A smart manufacturing method and system for monodisperse polymer-based chromatographic microspheres

Abstract

This invention discloses an intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres, relating to the field of intelligent manufacturing. It utilizes a machine learning model to establish a quantitative mapping relationship between chromatographic microsphere process parameters and product microsphere performance parameters. Based on the user-inputted target microsphere performance, it inversely predicts and outputs the optimal process parameters. This, combined with a high-shear emulsification polymerization execution module, precisely completes the entire process of emulsification, polymerization, and functionalization, achieving customized monodisperse polymer-based chromatographic microspheres with the target microsphere performance. This invention realizes a process shift from experience-based trial and error to model prediction-driven processes for chromatographic microspheres, overcoming the challenge of precise control of multiple parameters. The resulting product microspheres have a narrow particle size distribution, controllable structure, and good batch consistency, making them suitable for high-performance chromatographic columns.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology, specifically relating to an intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres that deeply integrates high-shear emulsification technology with machine learning algorithms. Background Technology

[0002] Chromatographic separation technology is a core analytical tool in fields such as biomedicine, food safety, and environmental monitoring. Its separation efficiency is highly dependent on the chromatographic microspheres packed in the chromatographic column. High-performance chromatographic microspheres, especially polymer-based chromatographic microspheres used in analytical chromatography, are considered the "core" of high-end chromatographic separation. Their key structural parameters, including particle size and its distribution, pore structure, specific surface area, and surface chemical properties, directly determine the separation efficiency, selectivity, sample loading capacity, and stability of the chromatographic column. Therefore, achieving the large-scale, controllable preparation of polymer microspheres with highly uniform particle size, precisely adjustable pore size, and clean surface is crucial to enhancing the competitiveness of the high-end chromatography industry.

[0003] While traditional suspension polymerization and emulsion polymerization processes for preparing polymer-based chromatographic microspheres are relatively mature, their inherent thermodynamic and kinetic instabilities lead to wide product particle size distributions and poor batch-to-batch reproducibility. Furthermore, these methods require large amounts of stabilizers, emulsifiers, and other additives, whose residues are difficult to remove completely, contaminating the microsphere surface and affecting column separation performance. This can even cause column bed collapse, severely shortening column lifespan. In recent years, microfluidics and membrane emulsification technologies have emerged as new methods for preparing polymer microspheres, demonstrating high precision in droplet generation and particle size control, enabling their fabrication. However, these technologies still have significant drawbacks, such as high equipment costs, low throughput, susceptibility to channel blockage, difficulty in process scale-up, and limited applicability to continuous industrial production.

[0004] For example, US Patent 2015 / 053093 discloses a system and method for controlling polymer reactions and processing using automated continuous online monitoring, but this method is not suitable for the precise control of polymer-based chromatographic microspheres. Chinese Patent CN108704578B addresses the problem of poor production continuity in polymer-based chromatographic microsphere production equipment by designing a continuous microsphere production device, utilizing an emulsification module with injection head arrays and baffles, and a constant flow pump system; however, the uniformity of the microspheres produced is poor, making them unsuitable for chromatographic analysis, and the structural characteristics of the microspheres cannot be precisely controlled during the reaction process. Chinese Patent Application CN115662544A discloses a method for synthesizing polystyrene microspheres with controllable particle size based on interpretable machine learning, but this method does not control the polymerization process, and the product cannot be used for chromatographic analysis.

[0005] High-speed shear mixers, as highly efficient continuous mixing and dispersing devices, generate extremely strong shear fields through rotor-stator systems, enabling efficient breakup and homogenization of dispersed phases into micron / submicron droplets. They exhibit stability even in high-temperature, high-viscosity systems, providing a highly promising engineering approach for the continuous and controllable synthesis of polymer-based chromatographic microspheres. However, current processes based on high-speed shear mixers are mostly based on non-closed-loop, experience-based operations. The core problem lies in the lack of systematic understanding and mathematical models of the complex quantitative relationships between process parameters, microsphere structural characteristics, and chromatographic performance. Synthetic process development and optimization heavily rely on operator experience and repeated trial and error, making it impossible to deduce and precisely set optimal process parameters from preset chromatographic performance targets, thus hindering the precise design and manufacturing of high-performance microspheres.

[0006] While machine learning technology has demonstrated its powerful capabilities in modeling and multi-objective optimization of chemical processes, enabling the uncovering of complex nonlinear relationships in high-dimensional data, its deep integration into the microstructure formation process of chemical polymerization to construct an intelligent chromatographic microsphere preparation system that combines real-time sensing, intelligent decision-making, and closed-loop control remains a technological gap. Existing technologies, whether traditional chemical methods or emerging physical devices, have failed to systematically solve the challenges across the entire chain, from performance-oriented design to precise process control and ensuring consistent product quality.

[0007] Therefore, given the prominent problems of existing chromatographic microsphere preparation technologies, such as poor particle size distribution control, poor batch stability, disconnect between process and performance, excessive reliance on human experience, and difficulty in large-scale continuous production of high-quality products, there is an urgent need for an innovative system solution. This solution can deeply integrate efficient physical preparation methods with advanced intelligent optimization algorithms, breaking through the gap from "experience-based trial and error" to "model prediction and precise control," thereby achieving reliable, continuous, and intelligent manufacturing of high-performance polymer-based chromatographic microspheres. Summary of the Invention

[0008] 1. The problem to be solved In the research of this invention, the applicant found that traditional processes for preparing polymer-based chromatographic microspheres suffer from problems such as wide product particle size distribution, poor batch-to-batch repeatability, and numerous residues affecting usability. Emerging microfluidic and membrane emulsification technologies can achieve high-precision control of the particle size of monodisperse microspheres in a continuous manner, but these processes suffer from problems such as high equipment costs, low throughput, easy channel blockage, and difficulty in industrial continuous production. Furthermore, both traditional and emerging technologies lack a quantitative understanding of the relationship between process parameters, microsphere structural characteristics, and chromatographic performance. Preparation relies on operational experience and repeated trial and error, resulting in poor reliability, high efficiency and cost, and an inability to meet the needs of technological development.

[0009] Therefore, addressing the problems of wide particle size distribution, poor batch consistency, lack of quantitative mapping between process and performance, and over-reliance on human experience in existing polymer-based chromatographic microsphere preparation technologies, a systematic solution that integrates efficient physical preparation methods with intelligent algorithm optimization capabilities is urgently needed. This paper proposes an intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres that deeply integrates high-shear emulsification technology with machine learning algorithms. By constructing an "execution-perception-decision" intelligent route, it achieves reverse precision design from performance targets to process parameters. This not only significantly improves the monodispersity and structural controllability of the microspheres but also provides a reliable technical path for their industrial continuous production, possessing significant scientific and engineering application value.

[0010] 2. Technical Solution To solve the above problems, the technical solution adopted by the present invention is as follows: In a first aspect, this invention provides a smart manufacturing method for monodisperse polymer-based chromatographic microspheres, comprising: Historical synthesis data is acquired and a historical database is established. The historical synthesis data consists of process parameter data and microsphere structural characteristic data corresponding to the preparation of several monodisperse polymer-based chromatographic microspheres by the high-shear emulsification polymerization execution module. The process parameter data includes high-shear emulsification speed, shear time, mass ratio of polymer monomer to porogen, reaction temperature program, concentration of functionalizing reagent and reaction time. The microsphere structural characteristic data includes the average particle size, particle size variation coefficient, average pore size, specific surface area and surface functional group density of the microspheres. Based on historical databases, an ensemble learning model based on machine learning is trained; wherein the input of the ensemble learning model is the process parameter data of the chromatographic microspheres, and the output is the microsphere structural feature data corresponding to the process parameter data; Obtain the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized, perform reverse reasoning calculations based on the integrated learning model, and output the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized. Based on the theoretical process parameters, the high-shear emulsification polymerization execution module is used to perform the synthesis steps to synthesize monodisperse polymer-based chromatographic microspheres; wherein, the high-shear emulsification polymerization execution module includes at least one high-shear emulsification reactor; The integrated learning model is a polymer-based chromatographic microsphere process-structure mapping network, which includes a feature embedding layer, a base learner group layer, a dynamic weight fusion layer and a microsphere structure feature output layer connected in series, as well as an uncertainty estimation module connected in parallel with the base learner group layer. The feature embedding layer is used to normalize continuous variables in the input process parameter data, perform temporal encoding of the reaction temperature program, perform embedding mapping on categorical variables, and generate explicit feature crosses, outputting a dense vector. The base learner group layer includes several base learner sub-models, each of which predicts a microsphere structure feature. Each sub-model shares the underlying feature representation and introduces physical prior constraint penalties during training. The uncertainty estimation module is used to output the predicted mean and 95% confidence interval of each microsphere structure feature based on Monte Carlo Dropout. The dynamic weight fusion layer includes a gating network, which adaptively adjusts the fusion weights output by each sub-model according to the input process parameter data. The microsphere structure feature output layer is used to output the microsphere structure feature data corresponding to the process parameter data.

