Proportioning optimization method of nano-composite flame-retardant master batch in engineering plastic based on machine learning

Abstract

The invention belongs to the technical field of crossing of high polymer materials and artificial intelligence, and discloses a method for optimizing the proportion of a nano-composite flame-retardant master batch in engineering plastic based on machine learning. The method is used for solving the problems that a traditional experience trial and error method is low in efficiency in multi-target performance collaborative optimization, high-dimensional nonlinear parameter space search is difficult, and microscopic coupling effect modeling is missing. According to the method, a multi-dimensional input feature set containing material composition and process parameters is constructed, six key performance indexes are combined to form training data, a mixed model formed by cascading a multi-layer perceptron and Gaussian process regression is trained, an improved non-dominated sorting genetic algorithm (NSGA-II) is adopted for multi-target global optimization, and an optimal ratio meeting preset constraints is obtained. The method is an industrial internet technology system service, and solves the technical problem that multi-target performance is difficult to collaboratively optimize due to the fact that a traditional experience trial-and-error method cannot perform modeling and searching on a high-dimensional nonlinear parameter space of a nano-composite flame-retardant system.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of polymer materials and artificial intelligence, specifically involving a method for optimizing the proportion of nanocomposite flame retardant masterbatch in engineering plastics based on machine learning. Background Technology

[0002] Engineering plastics, due to their excellent mechanical strength, thermal stability, and processability, have been widely used in high-end industries with stringent material performance requirements, such as electronics, automotive manufacturing, and aerospace. In these applications, materials not only need to withstand complex mechanical loads and high-temperature environments but also must meet increasingly stringent fire safety regulations. Therefore, multi-dimensional and highly synergistic comprehensive requirements are placed on their flame retardant properties, mechanical properties, and processing rheological properties. Nanocomposite flame retardant masterbatches, as a novel functional additive, are gradually becoming a key technological path for improving the overall performance of engineering plastics due to their high specific surface area, excellent interfacial compatibility, and ability to achieve high-efficiency flame retardancy with low addition amounts. These masterbatches typically contain nanoscale flame retardant components such as layered double hydroxides, nanoclay, and carbon nanotubes, and are composited with a carrier resin through surface modification to improve their dispersion and interfacial bonding in the matrix, thereby maintaining or even enhancing the mechanical and thermal properties of the material while inhibiting combustion.

[0003] Currently, the formulation design of engineering plastic composite systems mainly relies on empirical trial-and-error methods or orthogonal experimental designs based on finite factorial methods. Patent CN111518391B proposes a polyphenylene sulfide resin composition that optimizes the material's weather resistance and overall mechanical properties by setting a weight percentage range for polyphenylene sulfide resin, polyamide resin, glass fiber, and various additives, and utilizing the synergistic effects between components (such as the interaction between polyamide resin and stress-relieving agents and anti-aging agents). This approach effectively improves the material's environmental adaptability in specific application contexts, and its formulation strategy is based on long-term engineering experience with the material system, possessing certain practical value. Patent CN102886880B also limits the ratio range (15%–40%) of engineering plastics and glass fiber in the one-time molding process of spiral wire guides, and combines mold structure and injection molding process parameter control to ensure product dimensional accuracy. The above-mentioned technical solutions demonstrate a deep understanding of material proportioning and process synergy in their respective application scenarios, reflecting the effectiveness of traditional material development paradigms in solving local performance optimization problems.

[0004] As the requirements for material performance in high-end manufacturing continue to upgrade, especially the need for simultaneous optimization of multiple objectives such as flame retardant efficiency, mechanical strength, heat distortion temperature, and melt flow rate, the traditional proportioning methods relying on experience and limited experiments are gradually revealing their inherent limitations. Nanocomposite flame retardant systems themselves possess highly nonlinear, strongly coupled, and multi-scale characteristics. On the one hand, the dispersion state, interfacial interactions, and physicochemical behavior of nano flame retardants during combustion are closely related to their content in the matrix, surface modification methods, and compatibility with other additives. On the other hand, the improvement of flame retardant properties is often accompanied by the deterioration of mechanical properties or the decrease in processing fluidity, and there is a complex inverse relationship among the three.

