optimal next experimental point parameter combination and feeding back and fine-tuning the prediction unit.The application can effectively solve the pain points that a traditional black box model is inaccurate and cannot be explained.

Description

Technical Field

[0001] This invention relates to the field of nanofiber membrane process optimization and intelligent control technology, specifically a nanofiber membrane production quality prediction and process optimization system based on digital twins. Background Technology

[0002] Electrospinning technology is a key process for preparing nanofiber membranes. The quality of the final product, such as fiber diameter and porosity, is subject to complex influences from material formulation and process parameters. In actual research and development and production, the search space for formulation-process parameter combinations is extremely high, and the internal physicochemical interactions at multiple scales make it extremely difficult to accurately predict and control the macroscopic quality of nanofiber membranes.

[0003] Current technologies mainly rely on expert experience to conduct a large number of trial and error experiments. This approach not only has a long development cycle and high cost, but also makes it difficult to systematically find the globally optimal combination of process parameters. In addition, traditional pure data-driven models are often black box models, and their prediction process lacks adherence to and explanation of physical laws, resulting in weak model generalization ability and difficulty in truly guiding complex process optimization processes, which leads to bottlenecks in product quality stability and performance improvement.

[0004] Therefore, there is an urgent need for a solution that can accurately, efficiently, and interpretably predict quality and optimize processes to address the problems existing in current technologies.

[0005] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the aforementioned technical problems, the purpose of this invention is to provide a digital twin-based system for predicting the production quality and optimizing the process of nanofiber membranes. Specifically, the technical solution of this invention includes:

[0007] Multiscale property gene coding units are used to collect material parameters and process parameters in the nanofiber membrane preparation process, and to perform structuring on the material parameters and process parameters in order to construct multiscale property genes that characterize physicochemical properties.

[0008] Physical interaction heterogeneity graph construction unit, used to construct physical interaction heterogeneity graphs for specific formulation and process combinations based on multi-scale physical property genes;

[0009] The physical grammar graph neural network prediction unit is used to receive the physical interaction heterogeneous graph and perform message passing and node update processing to generate the final node feature vector containing global physical interaction information. Then, the final node feature vector is decoded to generate the macroscopic quality prediction index of the nanofiber membrane and the corresponding prediction uncertainty assessment.

[0010] The active learning parameter optimization unit is used to calculate the optimal combination of parameters for the next experimental point based on macroscopic quality prediction indicators and prediction uncertainty assessment.

[0011] The model closed-loop adaptive correction unit is used to obtain the real experimental data corresponding to the optimal next experimental point parameter combination, and to perform feedback fine-tuning of the physical grammar graph neural network prediction unit based on the real experimental data.

[0012] Furthermore, the process of structuring multi-scale property gene coding units is as follows:

[0013] The molecular-scale chemical properties of the material are analyzed to determine the molecular weight descriptor, chain structure information, and functional group information; the macroscopic physicochemical properties of the solution-scale layer are analyzed to determine the Hansen solubility parameter, intrinsic viscosity, and dielectric constant; and the process response layer characteristics of the material under the process environment are analyzed to determine the model coefficients of polarizability and solvent evaporation rate under an electric field.

[0014] Furthermore, the construction process of the physical interaction heterogeneous graph is as follows:

[0015] Material components or process parameters are set as nodes, and the initial feature vector is generated by multi-scale property gene encoding; the physicochemical interactions between nodes are set as edges, and the type and weight of the edges are calculated based on the parameters in the multi-scale property gene of the corresponding node.

[0016] Furthermore, the process of message passing and node update processing by the physical grammar graph neural network prediction unit is as follows:

[0017] The message passed from the neighboring node to the central node is defined as a weighted sum of multiple physical action terms; the physical action terms are constructed by modulating the interaction strength between node feature vectors by using physical parameters derived from multi-scale physical property genes as adjustment factors; all messages received from the neighboring nodes are aggregated and combined with the state of the node itself in the previous layer to generate an updated node state.

[0018] Furthermore, the physical parameters include the normalized intrinsic viscosity of the solution, which is related to viscoelasticity, and the normalized dielectric constant of the solution, which is related to charge conduction.