[0011] This invention combines the excellent physical regulation capability of the high-shear emulsification polymerization execution module with the intelligent prediction capability of machine learning, realizing the reverse intelligent design of chromatographic microspheres from "performance target" to "process parameters", and overcoming the problem of multi-parameter synergistic and precise control of polymer-based chromatographic microspheres.

[0012] Furthermore, the feature embedding layer's time-series encoding of the reaction temperature program includes: The reaction temperature program is split into a heating rate sequence, a holding time sequence, and a target temperature sequence. A one-dimensional convolutional layer is used to extract the local temporal patterns of the heating rate sequence, the holding time sequence, and the target temperature sequence, and then global average pooling is used to extract the temperature feature vector.

[0013] Furthermore, the ensemble learning model employs a multi-task loss function to optimize the target prediction error during training; the calculation formula for the multi-task loss function is as follows: L total = Σ k ( α k * MSE k )+ λ physics * L physics + λ reg * L reg In the formula, L total Represents the total loss function; k Indicates the first k Individual microsphere structural features; MSE kIndicates the first k Mean square prediction error of individual microsphere structural features; α k For the first k The fusion weights of individual microsphere structural features are determined based on the historical variation coefficients of the microsphere structural features. CV k set up, α k = 1 / CV k ; L physics As a physical constraint penalty term, calculate the degree of violation of physical trends; L reg This is an L2 regularization term to prevent overfitting; λ physics , λ reg These are hyperparameters, with values ​​of 0.1 and 0.001 respectively.

[0014] Furthermore, the process of obtaining the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized, performing reverse reasoning calculations based on the ensemble learning model, and outputting the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized is as follows: Based on the uncertainty estimation module, a coarse-grained scan of the overall process parameter space is performed, and contour lines of the predicted standard deviation of each microsphere structural feature are drawn. Based on the predicted standard deviation contour lines, the range of process parameters with predicted standard deviations lower than a preset threshold is selected as the feasible region. Within the feasible region, an optimizer combining Bayesian optimization and covariance matrix adaptive evolution strategy is employed. The objective is to minimize the weighted Euclidean distance between the predicted microsphere structural feature data and the target microsphere structural feature data, and to search for the optimal process parameters, which are the theoretical process parameters. Specifically, the optimizer uses a Gaussian process surrogate model and an expected improvement acquisition function for iteration during the Bayesian optimization phase. When the Bayesian optimization converges and stalls or reaches the maximum number of iterations, it switches to the covariance matrix adaptive evolution strategy for fine-grained search. Output theoretical process parameters and their corresponding prediction confidence intervals and uncertainty scores to guide production.

[0015] Furthermore, the actual microsphere structural feature data of the product microspheres synthesized by the high-shear emulsification polymerization execution module is collected and fed back to the ensemble learning model so that the ensemble learning model can be optimized in conjunction with the actual microsphere structural feature data. This solution provides closed-loop feedback, comparing the actual performance data of the product microspheres obtained according to reverse intelligent design and production with the target performance data, and feeding this data back to the ensemble learning model as new samples for incremental learning and parameter optimization. This guides the synthesis of subsequent batches, achieving continuous optimization of the model in the preparation process.

[0016] Furthermore, the intelligent manufacturing method further includes: collecting dynamic data of the reaction process of the high-shear emulsification polymerization execution module, and dynamically fine-tuning the high-shear emulsification rotation speed and shear time based on the dynamic data; wherein the dynamic data is the droplet size of the emulsion after high-shear emulsification. The solution dynamically fine-tunes the shear rotation speed and shear time by monitoring the droplet size data in real time to ensure that the emulsion droplet size distribution meets a preset target.

[0017] Preferably, the monodisperse polymer-based chromatographic microspheres prepared by the intelligent manufacturing method have the following customizable microsphere structural features, including: an average particle size of 1-20 μm; a particle size variation coefficient of less than 5%, preferably less than 3%; an average pore size of 5-100 nm; a specific surface area of ​​10-800 m² / g; and a surface functional group density of 0.1-2.0 mmol / g, wherein the surface functional groups include one or more of tertiary amine groups, sulfonic acid groups, carboxyl groups, hydroxyl groups, C18 alkyl chains, and phenyl groups. Preferably, the average relative deviation of the particle size and the average relative deviation of the pore size of the monodisperse polymer-based chromatographic microspheres produced by the intelligent manufacturing method are less than 2% for three consecutive production batches.

[0018] In a second aspect, the present invention provides an intelligent manufacturing system for monodisperse polymer-based chromatographic microspheres, comprising: The acquisition and establishment module is used to acquire historical synthesis data and establish a historical database. The historical synthesis data consists of process parameter data and microsphere structural characteristic data corresponding to the preparation of several monodisperse polymer-based chromatographic microspheres by the high-shear emulsification polymerization execution module. The process parameter data includes high-shear emulsification speed, shear time, mass ratio of polymer monomer to porogen, reaction temperature program, concentration of functionalizing reagent and reaction time. The microsphere structural characteristic data includes the average particle size, particle size variation coefficient, average pore size, specific surface area and surface functional group density of the microspheres. The model training module is used to train an ensemble learning model based on machine learning using a historical database; wherein the input of the ensemble learning model is the process parameter data of the chromatographic microspheres, and the output is the microsphere structural feature data corresponding to the process parameter data; The reverse reasoning module is used to obtain the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized, perform reverse reasoning calculations based on the integrated learning model, and output the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized. A high-shear emulsification polymerization execution module includes at least one high-shear emulsification reactor for synthesizing monodisperse polymer-based chromatographic microspheres based on the theoretical process parameters. The ensemble learning model trained by the model training module is a polymer-based chromatographic microsphere process-structure mapping network, which includes a feature embedding layer, a base learner group layer, a dynamic weight fusion layer and a microsphere structure feature output layer connected in series, as well as an uncertainty estimation module connected in parallel with the base learner group layer. The feature embedding layer is used to normalize continuous variables in the input process parameter data, perform temporal encoding of the reaction temperature program, perform embedding mapping on categorical variables, and generate explicit feature crosses, outputting a dense vector. The base learner group layer includes several base learner sub-models, each of which predicts a microsphere structure feature. Each sub-model shares the underlying feature representation and introduces physical prior constraint penalties during training. The uncertainty estimation module is used to output the predicted mean and 95% confidence interval of each microsphere structure feature based on Monte Carlo Dropout. The dynamic weight fusion layer includes a gating network, which adaptively adjusts the fusion weights output by each sub-model according to the input process parameter data. The microsphere structure feature output layer is used to output the microsphere structure feature data corresponding to the process parameter data.

[0019] In preparing monodisperse polymer-based chromatographic microspheres, this system combines the excellent physical control capabilities of the high-shear emulsification polymerization execution module with the intelligent prediction capabilities of machine learning. It uses an integrated learning model trained by the model training module to inversely infer the process parameters for preparing chromatographic microspheres with target performance. By constructing an intelligent manufacturing scenario, it achieves precise control of multiple parameters of polymer-based chromatographic microspheres, ultimately obtaining polymer-based chromatographic microspheres with target performance.

[0020] Furthermore, the reverse reasoning module, based on the ensemble learning model, performs reverse reasoning calculations and outputs the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized. This process execution unit includes: The scanning and filtering unit is used to perform coarse-grained scanning of the overall process parameter space based on the uncertainty estimation module, draw the predicted standard deviation contour lines of each microsphere structural feature, and filter the process parameter range with the predicted standard deviation below a preset threshold as the feasible region based on the predicted standard deviation contour lines. The optimization search unit employs a hybrid optimizer combining Bayesian optimization and covariance matrix adaptive evolution strategy within the feasible region. Its objective is to minimize the weighted Euclidean distance between the predicted and target microsphere structural feature data, searching for the optimal theoretical process parameters. Specifically, the optimizer iterates using a Gaussian process surrogate model and an expected improvement acquisition function during the Bayesian optimization phase. Once the Bayesian optimization converges or reaches its maximum iteration count, it switches to the covariance matrix adaptive evolution strategy for a refined search. The output unit outputs the theoretical process parameters along with their corresponding prediction confidence intervals and uncertainty scores to guide production.

[0021] Furthermore, the high-shear emulsification polymerization execution module is used to precisely execute the emulsification, polymerization and surface functionalization reactions of microspheres, and also includes: a precision metering feeding unit, a programmable temperature-controlled polymerization reactor and an automatic surface functionalization reaction unit; The precision metering and feeding unit is used to precisely control the addition ratio and timing of the aqueous phase and oil phase according to the theoretical process parameters. The programmed temperature-controlled polymerization reactor is used for multi-stage precise heating, holding and cooling processes for the emulsification and polymerization of reactants. The automated surface functionalization reaction unit is used to perform surface functionalization reactions based on the surface functional group density in the target microsphere structural feature data.