[0005] Traditional methods lack the ability to efficiently model and actively explore the nonlinear mapping relationships between variables in a high-dimensional parameter space. This makes it difficult to capture the global optimal solution within an acceptable experimental cost and cycle, and even more difficult to dynamically adjust the proportioning strategy for different end-use scenarios (such as thin-walled injection molded parts with high flowability requirements and structural parts with high rigidity requirements).

[0006] Existing technologies generally neglect the multi-physics coupling effect of nano-flame retardant systems at the microscale; for example, the orientation behavior of nanoparticles in shear fields, the kinetics of carbon layer formation during pyrolysis and their regulatory effect on heat and mass transfer. Although these effects are not directly reflected in the macroscopic proportions, they profoundly affect the overall performance of the final material.

[0007] Therefore, in the absence of data-driven intelligent modeling methods, even if a local "better" ratio is obtained through a large number of experiments, it is difficult to guarantee the consistency and transferability of its performance in different batches and under different process conditions.

[0008] The core technical bottleneck in optimizing the formulation of nanocomposite flame retardant masterbatches in engineering plastics lies in how to overcome the limitations of traditional empirical methods in searching high-dimensional nonlinear parameter spaces while taking into account flame retardant performance, mechanical properties and processing performance, and achieve systematic modeling and intelligent decision-making for multi-scale and multi-physics coupling effects. Summary of the Invention

[0009] This invention discloses a machine learning-based method for optimizing the proportion of nanocomposite flame retardant masterbatch in engineering plastics. It solves the technical problem that traditional empirical trial-and-error methods are unable to model and search the high-dimensional nonlinear parameter space of nanocomposite flame retardant systems, which makes it difficult to coordinate the optimization of multi-objective performance (flame retardancy, mechanical properties, and processing properties).

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A machine learning-based method for optimizing the proportion of nanocomposite flame-retardant masterbatch in engineering plastics includes the following steps: Establish a multi-dimensional input feature set, which includes the type of engineering plastic matrix, the proportion of nanocomposite flame retardant masterbatch, the type and content of auxiliary additives, processing temperature, screw speed and injection pressure; Based on the multidimensional input feature set, composite material samples are prepared through a standardized sample preparation process, and six performance indicators of the composite material samples—limiting oxygen index, vertical burning rating, tensile strength, flexural modulus, heat distortion temperature, and melt flow rate—are measured as output response variables. The multidimensional input feature set and the output response variable constitute a training dataset, and a hybrid machine learning model is trained using the training dataset. The hybrid machine learning model is used to establish a mapping relationship between the multidimensional input feature set and the output response variable. Based on the hybrid machine learning model, a multi-objective optimization function is constructed, and an improved non-dominated sorting genetic algorithm is used to globally optimize the proportioning parameters to obtain the optimal proportioning scheme that meets the preset performance constraints.

[0011] Furthermore, the preparation method of the nanocomposite flame retardant masterbatch includes: Layered double hydroxides, organically modified montmorillonite, and surface-functionalized carbon nanotubes were compounded at a mass ratio of 1:0.8-1.2:0.3-0.5 to obtain nano flame-retardant fillers. The nano-flame retardant filler and carrier resin are melt-blended at 220-280°C using a twin-screw extruder to obtain the nano-composite flame retardant masterbatch. The layered double hydroxide has an average particle size of 50-100 nm; the organically modified montmorillonite has an interlayer spacing of not less than 2.5 nm; the surface-functionalized carbon nanotubes have an aspect ratio of not less than 200, and their surfaces are grafted with phosphorus- or nitrogen-containing functional groups via acylation reactions, the functional groups being selected from amino (-N) groups. ), phosphate ester group (-PO(OR) R is a C1-C4 alkyl group or a melamine group, the grafting density is 5-15 wt%, and the carbon nanotubes are dispersed in formamide at a concentration of 1 mg / mL without precipitation after standing for 24 hours. The carrier resin is one of polyphenylene sulfide, polyamide 66, or polycarbonate. The nano flame-retardant filler has a mass fraction of 15%-30% in the nanocomposite flame-retardant masterbatch, and the carrier resin has a mass fraction of 70%-85%.