[0019] Furthermore, the decoding process of the physical grammar graph neural network prediction unit is as follows:

[0020] Using a preset aggregation function, the final node feature vectors of all nodes are fused into a single vector representing the entire graph. Through a readout function, the single vector representing the entire graph is mapped to the macroscopic quality prediction index of the nanofiber membrane. The macroscopic quality prediction index includes the average fiber diameter and porosity.

[0021] Furthermore, the process by which the active learning parameter optimization unit calculates the optimal combination of parameters for the next experimental point is as follows:

[0022] Define a scalarized comprehensive objective function based on preset requirements; input the comprehensive objective score generated by the macro quality prediction index and the prediction uncertainty assessment into the preset upper confidence bound acquisition function for calculation, and determine the parameter combination that maximizes the calculation result as the optimal parameter combination for the next experimental point.

[0023] Furthermore, the feedback fine-tuning process of the model's closed-loop adaptive correction unit is as follows:

[0024] The optimal combination of parameters for the next experimental point is combined with real experimental data to form a new data point; the new data point is added incrementally to the training dataset; the stochastic gradient descent algorithm is used to calculate the gradient of the loss function between the current model's predicted output for the new experimental point and the real experimental data; and the model parameters of the prediction unit of the physical grammar graph neural network are updated based on the gradient of the loss function and the preset learning rate.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. This invention addresses the pain points of inaccurate and uninterpretable predictions in traditional black-box models. This system constructs a multi-scale property gene containing physicochemical significance and embeds core physical laws such as viscoelasticity and charge conduction as grammars into the computation of a graph neural network, deeply integrating domain knowledge with artificial intelligence. This results in high model prediction accuracy, strong generalization ability, and physical consistency in the computation process, enhancing the interpretability and reliability of the results and reducing reliance on massive amounts of training data.

[0027] 2. This invention significantly improves R&D efficiency and reduces costs. The system constructs a closed-loop intelligent process of prediction-optimization-experimentation-correction. Through an active learning parameter optimization unit, it can intelligently recommend the optimal next experimental point based on model prediction and uncertainty assessment. This data-driven strategy replaces the traditional trial-and-error method that relies on expert experience, greatly reducing the number of blind experiments and thus significantly shortening the R&D cycle of new material formulations and processes, saving valuable material and human resource costs.

[0028] 3. This invention achieves global optimization of a high-dimensional, complex process space. The formulation-process parameter space of nanofiber membranes is highly dimensional and involves complex interactions, making optimization difficult. This system uses a physical interaction heterogeneity graph to structurally model the complex physical system and leverages an upper confidence bound acquisition function to balance exploration and utilization, efficiently performing a global search within the vast parameter space. This not only quickly locates the process parameter combinations that meet the requirements but also has the potential to discover superior, non-explicit process windows that transcend human intuition.

[0029] 4. This invention possesses powerful adaptive and self-learning capabilities, ensuring the system's long-term effectiveness. Through the model closed-loop adaptive correction unit, the system can continuously utilize the latest real experimental data to incrementally and automatically fine-tune the prediction model. This makes it not only a static prediction tool but also a digital twin system that can interact with the physical world and continuously evolve. This capability ensures that the model maintains high accuracy when facing new material systems, environmental changes, or equipment aging, exhibiting strong robustness and industrial applicability. Attached Figure Description

[0030] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0031] Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0033] Example 1:

[0034] Please see Figure 1 A digital twin-based nanofiber membrane production quality prediction and process optimization system includes:

[0035] Multiscale property gene coding units are used to collect material parameters and process parameters in the nanofiber membrane preparation process, and to perform structuring on the material parameters and process parameters in order to construct multiscale property genes that characterize physicochemical properties.

[0036] Physical interaction heterogeneity graph construction unit, used to construct physical interaction heterogeneity graphs for specific formulation and process combinations based on multi-scale physical property genes;

[0037] The physical grammar graph neural network prediction unit is used to receive the physical interaction heterogeneous graph and perform message passing and node update processing to generate the final node feature vector containing global physical interaction information. Then, the final node feature vector is decoded to generate the macroscopic quality prediction index of the nanofiber membrane and the corresponding prediction uncertainty assessment.

[0038] The active learning parameter optimization unit is used to calculate the optimal combination of parameters for the next experimental point based on macroscopic quality prediction indicators and prediction uncertainty assessment.