[0022] Furthermore, the intelligent manufacturing system also includes: The first data feedback module is used to collect the actual microsphere structural feature data of the product microspheres synthesized by the high-shear emulsification polymerization execution module, and feed the actual microsphere structural feature data back to the integrated learning model so that the integrated learning model can optimize the model by combining the actual microsphere structural feature data. The first data feedback module is connected online to the high-shear emulsification polymerization execution module and the model training module, and includes: an automatic sampling and conveying device, and a product characterization unit used in conjunction with the automatic sampling and conveying device; the automatic sampling and conveying device is used to automatically sample the product microspheres and convey them to the product characterization unit; the product characterization unit includes a scanning electron microscope, a specific surface area and pore size analyzer, and an infrared spectroscopy, used to determine the final particle size, particle size distribution, morphology, specific surface area, pore size distribution, and infrared spectrum of the product microspheres. The second data feedback module is connected online to the high-shear emulsification polymerization execution module. It is used to collect dynamic data of the reaction process of the high-shear emulsification polymerization execution module and dynamically fine-tune the high-shear emulsification speed and shear time based on the dynamic data. The dynamic data is the droplet size of the emulsion after high-shear emulsification.

[0023] The first data feedback module is used to construct the intelligent closed loop of the intelligent manufacturing system. By collecting and feeding back actual microsphere structural characteristic data, it updates the model training module and optimizes the model. This positive feedback enables the high-shear emulsification polymerization execution module to obtain target product microspheres with smaller errors in performance through more precise parameter control. The second data feedback module further reduces the error between the product microspheres and the target performance microspheres by collecting real-time data from the synthesis process and fine-tuning the process parameters.

[0024] Furthermore, the second data feedback module includes at least one online laser particle size analyzer, which is used for real-time in-situ monitoring of the particle size distribution and dispersion stability of emulsion droplets after high-shear emulsification.

[0025] Preferably, the high-shear emulsification reactor adopts a rotor-stator structure with a shearing speed of 2000~20000 rpm, a structural gap of 0.1~0.3 mm between the rotor and stator, and 6~12 teeth. Preferably, the high-shear emulsification reactor dynamically fine-tunes the shearing speed and shearing time based on real-time monitoring data from an online laser particle size analyzer to ensure that the emulsion droplet particle size distribution meets the preset target.

[0026] Furthermore, the monodisperse polymer-based chromatographic microspheres produced by the intelligent manufacturing system have the following customizable microsphere structural features, including: an average particle size of 1-20 μm; a particle size variation coefficient of less than 5%, preferably less than 3%; an average pore size of 5-100 nm; a specific surface area of ​​10-800 m² / g; and a surface functional group density of 0.1-2.0 mmol / g, wherein the surface functional groups include one or more of tertiary amine groups, sulfonic acid groups, carboxyl groups, hydroxyl groups, C18 alkyl chains, and phenyl groups. Furthermore, the average relative deviation of the particle size and the average relative deviation of the pore size of the monodisperse polymer-based chromatographic microspheres produced by the intelligent manufacturing system in three consecutive production batches are less than 2%.

[0027] In a third aspect, the present invention proposes the application of monodisperse polymer-based chromatographic microspheres prepared by the intelligent manufacturing method of the first aspect of the present invention or monodisperse polymer-based chromatographic microspheres prepared by the intelligent manufacturing system of the second aspect of the present invention in reversed-phase chromatography, normal-phase chromatography, hydrophilic interaction chromatography, ion exchange chromatography, size exclusion chromatography, and affinity chromatography.

[0028] Preferably, the application of the monodisperse polymer-based chromatographic microspheres in any chromatographic process involves packing the monodisperse polymer-based chromatographic microspheres into a chromatographic column using a high-pressure homogenization method. This chromatographic column exhibits high column efficiency, high selectivity, and high stability in separation and analysis.

[0029] 3. Beneficial effects Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres disclosed in this invention combines the physical precision control advantage of the high-shear emulsification polymerization execution module with the self-learning intelligent optimization system of machine learning for the first time. By using the integrated learning model to learn from historical synthesis data, a quantitative mapping relationship is established between process parameters and product microsphere performance parameters. Based on the target microsphere performance, the synthesis process parameters of the high-shear emulsification polymerization execution module are inversely inferred through model calculation, thus solving the problem of multi-parameter coupling and inability to coordinate and precisely control during the preparation of monodisperse polymer-based chromatographic microspheres.

[0030] (2) The intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres disclosed in this invention overturns the traditional R&D model, which is driven by operational experience and trial and error, and is driven by model prediction. This greatly shortens the R&D cycle, reduces the dependence on skilled engineers, and realizes the reverse design and digital intelligent manufacturing of chromatographic packings. Specifically, the ensemble learning model proposed in this invention is a polymer-based chromatographic microsphere process-structure mapping network, comprising a feature embedding layer, a base learner group layer, a dynamic weight fusion layer, and a microsphere structure feature output layer connected in series, as well as an uncertainty estimation module connected in parallel with the base learner group layer. The feature embedding layer is used to normalize continuous variables in the input process parameter data, perform time-series encoding of the reaction temperature program, embedding and mapping categorical variables, and generate explicit feature crosses, outputting a dense vector. The base learner group layer includes several sub-models of base learners, each sub-model predicting a microsphere structure feature. Each sub-model shares the underlying feature representation and introduces physical prior constraint penalties during training. The uncertainty estimation module is used to output the predicted mean and 95% confidence interval of each microsphere structure feature based on Monte Carlo Dropout. The dynamic weight fusion layer includes a gating network, which adaptively adjusts the fusion weights output by each sub-model according to the input process parameter data. The microsphere structure feature output layer is used to output the microsphere structure feature data corresponding to the process parameter data. The integrated learning model designed in this invention can simultaneously optimize the prediction errors of multiple targets such as particle size, distribution, and pore size during training. Compared with models such as XGBoost, the average absolute error in the particle size variation coefficient prediction task is reduced by more than 20%, and the prediction accuracy is higher.

[0031] (3) The intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres disclosed in this invention updates the model with the actual process parameters and actual microsphere structural characteristics of the previous batch of product microspheres each time, resulting in monodisperse polymer-based chromatographic microsphere products with excellent and consistent performance; the microsphere structural characteristics that can be co-customized include: the average particle size is 1~20μm; the particle size variation coefficient is less than 5%, preferably less than 3%; the average pore size is 5~100 nm; the specific surface area is 10~800m² / g; the surface functional group density is 0.1~2.0 mmol / g, and the surface functional groups include one or more of tertiary amine groups, sulfonic acid groups, carboxyl groups, hydroxyl groups, C18 alkyl chains, and phenyl groups; that is, the product microspheres obtained by this invention not only have a precisely customizable structure and performance, but also excellent monodispersity, fully meeting the stringent requirements of high-end chromatographic applications for packing consistency.

[0032] (4) The intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres disclosed in this invention significantly improves production efficiency compared to existing technologies. The high yield characteristics of the high-shear emulsification polymerization execution module, combined with the optimization of the integrated learning model, enable each batch of samples manufactured to meet the usage requirements, making efficient use of resources, significantly reducing overall production costs, and possessing the development prospects of sustainable and industrialized production. Attached Figure Description

[0033] Figure 1 This is a flowchart of the intelligent manufacturing method for monodisperse polymer-based chromatographic microspheres disclosed in this invention; Figure 2 This is a schematic diagram of the composition structure of the ensemble learning model disclosed in this invention; Figure 3 This is a correlation diagram between the predicted and measured values ​​of the microsphere size of the test set by the ensemble learning model disclosed in this invention. Figure 4 This is a schematic diagram illustrating the working principle of the high-shear emulsification reactor disclosed in the embodiment. Figure 5 This is a flowchart illustrating the implementation of the intelligent manufacturing system for monodisperse polymer-based chromatographic microspheres disclosed in this invention. Figure 6 This is a SEM image of the monodisperse polymer-based chromatographic microspheres prepared in Example 1; Figure 7 A comparison of glyphosate standard analysis using a chromatographic column supported by the product of Example 1 and a commercial chromatographic column. Detailed Implementation

[0034] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0035] Unless otherwise specified in the examples, the procedures should be performed under standard conditions or conditions recommended by the manufacturer. Reagents or instruments whose manufacturers are not specified are all commercially available products.

[0036] As used herein, the term “about” is used to provide for the flexibility and imprecision associated with a given term, measure, or value. Those skilled in the art can readily determine the degree of flexibility for a particular variable. Molecular weights and other numerical data may be presented in range format herein. It should be understood that such range format is used solely for convenience and brevity and should be flexibly interpreted to include not only the values ​​explicitly stated as the limits of the range, but also all individual values ​​or subranges encompassed within the range, as if each value and subrange were explicitly stated. For example, a range of values ​​from about 1 to about 4.5 should be interpreted to include not only the explicitly stated limits of 1 to 4.5, but also individual numbers (such as 2, 3, 4) and subranges (such as 1 to 3, 2 to 4, etc.). The same principle applies to ranges that describe only a single value, such as “less than about 4.5,” which should be interpreted to include all the aforementioned values ​​and ranges. Furthermore, this interpretation should apply regardless of the breadth of the range or characteristic described.