[0012] Furthermore, the multidimensional input feature set is a vector of fixed dimensions; The types of engineering plastic matrices are represented by three-dimensional unique thermal encoding vectors; The proportion of nanocomposite flame retardant masterbatch added is input as a scalar value. The characteristic representation method of the auxiliary agent is as follows: A set of J possible adjuvant components is predefined. For each adjuvant component, two sub-features and a binary indicator variable are used for representation. and a continuous variable This indicates its mass percentage content; the multidimensional input feature set contains all J types of excipients. right; The processing temperature, screw speed and injection pressure are input as scalar values; All input features are Z-score normalized before entering the hybrid machine learning model.

[0013] Furthermore, the construction process of the training dataset includes two stages: initial dataset construction and active learning augmentation. The initial dataset was generated by sampling within the feasible region of the multidimensional input feature set using an optimal Latin hypercube design, with a sample size of [number missing]. satisfy Preliminary experiments have verified that this sample size range can provide a reliable initial data foundation for subsequent model training while controlling experimental costs. For each set of sampling points in the initial dataset, a composite material sample is prepared using a standardized sample preparation process and tested three times. The average value of the six performance indicators is taken as the output response variable of the sample. The active learning augmentation phase, based on the Gaussian process regression component of the trained hybrid machine learning model, samples new experimental points in regions of high prediction uncertainty or near the Pareto front in the feature space. The number of new sample points M in each iteration is typically set to the initial sample size. The data is allocated to 5%-10% of the model update efficiency to balance experimental costs; after the new experimental points are verified by experiments, their data are added to the training dataset to iteratively update the model.

[0014] Furthermore, the hybrid machine learning model is composed of a cascaded multilayer perceptron and a Gaussian process regression model; The multilayer perceptron has three hidden layers, with the number of neurons in each hidden layer being 128, 64, and 32 respectively. The activation function is a modified linear unit, and the output layer uses a linear activation function. The covariance function of the Gaussian process regression model is the Matrn kernel function, and its hyperparameters are determined by maximizing the marginal likelihood function. The output of the multilayer perceptron is used as the input of the Gaussian process regression model. The two are coupled through a residual connection mechanism. The Gaussian process regression model fits the residual between the predicted value and the true value of the multilayer perceptron. The final prediction result is the sum of the output of the multilayer perceptron and the output of the Gaussian process regression model.

[0015] Furthermore, the initial value of the noise variance of the Gaussian process regression model is set to 0.01, and it is dynamically decayed with the number of incremental learning iterations, with a decay coefficient of 0.95. The weights of the multilayer perceptron are initialized using a uniform Xavier distribution, and the bias term is initialized to zero. The loss function of the hybrid machine learning model is a weighted sum of mean squared error and mean absolute error, with a weight ratio of 2:1.

[0016] Furthermore, the multi-objective optimization function is defined as a six-dimensional vector function. Where x is the matching parameter vector, This is a predicted value for the limiting oxygen index. The numerical rating corresponds to the vertical flammability rating, where V-0 corresponds to 3, V-1 to 2, V-2 to 1, and Fail to 0. This is the predicted value of tensile strength. This is the predicted value of the flexural modulus. This is the predicted value for heat distortion temperature. This is a predicted value for the melt flow rate; During the training process of the hybrid machine learning model, to handle discrete vertical burning levels, ordinal regression or well-fit coding methods are used to convert the level labels (V-0, V-1, V-2, Fail) into continuous numerical labels for model training; this ensures that the model's actual predictions are accurate. It is a continuous score that reflects the trend of a sample to reach a higher flame retardancy rating, and its optimization goal is to maximize the value (i.e., a higher flame retardancy rating). The optimization objective is to maximize At the same time Controls are placed within the preset application window.