[0039] The model closed-loop adaptive correction unit is used to obtain the real experimental data corresponding to the optimal next experimental point parameter combination, and to perform feedback fine-tuning of the physical grammar graph neural network prediction unit based on the real experimental data.

[0040] This embodiment provides a digital twin-based nanofiber membrane production quality prediction and process optimization system. The system aims to solve the technical pain points in the existing nanofiber membrane preparation process, which are characterized by high spatial dimensions of formula-process parameters and complex physical processes, resulting in long R&D cycles, high costs, and difficulties in quality control. The system constructs a closed-loop intelligent system from microscopic physical genes to macroscopic quality prediction and then to experimental parameter optimization, thereby achieving accurate prediction and efficient optimization of the nanofiber membrane manufacturing process.

[0041] In this embodiment, the system comprises five core units that work together: a multi-scale property gene encoding unit, a physical interaction heterogeneous graph construction unit, a physical grammar graph neural network prediction unit, an active learning parameter optimization unit, and a model closed-loop adaptive correction unit. These five units together form a complete technical closed loop from data input, physical modeling, quality prediction, parameter optimization to model adaptive correction.

[0042] The multi-scale property gene encoding unit transforms the unstructured, multi-source heterogeneous material and process parameters in the nanofiber membrane preparation process into a structured digital object containing physicochemical meaning, providing high-quality input for subsequent physical modeling. In this embodiment, the unit collects the material and process parameters required for the nanofiber membrane preparation process and performs systematic structuring on these parameters, ultimately constructing a multi-scale property gene that can comprehensively characterize the physicochemical properties of each element. Here, the multi-scale property gene is a multi-level digital feature vector that provides a precise and calculable digital identity for each material component or process parameter. It represents a deep analysis and encoding of the physicochemical properties of materials and processes at multiple scales, including molecular, solution, and process response.

[0043] The physical interaction heterogeneous graph construction unit organizes discrete material property genes into a global network structure that can express the physical causal relationships between various elements. In this embodiment, based on the multi-scale material property genes generated in the preceding steps, this unit dynamically constructs a physical interaction heterogeneous graph for each specific formulation-process combination. The physical interaction heterogeneous graph here is a special type of graph structure data, where nodes represent material or process parameters and edges represent the physicochemical interactions between them. It transforms complex physical systems into mathematical objects that can be processed by graph neural networks. The initial characteristics of the nodes are defined according to the parameters in the multi-scale material property genes, and the types and weights of the edges between the nodes are calculated.

[0044] The physical grammar graph neural network prediction unit simulates the core physical laws in the nanofiber membrane preparation process, processes the constructed heterogeneous graph, and thus accurately predicts the macroscopic quality of the nanofiber membrane. In this embodiment, the unit receives the physical interaction heterogeneous graph as input and simulates key physical processes such as solution stretching and charge conduction through a specially designed message passing and node update mechanism to generate a final node feature vector containing global physical interaction information. The decoding module processes this feature vector and finally outputs the predicted value of the macroscopic quality of the nanofiber membrane, while simultaneously providing a prediction uncertainty.

[0045] The uncertainty assessment here is a quantitative evaluation of the model's confidence in its prediction results. Its role is to guide the subsequent parameter optimization process, enabling it to explore the region where the model's prediction is most uncertain. This is a characteristic of Bayesian optimization theory or neural network models themselves.

[0046] The active learning parameter optimization unit aims to efficiently find the optimal combination of formulation and process parameters in a vast parameter space with the fewest number of experiments, based on the results of the prediction model. In this embodiment, the unit calculates the optimal next experimental point parameter combination that maximizes the function value based on the macroscopic quality prediction index and prediction uncertainty assessment output from the previous steps, using a preset upper confidence bound acquisition function. Here, the optimal next experimental point parameter combination is a specific formulation and process parameter setting that is recommended for immediate real-world experiments. Its purpose is to efficiently guide experimental design and avoid blind trial and error. It is the result of calculation by active learning and Bayesian optimization algorithms, and the result aims to balance the use of validation in the currently predicted optimal region and exploration in the region where the model is most uncertain.

[0047] The model closed-loop adaptive correction unit's function is to continuously optimize and iterate the prediction model using new real experimental data, ensuring the accuracy and long-term effectiveness of the system's predictions. In this embodiment, the unit acquires experimental data generated in a real environment based on the optimal next experimental point parameter combination. Based on the difference between the real data and the model's predicted values, the model parameters of the physical grammar graph neural network prediction unit are fine-tuned. This process forms a complete closed loop of prediction-optimization-experiment-correction, enabling the system to have self-learning and adaptive capabilities.