[0037] Combination Figure 1 As shown, this invention proposes a smart manufacturing method for monodisperse polymer-based chromatographic microspheres, comprising the following steps: Step S100: Obtain historical synthesis data and establish a historical database. The historical synthesis data consists of process parameter data and microsphere structural characteristic data corresponding to the preparation of several monodisperse polymer-based chromatographic microspheres by the high-shear emulsification polymerization execution module in the past. The process parameter data includes high-shear emulsification speed, shear time, mass ratio of polymer monomer to porogen, reaction temperature program, concentration of functionalizing reagent and reaction time. The microsphere structural characteristic data includes the average particle size, particle size variation coefficient, average pore size, specific surface area and surface functional group density of the microspheres. Step S200: Train an ensemble learning model based on machine learning according to the historical database; wherein the input of the ensemble learning model is the process parameter data of the chromatographic microspheres, and the output is the microsphere structural feature data corresponding to the process parameter data; In this embodiment, the ensemble learning model is a polymer-based chromatographic microsphere process-structure mapping network for high-dimensional small sample data, denoted as PS-MapNet. PS-MapNet adopts a multi-task loss function during training and optimizes the prediction errors of multiple targets such as particle size, distribution, and pore size. Compared with models such as XGBoost, the average absolute error in the particle size variation coefficient prediction task is reduced by more than 20%.

[0038] Combination Figure 2 As shown, the integrated learning model is a polymer-based chromatographic microsphere process-structure mapping network, which includes a feature embedding layer, a base learner group layer, a dynamic weight fusion layer, and a microsphere structure feature output layer connected in series, as well as an uncertainty estimation module connected in parallel with the base learner group layer; the detailed architecture and principle of each structural layer are described in detail below.

[0039] The feature embedding layer is responsible for converting the original process parameters into feature vectors suitable for deep learning processing. It normalizes continuous variables in the input process parameter data, performs temporal encoding of the reaction temperature program, embeds and maps categorical variables, generates explicit feature crosses, and outputs a dense vector. Specifically, the normalization of continuous variables includes Z-score normalization of high-shear emulsification speed (2000~20000rpm), shear time (0~600s), polymer monomer to porogen mass ratio (0.5~2.0), functionalized reagent concentration (0~30%v / v), and reaction time (0~24h) to ensure that all features are on the same order of magnitude. The temporal encoding of the reaction temperature program involves splitting it into a heating rate sequence, a holding time sequence, and a target temperature sequence. A one-dimensional convolutional layer is used to extract the local temporal patterns of the heating rate sequence, holding time sequence, and target temperature sequence, and global average pooling is used to extract the temperature feature vector. The kernel size of the one-dimensional convolutional layer is 3. Embedding mapping for categorical variables includes: 1) for monomer types, such as St, DVB, GMA, and VBC; 2) for porogen types, such as toluene, dodecanol, and cyclohexane; and 3) for functionalized reagent types, etc., which are discrete variables. A learnable embedding matrix is ​​used to map these variables into dense vectors. The process of generating explicit feature crosses involves generating cross features as additional inputs through explicit second-order feature combinations, such as shear speed × shear time and monomer ratio × porogen ratio.

[0040] Finally, the feature embedding layer outputs a dense vector with a dimension of 128 for use by subsequent units.

[0041] The base learner layer comprises several sub-models of the base learners, each sub-model corresponding to predict a microsphere structural feature. Each sub-model shares the underlying feature representation and incorporates physical prior constraint penalties during training. Specifically, it includes multiple sub-models designed for different microsphere structural features, including particle size, pore size, and functional group density. Each sub-model shares the underlying features but has an independent output head. This unit is a multi-task learning architecture containing K independent base learners, where K equals the number of microsphere structural features to be predicted, such as... Figure 2 The K=5 shown corresponds to the average particle size, particle size variation coefficient, average pore size, specific surface area, and surface functional group density. Each base learner submodule employs a three-layer fully connected network with hidden layer dimensions of 64, 32, and 16, and possesses the following characteristics: 1) Shared underlying features: The first layer of all base learners shares the same weight matrix, extracting a common representation from the feature embedding layer.

[0042] 2) Physically-guided weighted loss: During training, physical prior constraints are introduced for each base learner; for example, the prediction of average particle size should satisfy the following: particle size is negatively correlated with shear rotation speed and positively correlated with monomer concentration; if the model prediction violates this physical trend, an additional penalty term is added. Similarly, the prediction of specific surface area should satisfy the following: specific surface area is negatively correlated with pore size, roughly following an inverse relationship; violations are also penalized.

[0043] 3) Independent output head: Each base learner outputs a scalar value corresponding to a structural feature, and at the same time outputs the local confidence of the prediction based on the activation variance of the learner under the current input.

[0044] The design of the base learner layer enables the network to optimize multiple objectives simultaneously, avoiding conflicts between multiple tasks for a single model, while using physical knowledge to constrain the rationality of predictions.

[0045] The uncertainty estimation module is used to output the predicted mean and 95% confidence interval of each microsphere structural feature based on Monte Carlo Dropout; to perform variance estimation on the prediction results based on Monte Carlo Dropout, and to output the prediction interval corresponding to the process parameters for subsequent feasible region screening in reverse inference; the specific implementation process is as follows: during the training phase, neurons in the fully connected layer of each base learner are randomly dropped with probability p=0.1; during the inference phase, Dropout activation is maintained, and the same input is forward propagated T times, usually T is 30~50 times, to obtain T predicted values; the mean μ and standard deviation σ of the T predictions are calculated, where σ is used as the uncertainty measure of the prediction; the output prediction interval is [μ-1.96σ, μ+1.96σ], corresponding to the 95% confidence interval.

[0046] The output of the uncertainty estimation module is used in reverse reasoning to: filter high-confidence prediction regions; and prompt the user to increase experimental samples to reduce uncertainty when the target microsphere structure is in a high-uncertainty region.

[0047] The purpose of the dynamic weight fusion layer is to adaptively adjust the weights of each base learner in the final output based on the confidence level of the input process parameters. Specifically, it includes a gating network, which is used to adaptively adjust the fusion weights of the outputs of each sub-model based on the input process parameter data. Since the reliability of each base learner varies across different process parameter ranges—for example, the particle size predictor is more accurate in the high shear speed range, while the pore size predictor is more accurate in the low speed range—this is crucial.

[0048] The structure of the gated network is as follows: Input: a 128-dimensional vector output from the feature embedding layer; Hidden layer: 64-dimensional, ReLU activated; Output: A K-dimensional weight vector (K=5), normalized by Softmax, satisfying Σw i =1, w i is the fusion weight of the microsphere structure features corresponding to the i-th base learner.

[0049] The final predicted value of the structural feature of the j-th microsphere j for: j = Σ i (w i *h ij ); Among them, h ij Let i represent the prediction of the j-th microsphere structural feature by the i-th base learner. Since each base learner actually predicts only one microsphere structural feature, i=j. However, the gating network can independently assign weights to each output microsphere structural feature more accurately. The gating network outputs a K×K weight matrix, allowing cross-fusion. The dynamic weight fusion layer enables the ensemble learning model to automatically select the most reliable base learner based on the current process parameters, significantly improving the robustness of the prediction.

[0050] The microsphere structure feature output layer is used to output the microsphere structure feature data corresponding to the process parameter data.

[0051] The ensemble learning model selected in this invention employs the following training strategy: Since historical data is usually small, ranging from tens to hundreds of data points, PS-MapNet employs the following strategies to prevent overfitting and improve generalization ability: 1) Data augmentation: Gaussian noise is added to continuous variables to generate virtual samples; 2) Transfer learning pre-training: Feature embedding layers and base learner groups are pre-trained on large-scale public polymer synthesis datasets, such as literature data, and then fine-tuned on a self-built historical database; 3) Early stopping and model ensemble: Five PS-MapNet sub-models are trained using 5-fold cross-validation, and the final prediction is the mean of the predictions of the five sub-models. Uncertainty is the geometric mean of the standard deviation of the sub-model predictions and the standard deviation of Monte Carlo Dropout.

[0052] The above ensemble learning model uses a multi-task loss function to optimize the target prediction error during training. The formula for calculating the multi-task loss function is as follows: L total = Σ k ( α k * MSE k )+ λ physics * L physics + λ reg * L reg In the formula, L total Represents the total loss function; k Indicates the first k Individual microsphere structural features; MSE k Indicates the first k Mean square prediction error of individual microsphere structural features; α k For the first k The fusion weights of individual microsphere structural features are determined based on the historical variation coefficients of the microsphere structural features. CV k set up, α k = 1 / CV k ; L physics As a physical constraint penalty term, calculate the degree of violation of physical trends; L reg This is an L2 regularization term to prevent overfitting; λ physics , λ reg These are hyperparameters, with values ​​of 0.1 and 0.001 respectively.