[0017] Furthermore, the improved non-dominated sorting genetic algorithm (NSGA-II) introduces a dynamic crowding distance calculation mechanism and an elite retention strategy on the basis of the traditional NSGA-II algorithm; The population size is set at 200, the maximum number of generations is 100, the crossover probability is 0.9, and the mutation probability is 0.1. In each generation of non-dominated solution set, a performance constraint threshold is set according to the application scenario requirements, and only individuals that meet all constraints are retained to enter the next generation. The performance constraint thresholds include: limiting oxygen index not less than 28%, vertical burning rating not less than V-1, tensile strength not less than 80 MPa, flexural modulus not less than 3000 MPa, heat distortion temperature not less than 220°C, and melt flow rate between 8 and 25 g / 10 min.

[0018] Furthermore, the improved Non-Dominated Sorting Genetic Algorithm (NSGA-II) introduces a performance stability index in its convergence determination. This performance stability index is defined as the performance stability index of the objective function values ​​in the most recent ten generations of non-dominated solution sets, after... The sum of standard deviations after interval normalization; When the performance stability index is less than the preset tolerance of 0.005 for five consecutive generations, the algorithm is determined to have converged and the evolution is terminated.

[0019] Furthermore, the present invention also includes an online feedback correction step: The measured values ​​of the melt flow rate of the product are collected in real time by melt flow rate meters and process sensors deployed on the production line. Processing temperature and injection pressure ; The collected process parameters are input into the hybrid machine learning model to obtain the corresponding predicted melt flow rate. ; Calculate K consecutive production batches ( The mean relative deviation of melt flow rate : ; when When this is triggered, the model's incremental learning mechanism is activated, incorporating recently collected data... The data pairs are added to the training dataset to perform local retraining on the Gaussian process regression part of the hybrid machine learning model.

[0020] Furthermore, the equipment used to determine the output response variable includes a cone calorimeter, a universal testing machine, a thermomechanical analyzer, and a melt flow rate meter; In the online feedback correction step, the deviation signal is smoothed by a low-pass filter and then input to the model update interface; The incremental learning mechanism employs a Bayesian online update algorithm to ensure that the adjustment of model parameters does not exceed twice the standard deviation of the statistical characteristics of the original training set.

[0021] Furthermore, the preset application window and the performance constraint threshold are automatically switched for different application scenarios: for electronic and electrical appliance housing application scenarios, the preset application window for melt flow rate is set to 15–25 g / 10 min; for automotive engine peripheral component application scenarios, the constraint threshold for heat distortion temperature is increased to 240°C, and the constraint threshold for tensile strength is increased to 90 MPa; for aerospace interior component application scenarios, the constraint threshold for vertical combustion rating is forcibly set to V-0 level.

[0022] Furthermore, the proportion of nanocomposite flame-retardant masterbatch added in the optimal formulation is a continuous value, rounded to the nearest 0.5% multiple as the final instruction value; the final instruction value is executed through a loss-in-weight feeding system, with feeding accuracy controlled within ±0.2%; the hybrid machine learning model is deployed on an edge computing node equipped with a dedicated neural network acceleration chip, with an inference latency of less than 50 milliseconds; all experimental data, model parameters, and optimization results are stored in a blockchain distributed ledger, with each data block containing a timestamp, operator identifier, device number, and hash checksum to ensure data immutability and full-process traceability.

[0023] Compared with the prior art, the present invention has the following beneficial effects: This invention achieves precise, efficient, and adaptive optimization of the proportion of nanocomposite flame retardant masterbatch in engineering plastics by constructing a systematic technical solution that integrates high-dimensional nonlinear mapping modeling, multi-objective intelligent optimization, and online feedback correction. It effectively overcomes the fundamental defects of traditional empirical methods in multi-scale coupling effect modeling and global optimal solution search, and provides a reliable technical path for the intelligent formulation development of high-end engineering plastics. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the overall process of the ratio optimization method described in this invention.