[0048] This embodiment constructs a complete, closed-loop intelligent optimization system through the collaborative work of the above five units. Compared with the traditional R&D model that relies on expert experience or a lot of trial and error, this invention can significantly shorten the new product development cycle, reduce experimental costs, and discover better combinations of process parameters that surpass human experience. By embedding physical laws into neural networks, the model not only has high prediction accuracy but also better interpretability of physical mechanisms, solving the pain points of traditional black box models and realizing efficient, accurate, and interpretable prediction and optimization of nanofiber membrane preparation processes.

[0049] Example 2:

[0050] The process of structuring multi-scale physical property gene coding units is as follows:

[0051] The molecular-scale chemical properties of the material are analyzed to determine the molecular weight descriptor, chain structure information, and functional group information; the macroscopic physicochemical properties of the solution-scale layer are analyzed to determine the Hansen solubility parameter, intrinsic viscosity, and dielectric constant; and the process response layer characteristics of the material under the process environment are analyzed to determine the model coefficients of polarizability and solvent evaporation rate under an electric field.

[0052] Based on Example 1, this embodiment specifies the structured processing of the multi-scale property gene coding unit to ensure that the encoded multi-scale property genes can comprehensively and deeply capture all key physicochemical information affecting the nanofiber membrane preparation process, laying a solid foundation for constructing a high-fidelity physical interaction isomer diagram and a high-precision prediction model.

[0053] In this embodiment, the process of structuring the multi-scale physical property gene coding unit is precisely divided into three hierarchical analysis steps: analyzing the molecular-scale chemical properties of the material. This level aims to capture the intrinsic characteristics of the material from the most basic chemical structure. Specifically, it involves determining the molecular weight descriptor, chain structure information, and functional group information of the polymer.

[0054] The macroscopic physicochemical properties of the solution-scale layer are analyzed. This layer aims to characterize the macroscopic properties exhibited in solution state, which are determined by molecular-scale attributes. Specifically, this involves determining the Hansen solubility parameter, intrinsic viscosity, and dielectric constant of the solution. The process response layer characteristics of the material under the process environment are analyzed. This layer aims to describe the dynamic response behavior of the material system under the actual spinning electric field and solvent evaporation environment. Specifically, this involves determining the polarizability of the system under the electric field and the model coefficients of the solvent evaporation rate.

[0055] Through this multi-scale, hierarchical encoding method from molecules and solutions to process responses, the property genes constructed in this invention are no longer simple parameter listings, but form a structured knowledge system containing inherent physical logic. This refined feature engineering enables subsequent graph neural network models to understand and learn the complex mapping relationship between formulation, process, structure, and performance from a more fundamental physical level, thereby improving the model's prediction accuracy, generalization ability, and physical interpretability.

[0056] Example 3:

[0057] The construction process of the physical interaction heterogeneous graph is as follows:

[0058] Material components or process parameters are set as nodes, and the initial feature vector is generated by multi-scale property gene encoding; the physicochemical interactions between nodes are set as edges, and the type and weight of the edges are calculated based on the parameters in the multi-scale property gene of the corresponding node.

[0059] Based on Example 1, this embodiment further specifies the construction process of the physical interaction heterogeneous graph. The purpose of this is to organically organize the discrete property genes encoded in the previous step into a network topology that can explicitly express the physical and chemical interactions between elements in the system, thereby injecting domain knowledge into the model in a structured manner.

[0060] In this embodiment, the construction process of the physical interaction heterogeneous graph follows the following explicit rules: material components or process parameters are set as nodes, and each node in the graph represents a specific physical entity or process condition. Its initial feature vector is generated by multi-scale property gene encoding, which means that the initial state of each node contains rich physical information at multiple scales; the physical and chemical interactions between nodes are set as edges, and each edge in the graph represents a known physical or chemical relationship between nodes. The type and weight of the edge are calculated based on the parameters in the multi-scale property gene of the corresponding node.