[0053] Step S300: Obtain the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized, perform reverse reasoning calculation based on the integrated learning model, and output the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized. The steps for performing reverse reasoning calculation based on the integrated learning model include: performing a coarse-grained scan of the overall process parameter space based on the uncertainty estimation module, drawing the prediction standard deviation contour lines of each microsphere structural feature, and selecting the process parameter range with a prediction standard deviation lower than a preset threshold as the feasible region based on the prediction standard deviation contour lines. Within the feasible region, an optimizer combining Bayesian optimization and covariance matrix adaptive evolution strategy is employed. The objective is to minimize the weighted Euclidean distance between the predicted microsphere structural feature data and the target microsphere structural feature data, and to search for the optimal process parameters, which are the theoretical process parameters. Specifically, the optimizer uses a Gaussian process surrogate model and an expected improvement acquisition function for iteration during the Bayesian optimization phase. When the Bayesian optimization converges and stalls or reaches the maximum number of iterations, it switches to the covariance matrix adaptive evolution strategy for fine-grained search. Output theoretical process parameters and their corresponding prediction confidence intervals and uncertainty scores to guide production.

[0054] Specifically, the implementation process of the reverse inference method based on PS-MapNet is detailed below.

[0055] First, let's define the problem of reverse inference based on PS-MapNet: given the user's desired microsphere structure feature vector y target = [d target CV target pore target SSA targe t, func target The optimal process parameters x = [rotation speed, time, single-component ratio, temperature program parameters, functionalization concentration, functionalization time] need to be solved to ensure that the PS-MapNet predictions are optimized. =f(x) should be as close as possible to y target It also meets equipment constraints, such as rotational speed range and maximum reactor temperature; among which, the monomer-porogen ratio is the ratio of monomer to porogen.

[0056] Secondly, before reverse inference, the uncertainty estimation module of PS-MapNet is used to perform grid sampling on the entire process parameter space, i.e. coarse-grained scanning, and draw contour lines of the predicted standard deviation of each output feature. Only the process parameter range with the predicted standard deviation less than the preset threshold is retained as the feasible region, and the above optimization is performed within the feasible region. The feasible region screening significantly improves the success rate of reverse design. The preset threshold can be set to 0.1 × target value.

[0057] During subsequent optimization, a hybrid optimizer combining physically constrained Bayesian optimization and an adaptive evolutionary strategy based on the covariance matrix was employed. Optimizers were selected from the historical database that were consistent with y. target The five closest samples were used as the initial population for initialization; Bayesian optimization phase: A Gaussian process surrogate model is used to model the objective function, i.e., the negative Euclidean distance between the prediction and the target; the acquisition function adopts a weighted combination of expected improvement and prediction uncertainty to encourage exploration of high uncertainty regions; each iteration recommends a candidate process parameter x. candidate .

[0058] The covariance matrix adaptive evolution strategy stage: A Gaussian perturbation population with an adaptive covariance matrix is ​​generated, centered on the current optimal solution; the fitness of individuals in the Gaussian perturbation population is evaluated, i.e., the negative Euclidean distance from the target plus a penalty term for process parameters, such as penalties for extreme values; the optimal solution and the covariance matrix are updated; physical constraint handling includes: hard constraints, such as direct elimination of candidate solutions that exceed equipment limits (e.g., speed, time); soft constraints, such as multiplying the fitness by a decay factor of 0.9 when the product of shearing time and speed exceeds a preset upper limit.

[0059] The optimizer ultimately outputs a set of theoretical process parameters, along with predicted values ​​and 95% confidence intervals for each microsphere structural feature. The prediction uncertainty score for these parameters ranges from 0 to 1, with lower scores indicating higher reliability. Recommended production batch sizes are also provided; if uncertainty is high, a small-scale trial is recommended first.

[0060] Step S400: Based on the theoretical process parameters, the synthesis step is performed using the high-shear emulsification polymerization execution module to synthesize monodisperse polymer-based chromatographic microspheres; wherein, the high-shear emulsification polymerization execution module includes at least one high-shear emulsification reactor.

[0061] To reduce the deviation between the structural characteristics of the monodisperse polymer-based chromatographic microspheres actually prepared by the method and the structural characteristics of the target microspheres, this method further includes: step S500, collecting the actual microsphere structural characteristic data of the product microspheres synthesized by the high-shear emulsification polymerization execution module, and feeding the actual microsphere structural characteristic data back to the integrated learning model so that the integrated learning model can optimize the model by combining the actual microsphere structural characteristic data. That is, the intelligent manufacturing method of this invention forms an intelligent closed-loop optimization loop when applied. After each synthesis is completed using the high-shear emulsification polymerization execution module, the collected actual result data is compared with the set target, and the deviation is fed back to the integrated learning model, so that the model can continuously learn and self-optimize, automatically adjusting the process parameters for the next synthesis, forming an intelligent closed loop of "design-synthesis-detection-learning-optimization" that continuously approaches the optimal target. For example, Figure 3The correlation diagram shown in the figure illustrates the relationship between the predicted and measured values ​​of microsphere size on the test set by the ensemble learning model. When the ensemble learning model trained using the training strategy disclosed in this invention is tested on the test set, the coefficient of determination of the predicted particle size is higher than that of the actual particle size, indicating that the model is overfitting and requires further optimization by adding more training data.

[0062] As an optional practical approach, the intelligent manufacturing method for monodisperse polymer-based chromatographic microspheres further includes: collecting dynamic data of the reaction process of the high-shear emulsification polymerization execution module, and dynamically fine-tuning the high-shear emulsification rotation speed and shear time based on the dynamic data; wherein, the dynamic data is the droplet size of the emulsion after high-shear emulsification. That is, during the method, the shear rotation speed and shear time are dynamically fine-tuned based on the real-time monitored droplet size data to ensure that the droplet size distribution of the emulsion meets the preset target.

[0063] Based on the same inventive concept as the above-described method embodiments, this application also discloses an intelligent manufacturing system for monodisperse polymer-based chromatographic microspheres. The intelligent manufacturing system includes: an acquisition and establishment module for acquiring historical synthesis data and establishing a historical database. The historical synthesis data consists of process parameter data and microsphere structural characteristic data corresponding to the historical preparation of several monodisperse polymer-based chromatographic microspheres by a high-shear emulsification polymerization execution module. The process parameter data includes high-shear emulsification speed, shear time, mass ratio of polymer monomer to porogen, reaction temperature program, functionalizing reagent concentration, and reaction time. The microsphere structural characteristic data includes the average particle size, particle size variation coefficient, and average pore size of the microspheres. The system includes: a specific surface area and surface functional group density; a model training module for training an ensemble learning model based on a historical database; wherein the input of the ensemble learning model is the process parameter data of the chromatographic microspheres, and the output is the microsphere structural feature data corresponding to the process parameter data; a reverse inference module for obtaining the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized, performing reverse inference calculations based on the ensemble learning model, and outputting the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized; and a high-shear emulsification polymerization execution module, including at least one high-shear emulsification reactor, for synthesizing monodisperse polymer-based chromatographic microspheres based on the theoretical process parameters.

[0064] The ensemble learning model trained by the model training module is a polymer-based chromatographic microsphere process-structure mapping network, comprising a feature embedding layer, a base learner group layer, a dynamic weight fusion layer, and a microsphere structure feature output layer connected in series, as well as an uncertainty estimation module connected in parallel with the base learner group layer. The feature embedding layer normalizes continuous variables in the input process parameter data, performs time-series encoding of the reaction temperature program, embeds and maps categorical variables, generates explicit feature crosses, and outputs a dense vector. The base learner group layer includes several sub-models of base learners, each sub-model predicting a microsphere structure feature. Each sub-model shares the underlying feature representation and incorporates physical prior constraint penalties during training. The uncertainty estimation module outputs the predicted mean and 95% confidence interval of each microsphere structure feature based on Monte Carlo Dropout. The dynamic weight fusion layer includes a gating network that adaptively adjusts the fusion weights output by each sub-model according to the input process parameter data. The microsphere structure feature output layer outputs the microsphere structure feature data corresponding to the process parameter data.

[0065] The reverse reasoning module, based on the ensemble learning model, performs reverse reasoning calculations and outputs the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized. This process execution unit includes: a scanning and screening unit, used to perform a coarse-grained scan of the overall process parameter space based on the uncertainty estimation module, plotting the predicted standard deviation contour lines of each microsphere structural feature, and screening process parameter ranges with predicted standard deviations below a preset threshold as feasible regions based on the predicted standard deviation contour lines; an optimization search unit, used within the feasible region, employing an optimizer combining Bayesian optimization and a covariance matrix adaptive evolution strategy to minimize the weighted Euclidean distance between the predicted microsphere structural feature data and the target microsphere structural feature data, searching for the optimal process parameters as theoretical process parameters; wherein, the optimizer iterates using a Gaussian process surrogate model and an expected improvement acquisition function during the Bayesian optimization phase, and switches to a covariance matrix adaptive evolution strategy for fine-grained searching after the Bayesian optimization converges and stagnates or reaches the maximum number of iterations; and an output unit, used to output the theoretical process parameters and their corresponding predicted confidence intervals and uncertainty scores to guide production.