[0026] Figure 2 This is a flowchart of the training process for the hybrid machine learning model of this invention.

[0027] Figure 3 This is a flowchart of the multi-objective optimization process of the present invention. Detailed Implementation

[0028] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0029] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] Example 1: See Figures 1-3 This embodiment discloses a machine learning-based method for optimizing the proportion of nanocomposite flame retardant masterbatch in engineering plastics. It constructs a data-driven intelligent optimization framework that integrates the complex mapping relationship between material composition variables, processing parameters and multi-objective performance indicators, and on this basis, achieves precise control and adaptive decision-making of the proportion of nanocomposite flame retardant masterbatch in engineering plastics.

[0031] The technical solution of the present invention will be described in detail below with reference to specific embodiments.

[0032] In the preparation stage of nanocomposite flame retardant masterbatch, layered double hydroxide, organic modified montmorillonite and surface functionalized carbon nanotubes are compounded at a mass ratio of 1:0.8-1.2:0.3-0.5 to obtain nano flame retardant filler.

[0033] The average particle size of the layered double hydroxide is controlled within the range of 50-100 nanometers to ensure good dispersibility and interfacial interaction in the matrix; the interlayer spacing of the organically modified montmorillonite is not less than 2.5 nanometers to facilitate the insertion of polymer segments and the formation of intercalation or exfoliation structures; the aspect ratio of the surface-functionalized carbon nanotubes is not less than 200, and their surfaces are grafted with phosphorus- or nitrogen-containing functional groups to enhance their compatibility with the polymer matrix and synergistically improve flame retardant efficiency.

[0034] The nano-flame-retardant filler and carrier resin are melt-blended using a twin-screw extruder at 220-280°C to obtain the nano-composite flame-retardant masterbatch. The carrier resin is one of polyphenylene sulfide, polyamide 66, or polycarbonate. The mass fraction of the nano-flame-retardant filler in the nano-composite flame-retardant masterbatch is 15%-30%, and the mass fraction of the carrier resin is 70%-85%. This nano-composite flame-retardant masterbatch is used as a functional additive for subsequent formulation optimization.

[0035] During the input feature set construction phase, the definition of multidimensional input variables includes: The types of engineering plastic matrix, the proportion of nanocomposite flame retardant masterbatch, the types and contents of auxiliary additives, processing temperature, screw speed, and injection pressure.

[0036] The engineering plastic matrix type is represented using a unique thermal coding method, corresponding to three categories: polyphenylene sulfide, polyamide 66, and polycarbonate. The proportion of nanocomposite flame-retardant masterbatch is directly input as a mass percentage, ranging from 5% to 25%. Auxiliary additives are selected from categories such as antioxidants, lubricants, nucleating agents, and compatibilizers. To ensure a fixed input feature dimension, a predefined set of J=12 fixed auxiliary agent components is defined, selected from the following four categories, each containing three specific substances: (1) Antioxidants: Antioxidant 1010, Antioxidant 168, Antioxidant 1076; (2) Lubricants: calcium stearate, ethylene bis-stearamide (EBS), polyethylene wax; (3) Nucleating agents: talc, sodium benzoate, silicon dioxide; (4) Compatibilizers: maleic anhydride grafted polypropylene (PP-g-MAH), ethylene-acrylate-maleic anhydride terpolymer (E-AA-MAH), styrene-maleic anhydride copolymer (SMA). For each of the 12 adjuvant components, two sub-features are used: a binary indicator variable. (1 represents adding, 0 represents not adding) and a continuous variable This indicates its mass percentage content; the multidimensional input feature set contains all 12 excipients. Yes, this constitutes a fixed 24-dimensional input feature vector; for any specific formulation, the adjuvants used must be a subset of this predefined set.

[0037] All input features are Z-score standardized before entering the model. This involves calculating the mean and standard deviation of each feature dimension and then centering and normalizing it to eliminate the interference of different units on the model training process.