[0061] For example, the weight of the edge between the polymer node and the solvent node can be calculated by the Euclidean distance of their Hansen solubility parameters to quantitatively characterize the intensity of the dissolution effect; the weight of the edge between the salt ion additive node and the polymer node can be calculated based on the concentration and charge number of the salt ions to characterize their contribution to the overall charge conduction capacity of the solution.

[0062] In addition, parameters that affect the global process environment, such as ambient temperature and relative humidity, can also be set as special global context nodes. Their feature vectors will be used to adjust the interactions related to solvent evaporation or charge conduction during message passing, so that the model can adapt to different production environments.

[0063] Unlike traditional graph neural networks that treat all nodes and edges as homogeneous, the heterogeneous graph constructed in this embodiment directly models the physical causal relationships of the nanofiber membrane production system through its topology and weight distribution. This design hardcodes domain knowledge into the model input, reducing the model's dependence on massive training data and making the model's learning process more focused on key physical interactions, thereby improving learning efficiency and the accuracy of the final prediction.

[0064] Example 4:

[0065] The process of message passing and node update handling in the prediction unit of the physical grammar graph neural network is as follows:

[0066] The message passed from the neighboring node to the central node is defined as a weighted sum of multiple physical action terms; the physical action terms are constructed by modulating the interaction strength between node feature vectors by using physical parameters derived from multi-scale physical property genes as adjustment factors; all messages received from the neighboring nodes are aggregated and combined with the state of the node itself in the previous layer to generate an updated node state.

[0067] The physical parameters include the normalized intrinsic viscosity of the solution, which is related to viscoelasticity, and the normalized dielectric constant of the solution, which is related to charge conduction.

[0068] Based on Example 1, this embodiment specifies the message passing and node update processing of the physical grammar graph neural network prediction unit, as well as the key physical parameters therein. The core innovation of this design is that instead of using general graph neural network operators that are independent of specific problems, it creatively designs a physical grammar that directly embeds the most important physical laws in the nanofiber membrane preparation process, such as viscoelastic effect and charge conduction effect, into the operation structure of the neural network.

[0069] In this embodiment, the core process of message passing and node updating by the physical grammar graph neural network prediction unit is as follows;

[0070] During the message construction phase, the message passed from neighbor node j to central node i is defined as a weighted sum of multiple physical action terms. The construction of these physical action terms modulates the interaction strength between node feature vectors by using physical parameters derived from multi-scale property genes as adjustment factors. A non-restrictive message construction formula is as follows:

[0071] ;

[0072] in, : represents the message vector passed from neighbor node j to central node i in the l-th layer network. It is a numerical vector, and its value is calculated by this formula.

[0073] and : These are the feature vectors of nodes i and j in layer l, respectively, representing the digital state of the nodes at this stage. They are numerical vectors and are the output of the previous layer network or the initial node features.

[0074] and : is the core regulatory factor for realizing physical grammars; it is not a learnable parameter, but a scalar with clear physical meaning directly derived from the physical property gene; among which, The physical parameter represented is the normalized intrinsic viscosity of the solution, which is related to viscoelasticity and is used to adjust the interaction strength related to solution stretching and deformation. The physical parameter represented is the normalized solution dielectric constant, which is related to charge conduction and is used to adjust the interaction strength related to electric field action and charge migration.

[0075] Etc.: These are weight matrices learned by the model during training through optimization algorithms, used for linear transformation and aggregation of feature vectors;

[0076] : is the hyperbolic tangent activation function, used to introduce nonlinearity and ensure the stability of numerical computation;

[0077] During the node update phase, all messages received from neighboring nodes are aggregated and combined with the state of the node's own parent node to generate the updated node state. The update rules are as follows:

[0078] ;

[0079] in, It is the set of all neighboring nodes of node i. It is a non-linear activation function. It is a learnable weight matrix that acts on the features of the node itself;

[0080] This design, which uses physical parameters as interaction intensity modulation factors, forces the neural network to follow preset physical laws, making the model physically self-consistent. It can learn the correct physical dependencies with only a small amount of data, improving the model's prediction accuracy, stability, and generalization ability. Compared with traditional black-box models, the prediction results of this invention are not only accurate, but its internal computation process also has stronger interpretability due to the inclusion of physical grammar.