[0066] To further reduce the deviation between the actual product microsphere structural features obtained by the high-shear emulsification polymerization execution module and the target microsphere structural features, the intelligent manufacturing system of the present invention also includes a first data feedback module. The first data feedback module constructs an intelligent closed loop of the intelligent manufacturing system, which is used to collect the actual microsphere structural feature data of the product microspheres synthesized by the high-shear emulsification polymerization execution module, and feed back the actual microsphere structural feature data to the integrated learning model so that the integrated learning model can perform model optimization in combination with the actual microsphere structural feature data.

[0067] As an optional implementation, the intelligent manufacturing system further includes: a second data feedback module connected online to the high-shear emulsification polymerization execution module, used to collect dynamic data of the reaction process of the high-shear emulsification polymerization execution module, and dynamically fine-tune the high-shear emulsification speed and shear time based on the dynamic data; wherein, the dynamic data is the droplet size of the emulsion after high-shear emulsification; the second data feedback module fine-tunes the process parameters of the high-shear emulsification polymerization execution module by collecting real-time data of the synthesis process, further reducing the error between the product microspheres and the target performance microspheres.

[0068] One embodiment of the intelligent manufacturing system for the above-mentioned monodisperse polymer-based chromatographic microspheres specifically includes: The high-shear emulsification polymerization execution module is used to precisely execute the emulsification, polymerization, and surface functionalization reactions of the product microspheres. Its core includes a high-shear emulsification reactor equipped with a precision metering feed unit, a programmable temperature-controlled polymerization reactor, and an automated surface functionalization reaction unit. Its working principle is as follows: Figure 4 As shown, the high-shear emulsification reactor adopts a rotor-stator structure with a shearing speed of 2000~20000 rpm, a structural gap of 0.1~0.3 mm between the rotor and stator, and 6~12 teeth. Preferably, the high-shear emulsification reactor dynamically fine-tunes the shearing speed and shearing time based on real-time monitoring data from an online laser particle size analyzer to ensure that the emulsion droplet particle size distribution meets the preset target. The precision metering feed unit can accurately control the addition ratio and timing of the aqueous and oil phases according to instructions; wherein, the oil phase metered by the precision metering feed unit includes monomers, pore-forming agents, and initiators. The programmed temperature-controlled polymerization reactor has a multi-stage precise heating, holding, and cooling program for the emulsification and polymerization of the reactants. The automatic surface functionalization reaction unit is used to perform surface functionalization reactions based on the surface functional group density in the target microsphere structural characteristic data.

[0069] A multi-dimensional online monitoring and characterization module, connected online to the high-shear emulsification polymerization execution module, is used to collect dynamic data of the reaction process and performance data of the product microspheres in real time; including: The process monitoring unit includes at least one online laser particle size analyzer for real-time, in-situ monitoring of the particle size distribution (DSD) and dispersion stability of the emulsion droplets after high-shear emulsification; the process monitoring unit constitutes the second data feedback module mentioned above. The product characterization unit includes an automated sampling and delivery device, and a scanning electron microscope, a specific surface area and pore size analyzer, and an infrared spectroscopy system used in conjunction with the automated sampling and delivery device. The automated sampling and delivery device is used to automatically sample the product microspheres and deliver them to the product characterization unit, and is typically implemented using a robotic arm. The product characterization unit is used to determine the final particle size, particle size distribution, morphology, specific surface area, pore size distribution, and Fourier transform infrared (FTIR) spectrum of the synthesized product microspheres. The multi-dimensional online monitoring and characterization module constitutes the data acquisition part of the first data feedback module described above.

[0070] A machine learning optimization and prediction platform is used to establish a quantitative mapping model from process parameters to microsphere performance and to perform inverse parameter optimization based on target performance. The platform incorporates a trained ensemble learning algorithm model. This model takes high-shear emulsification speed, shear time, polymer monomer to porogen mass ratio, reaction temperature program, functionalized reagent concentration, and reaction time as inputs, and outputs the average particle size, particle size variation coefficient, average pore size, specific surface area, and surface functional group density of the microspheres, establishing a nonlinear relationship between them. The platform receives the target microsphere performance parameters set by the user, and uses the model to inversely predict and output a set of optimized process parameters. In other words, the aforementioned model training module and inverse inference module are integrated into the machine learning optimization and prediction platform.

[0071] The closed-loop feedback control module, acting as the central nervous system, is communicatively connected to the execution module, monitoring and characterization module, and machine learning platform. This module is configured to: distribute the predicted process parameter set output by the machine learning optimization and prediction platform to the high-shear emulsification polymerization execution module to drive the synthesis process; simultaneously, receive real-time process data and final product data collected by the multi-dimensional online monitoring and characterization module, and transmit these data back to the machine learning optimization and prediction platform as feedback signals; the machine learning optimization and prediction platform uses the feedback data to incrementally learn or retrain the model, thereby achieving continuous self-optimization of model prediction accuracy and adaptive adjustment of synthesis parameters, forming an intelligent closed loop of "target input - parameter prediction - accurate synthesis - online detection - feedback optimization".

[0072] The intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres disclosed in the present invention can collaboratively customize the microsphere structural features, including: an average particle size of 1-20 μm; a particle size variation coefficient of less than 5%, preferably less than 3%; an average pore size of 5-100 nm; a specific surface area of ​​10-800 m² / g; and a surface functional group density of 0.1-2.0 mmol / g, wherein the surface functional groups include one or more of tertiary amine groups, sulfonic acid groups, carboxyl groups, hydroxyl groups, C18 alkyl chains, and phenyl groups; and the average relative deviation of the particle size of the obtained monodisperse polymer-based chromatographic microspheres in three consecutive production batches is less than 2%, and the average relative deviation of the pore size is less than 5%, exhibiting excellent dispersibility and batch-to-batch consistency.

[0073] The intelligent manufacturing method and system for monodisperse polymer-based chromatographic microspheres disclosed in this invention will be further described in detail below with reference to specific embodiments.

[0074] Intelligent manufacturing system for monodisperse polymer-based chromatographic microspheres based on Figure 5 The publicly disclosed implementation process for preparing HILIC chromatographic microspheres for the direct detection of glyphosate is as follows: 1) Target input: The operator sets the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized on the interface of the machine learning optimization and prediction platform, including: the average particle size of the microspheres is 5 μm, the particle size variation coefficient is less than 5%, the average pore size is 10 nm, the specific surface area is 100 m² / g, the surface functional group is tertiary amine group, and the density is 200 μmol / g.

[0075] 2) Intelligent decision-making: The user inputs the target microsphere structure: average particle size 5 μm, particle size variation coefficient <5%, average pore size 10 nm, specific surface area 100 m² / g, and tertiary amine group density 200 μmol / g.

[0076] The system first calls the pre-trained PS-MapNet network. The feature embedding layer of the network encodes the features of 286 pairs of process-structure data from the historical database. The temperature program is automatically split into time-series sub-features such as "heating rate 2.5℃ / min → holding for 60min → heating rate 1.5℃ / min → holding for 120min". In the base learner group, the base learner specifically targeting the average particle size (sub-network A) was subjected to a physical penalty of "particle size is negatively correlated with shear rotation speed" during training, and its coefficient of determination R² for predicting particle size reached 0.96; the base learner targeting the particle size variation coefficient (sub-network B) was subjected to a constraint of "variance coefficient is negatively correlated with shear time", and the prediction error was ±0.7%.

[0077] The dynamic weighted fusion layer automatically sets the weights of the average particle size-based learner to 0.35, the pore size-based learner to 0.25, the specific surface area to 0.20, the functional group density to 0.15, and the coefficient of variation to 0.05 based on the current target (5μm particle size is in the medium particle size range).

[0078] The uncertainty estimation module scanned all candidate process parameter ranges and found that the prediction uncertainty (standard deviation <3%) was lowest in the range of shear speed 14000~16000 rpm, shear time 280~320s, and monomer / porogen ratio 2:3~2.5:3. The inverse reasoning optimizer used Bayesian optimization to search within this range and converged after 35 iterations, outputting the optimal theoretical process parameters: high-shear emulsification speed 15000 rpm, shear time 300s, monomer (St / DVB=1:1) to porogen (toluene) mass ratio 2:3, temperature program (70℃ 2h → 85℃ 4h), functionalizing reagent (chloromethyl methyl ether) concentration 15% v / v, and functionalization time 6h.