[0038] In the training dataset construction phase, the initial dataset is first constructed. An optimal Latin hypercube design is employed, and 300 initial experimental points are generated by uniform sampling within a multidimensional feasible domain comprised of engineering plastic matrix types (3 types), nanocomposite flame retardant masterbatch addition ratios (5%-25% continuous range), auxiliary additives (4 predefined combinations), processing temperature (250-280℃), screw speed (200-400rpm), and injection pressure (80-120MPa).

[0039] For each set of experimental points, composite material samples are prepared according to the following parameter-adjustable sample preparation process, the core parameters of which are determined by the sampling points: (1) Mixing equipment and parameters: A co-rotating twin-screw extruder is used, with a screw diameter of 35mm and a length-to-diameter ratio of L / D=40. The screw configuration from the feed port to the die head is as follows: conveying section (10D) - kneading block (5D, 30° staggered angle) - conveying section (5D) - kneading block (5D, 60° staggered angle) - conveying section (10D) - vacuum devouring section (5D) - metering section (5D); The barrel temperature zones are set to 7 zones, from the feed port to the die head: 220℃, 250℃, 260℃, 265℃, ℃, 275℃, 270℃, among which The input value represents the "processing temperature" at this sampling point; the screw speed is set to the input value of the "screw speed" at this sampling point, denoted as... rpm, feeding rate 20kg / h, vacuum degree -0.08MPa.

[0040] (2) Injection molding parameters: The extruded and granulated material was dried in a 100℃ forced-air drying oven for 4 hours until the moisture content was <0.02%; a horizontal injection molding machine was used, with a clamping force of 800kN and the injection temperature set to... ℃ (i.e., consistent with the temperature of zone 5 of the mixing process), the injection pressure is set to the "injection pressure" input value of this sampling point, denoted as . MPa, holding pressure 80MPa, holding time 20 seconds, cooling time 30 seconds, mold temperature 80℃.

[0041] (3) Sample post-treatment: The standard test specimens of injection molding are placed in a constant temperature and humidity environment of 23±2℃ and 50±5% relative humidity for more than 48 hours to eliminate internal stress and moisture absorption effects.

[0042] (4) Performance testing standards: Limiting oxygen index is tested according to ASTM D2863, vertical flammability rating is tested according to UL94, tensile strength is tested according to ISO527-2 (tensile rate 50 mm / min), flexural modulus is tested according to ISO178, heat distortion temperature is tested according to ISO75-2 (load 1.8 MPa), melt flow rate is tested according to ISO1133 (test conditions are consistent with the corresponding standard conditions of the base resin).

[0043] Samples were prepared for each sampling point according to this procedure, and the preparation and testing were repeated three times. The average value of six performance indicators, including limiting oxygen index and vertical combustion rating, was taken as the output response variable for that sample, forming the initial training dataset. .

[0044] Three independent batches of samples were prepared for each formulation. All performance indicators of each batch were tested three times. After removing outliers, the arithmetic mean was taken as the final output response variable for that sample.

[0045] After obtaining the initial model, the active learning expansion phase begins. Using the Gaussian process regression component of the hybrid machine learning model, the prediction uncertainty of the unexperimented regions in the feature space is calculated. The M points with the highest prediction uncertainty are selected (in this embodiment, M=20, approximately the initial sample size). 6.7% of the data (this percentage is based on an empirical balance between model update efficiency and the cost of new experiments) or regions close to the current Pareto front but with significant model prediction discrepancies are selected as new experimental points for validation. The new experimental data is added to the training set to update the model. This process can be iterated until the model's prediction accuracy on the independent validation set meets a preset threshold (e.g., root mean square error). (If the performance is less than 5% of the target performance range or the experimental budget is exhausted.)

[0046] Furthermore, a hybrid machine learning model is constructed to fit the nonlinear mapping relationship between input features and output response. The hybrid machine learning model consists of a two-level structure: The first stage is a multilayer perceptron (MLP) with three hidden layers. The number of neurons in each hidden layer is 128, 64, and 32 respectively. The activation function is the modified linear unit (ReLU), and the output layer uses a linear activation function to ensure the continuity of the predicted values.