[0081] Example 5:

[0082] The decoding process of the prediction unit in the physical grammar graph neural network is as follows:

[0083] Using a preset aggregation function, the final node feature vectors of all nodes are fused into a single vector representing the entire graph; through a readout function, the single vector representing the entire graph is mapped to the macroscopic quality prediction index of the nanofiber membrane; wherein, the macroscopic quality prediction index includes the average fiber diameter and porosity.

[0084] Based on Example 1, this embodiment specifically defines the decoding process of the physical grammar graph neural network prediction unit. The purpose of this is to effectively convert the microscopic physical interaction information contained in each node of the graph after passing through multiple layers of physical grammar messages into a direct and quantitative prediction of the macroscopic quality indicators of the final product.

[0085] After L layers of message passing and node updates, each node i in the graph obtains a final node feature vector. The decoding process of the Graph Prediction Unit (GNU) in a physical grammar includes the following two key steps: First, a pre-defined aggregation function is used to fuse the final node feature vectors of all nodes into a single vector representing the entire graph. This aggregation function, AGG, can be a summation function, a mean function, or a learnable attention mechanism. Its mathematical expression is as follows:

[0086] ;

[0087] Where V is the set of all nodes in the graph. It is a single vector representing the entire graph;

[0088] The readout function maps a single vector representing the entire image to a macroscopic quality prediction index for the nanofiber membrane; this readout function R is typically a multilayer perceptron that receives the global feature vector. As input, it outputs one or more specific macroscopic quality prediction values; its mathematical expression is:

[0089] ;

[0090] Where y is the predicted quality index vector; in this embodiment, the macroscopic quality prediction index is specifically defined as including at least the average fiber diameter. With porosity ;

[0091] Through the two-step decoding process of aggregation and readout, this invention successfully transforms the microscopic physical interaction patterns learned by the graph neural network into measurable macroscopic performance indicators with direct guiding significance. Specifically, the output indicators are defined as the average fiber diameter and porosity, ensuring that the prediction results of this invention are closely integrated with actual production needs and have extremely high industrial application value.

[0092] It should be noted that although this embodiment limits the macroscopic quality prediction indicators to the average fiber diameter and porosity, the framework of this invention has good scalability; through the readout function By adding output dimensions and providing corresponding real experimental data for training, this invention can also predict and optimize other key quality indicators such as fiber diameter distribution standard deviation, bead defect density, and membrane tensile strength, thereby achieving a more comprehensive evaluation and control of product quality.

[0093] Example 6:

[0094] The process by which the active learning parameter optimization unit calculates the optimal parameter combination for the next experimental point is as follows:

[0095] Define a scalarized comprehensive objective function based on preset requirements; input the comprehensive objective score generated by the macro quality prediction index and the prediction uncertainty assessment into the preset upper confidence bound acquisition function for calculation, and determine the parameter combination that maximizes the calculation result as the optimal parameter combination for the next experimental point;

[0096] Based on Example 1, this embodiment specifically defines the process by which the active learning parameter optimization unit calculates the optimal combination of parameters for the next experimental point. The aim is to realize a data-driven, efficient search strategy in the formula-process parameter space, so as to quickly locate the best parameter combination that meets the preset requirements with minimal experimental cost.

[0097] In this embodiment, the process by which the active learning parameter optimization unit calculates the optimal combination of parameters for the next experimental point is based on the Bayesian optimization theory framework. It intelligently selects the next exploration point using an upper confidence bound acquisition function. The process is as follows: A scalarized comprehensive objective function is defined according to preset requirements. This transforms multidimensional performance metrics into a scalar score; for example, an unrestricted comprehensive objective function can be defined as a weighted sum of expected metrics, such as... ,in, It is the target fiber diameter. and These are the predicted fiber diameter and porosity, respectively. and These are non-negative weighting coefficients set according to actual needs; the negative sign here indicates that the goal is to minimize the difference between the diameter and the target.

[0098] The comprehensive target score generated from the macroeconomic quality forecasting indicators, along with the forecast uncertainty assessment, is input into a pre-defined upper confidence bound acquisition function for calculation. This function aims to strike a balance between utilization and exploration; utilization is embodied in the function. The term refers to the comprehensive objective score predicted by the surrogate model for the parameter combination x; the exploration is reflected in the function. This refers to the uncertainty assessment of the model's prediction of the comprehensive objective score; and the parameter combination that maximizes the calculation result is determined as the optimal parameter combination for the next experimental point; the optimal next experimental point This is obtained by solving the following maximization problem:

[0099] ;

[0100] in, This refers to the combination of parameters that will be determined as the optimal next experimental point.