[0079] The system also provides the predicted microsphere structure: average particle size 5.03 μm (95% CI: 4.85–5.21), particle size variation coefficient 4.2% (95% CI: 3.5%–5.0%), average pore size 10.2 nm (95% CI: 9.5–11.0), specific surface area 98 m² / g (95% CI: 92–10⁵), and functional group density 195 μmol / g (95% CI: 182–20⁸). All predicted ranges meet the user requirements, with an uncertainty score of 0.12 (low). The system automatically sends the parameters to the high-shear emulsification polymerization execution module.

[0080] 3) Automatic Execution: The closed-loop feedback control module sends the theoretical process parameters to the high-shear emulsification polymerization execution module, which automatically completes the following operations: Precision feeding: Add an aqueous phase containing sodium dodecyl sulfate (SDS) and polyvinylpyrrolidone (PVP) according to the formula, and add oil phase A, which contains monomers DVB (divinylbenzene), toluene, and azobisisobutyronitrile (AIBN). Prepolymerization and high-shear emulsification: After prepolymerization at 60°C for 30 minutes, oil phase B, which contains the monomers ST (styrene), toluene, and azobisisobutyronitrile (AIBN) in the formulation, is added, and prepolymerization is carried out again at 70°C for 10 minutes; subsequently, the high-shear emulsification reactor is automatically sheared at 15000 rpm for 300 seconds. Polymerization and functionalization: The microspheres were transferred to a temperature-controlled polymerization reactor and reacted at 85°C for 2 hours to obtain matrix microspheres. The matrix microspheres were then automatically subjected to chloromethylation and diethylamine amination reactions in an automated surface functionalization reaction unit.

[0081] 4) Sensing and Characterization: During the synthesis process, the droplet size distribution (DSD) of the emulsion was monitored in real time using an online laser particle size analyzer. After the functionalization reaction, samples were automatically sampled and transported to the product characterization unit, where the actual structural characteristics of the microspheres were measured: the average particle size was 4.9 μm, the coefficient of variation was 4.5%, the average pore size was 10.5 nm, the specific surface area was 95.7 m² / g, the surface functional group was tertiary amine, and the density was 186 μmol / g. The microstructure and dispersibility of the HILIC chromatographic microspheres are shown in the figure. Figure 6 As shown, the monodispersity is excellent and the nanospheres are of complete size.

[0082] 5) Closed-loop feedback and continuous model optimization: The measured microsphere structural feature data is automatically transmitted back to the machine learning optimization and prediction platform. The platform calculates whether the deviation between the measured data and the target value exceeds the preset deviation threshold. When the calculated deviation does not exceed the preset deviation threshold, it indicates that the synthesis verification is successful. The synthesized "process parameters-microsphere structural features" data pair is stored in the database as a new sample for continuous model iteration and optimization.

[0083] Specifically, after synthesis, the product characterization unit measured the actual microsphere structure as follows: average particle size 4.9 μm, coefficient of variation 4.5%, average pore size 10.5 nm, specific surface area 95.7 m² / g, and functional group density 186 μmol / g. The system calculated the Euclidean distance between the actual value and the target value to be 0.23, which is less than the preset threshold of 0.5, thus classifying it as a qualified batch.

[0084] This set of "theoretical process parameters + actual microsphere structure" was added to the historical database as new samples, subsequently triggering the incremental learning process of PS-MapNet: The system first recalculates the adaptive loss weights for each feature using the existing database. α k Then, fine-tuning is performed for 10 epochs using the new samples as the primary focus. The first two layers of the feature embedding layer are frozen, and only the base learner group and the dynamic weight fusion layer are updated. After fine-tuning, the model's prediction error decreases by approximately 15% in the local region near the new samples. After the model update is complete, the system automatically records the updated model version number in the log and uses it for the next batch of reverse inference. This continuous learning mechanism allows the model to evolve continuously with the increase of production batches, achieving extremely high prediction accuracy after long-term operation.

[0085] The measured microsphere structural feature data is automatically transmitted back to the machine learning optimization and prediction platform. The platform calculates whether the deviation from the target value exceeds the preset deviation threshold. When the calculated deviation does not exceed the preset deviation threshold, it indicates that the synthesis verification is successful. The synthesized "process parameters-microsphere structural features" data pair is stored in the database as a new sample for continuous model iteration and optimization.

[0086] The HILIC chromatographic microspheres prepared above were then packed into chromatographic column tubes using a high-pressure homogenization method to construct a chromatographic column. Glyphosate standards were analyzed in HILIC mode, combined with... Figure 7 As shown, the chromatographic column corresponding to the HILIC chromatographic microspheres exhibits performance far exceeding that of commercial columns (Waters APPC): retention time RSD < 1.2% (n=100), symmetrical peak shape, and detection limit as low as 0.2 ppb, perfectly achieving high sensitivity, high stability, and derivatization-free direct detection.

[0087] Examples 2-5 follow the same process as Example 1 in preparing chromatographic microspheres, except that the chromatographic microspheres required for different chromatograms and the target microsphere structural characteristics are different. The theoretical process parameters are first output through reverse reasoning using a machine learning optimization and prediction platform. Then, the high-shear emulsification polymerization execution module is used to adjust the parameters in real time based on the theoretical process parameters and the preparation process. The actual structural characteristics and chromatographic performance verification of the final chromatographic microspheres are shown in Table 1 below. The actual process parameters and actual microsphere structural characteristics constitute a data pair that is fed back to the ensemble learning model for incremental learning and parameter updates, enabling continuous model optimization.

[0088] In summary, the intelligent manufacturing method for monodisperse polymer-based chromatographic microspheres disclosed in this invention utilizes a machine learning model to map historical chromatographic microsphere structural characteristics and process parameters, establishing an integrated learning model that deeply integrates high-shear emulsification technology with machine learning algorithms. This model enables reverse reasoning of the theoretical process parameters for preparing target microspheres based on their structural characteristics, guiding the actual preparation process and ultimately obtaining actual microspheres with acceptable deviations from the target microsphere structural characteristics. Technically, by constructing an "execution-perception-decision" intelligent route, the invention achieves the preparation of chromatographic microspheres with customized structural characteristics, providing a reliable technical path for industrial continuous production. In terms of future prospects, this invention successfully constructs an intelligent digital factory for chromatographic microspheres, providing a novel and industrially feasible solution to the challenges of high-end chromatographic materials through the deep integration and closed-loop feedback of high-shear emulsification and machine learning.

[0089] Table 1. Structural characteristics, process parameters, and performance verification data of the target chromatographic microspheres prepared in Examples 2-5

[0090] The above description provides an illustrative overview of the present invention and its embodiments. This description is not restrictive, and the embodiments shown are merely one example of the invention's implementation. Actual implementations are not limited to these examples. Therefore, if those skilled in the art are inspired by this description and design similar implementations and examples without departing from the spirit of the invention, such designs should fall within the scope of protection of the present invention.

Claims

1. A smart manufacturing method for monodisperse polymer-based chromatographic microspheres, characterized in that, include: Historical synthesis data is acquired and a historical database is established. The historical synthesis data consists of process parameter data and microsphere structural characteristic data corresponding to the preparation of several monodisperse polymer-based chromatographic microspheres by the high-shear emulsification polymerization execution module. The process parameter data includes high-shear emulsification speed, shear time, mass ratio of polymer monomer to porogen, reaction temperature program, concentration of functionalizing reagent and reaction time. The microsphere structural characteristic data includes the average particle size, particle size variation coefficient, average pore size, specific surface area and surface functional group density of the microspheres. Based on historical databases, an ensemble learning model based on machine learning is trained; wherein the input of the ensemble learning model is the process parameter data of the chromatographic microspheres, and the output is the microsphere structural feature data corresponding to the process parameter data; Obtain the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized, perform reverse reasoning calculations based on the integrated learning model, and output the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized. Based on the theoretical process parameters, the high-shear emulsification polymerization execution module is used to perform the synthesis steps to synthesize monodisperse polymer-based chromatographic microspheres; wherein, the high-shear emulsification polymerization execution module includes at least one high-shear emulsification reactor; The integrated learning model is a polymer-based chromatographic microsphere process-structure mapping network, which includes a feature embedding layer, a base learner group layer, a dynamic weight fusion layer and a microsphere structure feature output layer connected in series, as well as an uncertainty estimation module connected in parallel with the base learner group layer. The feature embedding layer is used to normalize continuous variables in the input process parameter data, perform time-series encoding of the reaction temperature program, perform embedding mapping of categorical variables, generate explicit feature crosses, and output a dense vector. The base learner layer includes several base learner sub-models, each of which predicts a microsphere structure feature; each sub-model shares the underlying feature representation and introduces physical prior constraint penalties during training; The uncertainty estimation module is used to predict the mean and 95% confidence interval of each microsphere structural feature based on the Monte Carlo Dropout output. The dynamic weight fusion layer includes a gating network, which is used to adaptively adjust the fusion weights output by each of the sub-models based on the input process parameter data. The microsphere structure feature output layer is used to output the microsphere structure feature data corresponding to the process parameter data.