[0101] S: Represents the search space for all possible formulation-process parameters;

[0102] The prediction result of the physical grammar graph neural network prediction unit for the parameter combination x, after being transformed by the comprehensive objective function, is calculated by the preceding prediction unit.

[0103] Prediction unit pair The uncertainty of the predicted value is output along with the preceding prediction unit;

[0104] κ is a non-negative hyperparameter used to balance exploitation and exploration. Its value is usually selected in the range of [0.1, 3]. In the early stages of optimization, to encourage more extensive exploration, a larger κ value can be set, such as κ = 2.58, corresponding to a 99% confidence interval. As optimization progresses and high-value regions are discovered, the κ value can be appropriately reduced to enhance local optimization.

[0105] By employing an upper confidence bound acquisition function, this invention avoids the enormous computational load and experimental costs of traditional optimization methods, and also overcomes the defect that simple greedy algorithms are prone to getting trapped in local optima. This method can intelligently and dynamically plan experimental paths, and efficiently and reliably find the globally optimal or near-optimal combination of process parameters with far fewer experiments than traditional methods.

[0106] Example 7:

[0107] The process of feedback fine-tuning by the model's closed-loop adaptive correction unit is as follows:

[0108] The optimal combination of parameters for the next experimental point is combined with real experimental data to form a new data point; the new data point is added incrementally to the training dataset; the stochastic gradient descent algorithm is used to calculate the gradient of the loss function between the current model's predicted output for the new experimental point and the real experimental data; the model parameters of the prediction unit of the physical grammar graph neural network are updated based on the gradient of the loss function and the preset learning rate.

[0109] Based on Example 1, this embodiment specifies the feedback fine-tuning process of the model closed-loop adaptive correction unit, aiming to build a dynamic and continuously evolving system to ensure that the prediction model can continuously optimize itself using newly acquired real experimental data.

[0110] Recommending the optimal parameter combination for the next experimental point in the preceding steps. After conducting a real experiment, corresponding real experimental data will be obtained. The feedback fine-tuning process of the model's closed-loop adaptive correction unit utilizes this new data to... This drives the model iteration; the process is as follows: combine the optimal next experimental point parameter combination with the real experimental data to form a new data point, and add the incremental data point to the training dataset;

[0111] The stochastic gradient descent algorithm is used to calculate the gradient of the loss function between the current model's predicted output for the new experimental point and the actual experimental data. To do this, a loss function L is first defined to quantify the current model's predicted output. Compared with real experimental data The difference between them; a commonly used loss function is the mean squared error, which has the form of Based on the gradient of the loss function and the preset learning rate, the model parameters of the prediction unit of the physical grammar graph neural network are updated; the parameter update follows the optimization rules of stochastic gradient descent, and the update formula is as follows:

[0112] ;

[0113] in, : Represents the set of all learnable parameters of the model before this update;

[0114] This is the updated set of model parameters after this fine-tuning.

[0115] : This is the learning rate, a preset hyperparameter used to control the step size of each parameter update. Its initial value is usually in the range of 0. to The range is selected; to ensure stable convergence of the model, a learning rate decay strategy can be adopted, such as multiplying the learning rate by a coefficient less than 1 after every few iterations.

[0116] : is the loss function with respect to the current parameters The gradient vector, which indicates the direction in which the loss function decreases the fastest, is calculated using the backpropagation algorithm;

[0117] Through this feedback fine-tuning mechanism, the present invention constructs a complete intelligent closed loop of prediction-optimization-experimentation-correction; the system is no longer a static model, but a digital twin system that can continuously interact with the physical world, learn and evolve; this adaptive correction capability ensures that the system can maintain high performance when facing new materials, new processes or environmental changes, and has strong robustness and industrial applicability.

[0118] To ensure the stability and reliability of the model, robustness testing can be performed on the trained physical grammar graph neural network prediction unit before deployment. This testing includes inputting physically extreme combinations of parameters into the model that exceed the normal training range, such as zero voltage, zero polymer concentration, extremely high or extremely low solution viscosity, and verifying whether the model's output conforms to physical common sense, such as predicting extremely large fiber diameters or states that cannot be formed. At the same time, the sensitivity of the model's predictions is tested by introducing input perturbations. Through these stress tests, the generalization ability and predictive behavior of the model in practical applications can be ensured.