2. The intelligent manufacturing method for monodisperse polymer-based chromatographic microspheres according to claim 1, characterized in that, The feature embedding layer performs time-series encoding of the reaction temperature program, including: The reaction temperature program is split into a heating rate sequence, a holding time sequence, and a target temperature sequence. A one-dimensional convolutional layer is used to extract the local temporal patterns of the heating rate sequence, the holding time sequence, and the target temperature sequence, and then global average pooling is used to extract the temperature feature vector.

3. The intelligent manufacturing method for monodisperse polymer-based chromatographic microspheres according to claim 1, characterized in that, The ensemble learning model employs a multi-task loss function to optimize target prediction error during training; the formula for calculating the multi-task loss function is as follows: L total = Σ k ( α k * MSE k )+ λ physics * L physics + λ reg * L reg In the formula, L total Represents the total loss function; k Indicates the first k Individual microsphere structural features; MSE k Indicates the first k Mean square prediction error of individual microsphere structural features; α k For the first k The fusion weights of individual microsphere structural features are determined based on the historical variation coefficients of the microsphere structural features. CV k set up, α k = 1 / CV k ; L physics As a physical constraint penalty term, calculate the degree of violation of physical trends; L reg This is an L2 regularization term to prevent overfitting; λ physics , λ reg These are hyperparameters, with values ​​of 0.1 and 0.001 respectively.

4. The intelligent manufacturing method for monodisperse polymer-based chromatographic microspheres according to claim 1, characterized in that, The process of obtaining the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized, performing reverse reasoning calculations based on the integrated learning model, and outputting the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized is as follows: Based on the uncertainty estimation module, a coarse-grained scan of the overall process parameter space is performed, and contour lines of the predicted standard deviation of each microsphere structural feature are drawn. Based on the predicted standard deviation contour lines, the range of process parameters with predicted standard deviations lower than a preset threshold is selected as the feasible region. Within the feasible region, an optimizer combining Bayesian optimization and covariance matrix adaptive evolution strategy is employed. The objective is to minimize the weighted Euclidean distance between the predicted microsphere structural feature data and the target microsphere structural feature data, and to search for the optimal process parameters, which are the theoretical process parameters. Specifically, the optimizer uses a Gaussian process surrogate model and an expected improvement acquisition function for iteration during the Bayesian optimization phase. When the Bayesian optimization converges and stalls or reaches the maximum number of iterations, it switches to the covariance matrix adaptive evolution strategy for fine-grained search. Output theoretical process parameters and their corresponding prediction confidence intervals and uncertainty scores to guide production.

5. The intelligent manufacturing method for monodisperse polymer-based chromatographic microspheres according to claim 1, characterized in that, Also includes: Dynamic data of the reaction process of the high-shear emulsification polymerization execution module are collected, and the high-shear emulsification speed and shear time are dynamically fine-tuned based on the dynamic data; wherein, the dynamic data is the droplet size of the emulsion after high-shear emulsification.

6. A smart manufacturing system for monodisperse polymer-based chromatographic microspheres, characterized in that, include: The acquisition and establishment module is used to acquire historical synthesis data and establish a historical database. The historical synthesis data consists of process parameter data and microsphere structural characteristic data corresponding to the preparation of several monodisperse polymer-based chromatographic microspheres by the high-shear emulsification polymerization execution module. The process parameter data includes high-shear emulsification speed, shear time, mass ratio of polymer monomer to porogen, reaction temperature program, concentration of functionalizing reagent and reaction time. The microsphere structural characteristic data includes the average particle size, particle size variation coefficient, average pore size, specific surface area and surface functional group density of the microspheres. The model training module is used to train an ensemble learning model based on machine learning using a historical database; wherein the input of the ensemble learning model is the process parameter data of the chromatographic microspheres, and the output is the microsphere structural feature data corresponding to the process parameter data; The reverse reasoning module is used to obtain the target microsphere structural feature data of the monodisperse polymer-based chromatographic microspheres to be synthesized, perform reverse reasoning calculations based on the integrated learning model, and output the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized. A high-shear emulsification polymerization execution module includes at least one high-shear emulsification reactor for synthesizing monodisperse polymer-based chromatographic microspheres based on the theoretical process parameters. The ensemble learning model trained by the model training module is a polymer-based chromatographic microsphere process-structure mapping network, which includes a feature embedding layer, a base learner group layer, a dynamic weight fusion layer and a microsphere structure feature output layer connected in series, as well as an uncertainty estimation module connected in parallel with the base learner group layer. The feature embedding layer is used to normalize continuous variables in the input process parameter data, perform time-series encoding of the reaction temperature program, perform embedding mapping of categorical variables, generate explicit feature crosses, and output a dense vector. The base learner layer includes several base learner sub-models, each of which predicts a microsphere structure feature; each sub-model shares the underlying feature representation and introduces physical prior constraint penalties during training; The uncertainty estimation module is used to predict the mean and 95% confidence interval of each microsphere structural feature based on the Monte Carlo Dropout output. The dynamic weight fusion layer includes a gating network, which is used to adaptively adjust the fusion weights output by each of the sub-models based on the input process parameter data. The microsphere structure feature output layer is used to output the microsphere structure feature data corresponding to the process parameter data.

7. The intelligent manufacturing system for monodisperse polymer-based chromatographic microspheres according to claim 6, characterized in that, The reverse reasoning module performs reverse reasoning calculations based on the ensemble learning model and outputs the theoretical process parameters of the monodisperse polymer-based chromatographic microspheres to be synthesized. This process execution unit includes: The scanning and filtering unit is used to perform coarse-grained scanning of the overall process parameter space based on the uncertainty estimation module, draw the predicted standard deviation contour lines of each microsphere structural feature, and filter the process parameter range with the predicted standard deviation below a preset threshold as the feasible region based on the predicted standard deviation contour lines. An optimization search unit is used to search for optimal process parameters (theoretical process parameters) within the feasible region by employing a hybrid optimizer combining Bayesian optimization and covariance matrix adaptive evolution strategy. The objective is to minimize the weighted Euclidean distance between the predicted microsphere structure feature data and the target microsphere structure feature data. The optimizer iterates using a Gaussian process surrogate model and an expected improvement acquisition function during the Bayesian optimization phase. When the Bayesian optimization converges and stalls or reaches the maximum number of iterations, it switches to the covariance matrix adaptive evolution strategy for a refined search. The output unit is used to output theoretical process parameters and their corresponding prediction confidence intervals and uncertainty scores to guide production.

8. The intelligent manufacturing system for monodisperse polymer-based chromatographic microspheres according to claim 6, characterized in that, The high-shear emulsification polymerization execution module is used to precisely execute the emulsification, polymerization and surface functionalization reactions of microspheres, and also includes: a precision metering feeding unit, a programmable temperature-controlled polymerization reactor and an automatic surface functionalization reaction unit; The precision metering and feeding unit is used to precisely control the addition ratio and timing of the aqueous phase and oil phase according to the theoretical process parameters. The programmed temperature-controlled polymerization reactor is used for multi-stage precise heating, holding, and cooling processes of microsphere emulsification and polymerization. The automated surface functionalization reaction unit is used to perform surface functionalization reactions based on the surface functional group density in the target microsphere structural feature data.

9. The intelligent manufacturing system for monodisperse polymer-based chromatographic microspheres according to claim 6, characterized in that, Also includes: The first data feedback module is used to collect the actual microsphere structural feature data of the product microspheres synthesized by the high-shear emulsification polymerization execution module, and feed the actual microsphere structural feature data back to the integrated learning model so that the integrated learning model can optimize the model by combining the actual microsphere structural feature data. The first data feedback module is connected online to the high-shear emulsification polymerization execution module and the model training module, and includes: an automatic sampling and conveying device, and a product characterization unit used in conjunction with the automatic sampling and conveying device; the automatic sampling and conveying device is used to automatically sample the product microspheres and convey them to the product characterization unit; the product characterization unit includes a scanning electron microscope, a specific surface area and pore size analyzer, and an infrared spectroscopy, used to determine the final particle size, particle size distribution, morphology, specific surface area, pore size distribution, and infrared spectrum of the product microspheres. The second data feedback module is connected online to the high-shear emulsification polymerization execution module. It is used to collect dynamic data of the reaction process of the high-shear emulsification polymerization execution module and dynamically fine-tune the high-shear emulsification speed and shear time based on the dynamic data. The dynamic data is the droplet size of the emulsion after high-shear emulsification.

10. The application of monodisperse polymer-based chromatographic microspheres prepared by the intelligent manufacturing method according to any one of claims 1-5 or monodisperse polymer-based chromatographic microspheres prepared by the intelligent manufacturing system according to any one of claims 6-9 in reversed-phase chromatography, normal-phase chromatography, hydrophilic interaction chromatography, ion exchange chromatography, size exclusion chromatography and affinity chromatography.

Images