[0119] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A digital twin-based system for predicting the production quality and optimizing the process of nanofiber membranes, characterized in that, include: Multi-scale property gene coding units are used to collect material and process parameters from the nanofiber membrane preparation process, and to structure these parameters to construct multi-scale property genes characterizing physicochemical properties. The structuring process is as follows: The molecular-scale chemical properties of the material are analyzed to determine the molecular weight descriptor, chain structure information, and functional group information; the macroscopic physicochemical properties of the solution-scale layer are analyzed to determine the Hansen solubility parameter, intrinsic viscosity, and dielectric constant; and the process response layer characteristics of the material under the process environment are analyzed to determine the model coefficients of polarizability and solvent evaporation rate under an electric field. The physical interaction heterogeneity graph construction unit is used to construct physical interaction heterogeneities for specific formulation and process combinations based on multi-scale physical property genes; the construction process is as follows: Material components or process parameters are set as nodes, and the initial feature vector is generated by multi-scale property gene encoding; the physicochemical interactions between nodes are set as edges, and the type and weight of the edges are calculated based on the parameters in the multi-scale property gene of the corresponding node. The physical grammar graph neural network prediction unit is used to receive the physical interaction heterogeneous graph and perform message passing and node update processing to generate the final node feature vector containing global physical interaction information. Then, the final node feature vector is decoded to generate the macroscopic quality prediction index of the nanofiber membrane and the corresponding prediction uncertainty assessment. The active learning parameter optimization unit is used to calculate the optimal combination of parameters for the next experimental point based on macroscopic quality prediction indicators and prediction uncertainty assessment. The model closed-loop adaptive correction unit is used to obtain the real experimental data corresponding to the optimal next experimental point parameter combination, and to perform feedback fine-tuning of the physical grammar graph neural network prediction unit based on the real experimental data.

2. The nanofiber membrane production quality prediction and process optimization system based on digital twins according to claim 1, characterized in that, The process of message passing and node update handling in the prediction unit of the physical grammar graph neural network is as follows: The message passed from the neighboring node to the central node is defined as a weighted sum of multiple physical action terms; the physical action terms are constructed by modulating the interaction strength between node feature vectors by using physical parameters derived from multi-scale physical property genes as adjustment factors; all messages received from the neighboring nodes are aggregated and combined with the state of the node itself in the previous layer to generate an updated node state.

3. The nanofiber membrane production quality prediction and process optimization system based on digital twins according to claim 2, characterized in that, The physical parameters include the normalized intrinsic viscosity of the solution, which is related to viscoelasticity, and the normalized dielectric constant of the solution, which is related to charge conduction.

4. The nanofiber membrane production quality prediction and process optimization system based on digital twins according to claim 1, characterized in that, The decoding process of the prediction unit in the physical grammar graph neural network is as follows: Using a preset aggregation function, the final node feature vectors of all nodes are fused into a single vector representing the entire graph. Through a readout function, the single vector representing the entire graph is mapped to the macroscopic quality prediction index of the nanofiber membrane. The macroscopic quality prediction index includes the average fiber diameter and porosity.

5. The nanofiber membrane production quality prediction and process optimization system based on digital twins according to claim 1, characterized in that, The process by which the active learning parameter optimization unit calculates the optimal parameter combination for the next experimental point is as follows: Define a scalarized comprehensive objective function based on preset requirements; input the comprehensive objective score generated by the macro quality prediction index and the prediction uncertainty assessment into the preset upper confidence bound acquisition function for calculation, and determine the parameter combination that maximizes the calculation result as the optimal parameter combination for the next experimental point.

6. The nanofiber membrane production quality prediction and process optimization system based on digital twins according to claim 1, characterized in that, The process of feedback fine-tuning by the model's closed-loop adaptive correction unit is as follows: The optimal combination of parameters for the next experimental point is combined with real experimental data to form a new data point; the new data point is added incrementally to the training dataset; the stochastic gradient descent algorithm is used to calculate the gradient of the loss function between the current model's predicted output for the new experimental point and the real experimental data; and the model parameters of the prediction unit of the physical grammar graph neural network are updated based on the gradient of the loss function and the preset learning rate.

Images