[0047] The second stage is a Gaussian process regression (GPR) model, whose covariance function uses the Matérn kernel, and the hyperparameters are determined by maximizing the marginal likelihood function. The two-stage models are coupled through a residual connection mechanism: the multilayer perceptron first performs a preliminary nonlinear mapping on the input features and outputs a preliminary predicted value; the Gaussian process regression model then uses this preliminary predicted value as input to fit the residual between it and the true performance index; the final prediction result is the sum of the multilayer perceptron output and the Gaussian process regression output.

[0048] In a preferred embodiment of the present invention, the initial noise variance of the Gaussian process regression model is set to 0.01 and dynamically decays with the number of incremental learning iterations, with a decay coefficient of 0.95; the weights of the multilayer perceptron are initialized using a Xavier uniform distribution, and the bias term is initialized to zero; the loss function is a weighted sum of mean squared error (MSE) and mean absolute error (MAE), with a weight ratio of 2:1, to balance large error penalty and overall bias control.

[0049] Based on this, a multi-objective optimization function is constructed and global optimization is implemented.

[0050] The multi-objective optimization function is defined as a six-dimensional vector function. Where x is the matching parameter vector, This is a predicted value for the limiting oxygen index. The numerical rating corresponds to the vertical flammability rating, where V-0 corresponds to 3, V-1 to 2, V-2 to 1, and Fail to 0. This is the predicted value of tensile strength. This is the predicted value of the flexural modulus. This is the predicted value for heat distortion temperature. This is a predicted value for the melt flow rate; During the training process of the hybrid machine learning model, to handle discrete vertical burning levels, ordinal regression or well-fit coding methods are used to convert the level labels (V-0, V-1, V-2, Fail) into continuous numerical labels for model training; this ensures that the model's actual predictions are accurate. It is a continuous score that reflects the trend of a sample to reach a higher flame retardancy rating, and its optimization goal is to maximize the value (i.e., a higher flame retardancy rating). The optimization objective is to maximize At the same time Controls are placed within the preset application window.

[0051] An improved non-dominated sorting genetic algorithm (improved NSGA-II algorithm) was used to solve the problem. The population size was set to 200, the maximum number of generations was 100, the crossover probability was 0.9, and the mutation probability was 0.1.

[0052] In each generation of non-dominated solution set, a performance constraint threshold is set according to the application scenario requirements, and only individuals that meet all constraints are retained to enter the next generation.

[0053] The performance constraint thresholds include: limiting oxygen index not less than 28%, and vertical combustion rating not less than V-1 (i.e., ...). The tensile strength is not less than 80 MPa, the flexural modulus is not less than 3000 MPa, the heat distortion temperature is not less than 220℃, and the melt flow rate is between 8 and 25 g / 10 min.

[0054] As another preferred embodiment of the present invention, the improved non-dominated sorting genetic algorithm introduces a performance stability index in the convergence determination. This index is defined as the sum of the standard deviations of the objective function values ​​in the non-dominated solution set of the most recent ten generations; when this index is less than a preset tolerance for five consecutive generations... When the value is 0.005 (e.g., 0.005), the algorithm is considered to have converged and the evolution process is terminated.

[0055] Furthermore, the present invention also includes an online feedback correction module to maintain the predictive accuracy and robustness of the model in actual production environments. The online feedback correction module collects data in real time through melt flow rate meters deployed on the production line and temperature and pressure sensors embedded in the processing equipment. Melt flow rate (MFR) is used as the core monitoring indicator because it is quick to test and can comprehensively reflect the material rheological properties and process stability.

[0056] The module calculates K consecutive production batches ( Measured melt flow rate Predicted values ​​corresponding to the model Mean of relative deviation between : ; To suppress high-frequency measurement noise, a low-pass filter (e.g., a first-order Butterworth filter) with a cutoff frequency lower than the production cycle frequency is used to smooth the deviation signal before it is input to the model update interface. At that time, the incremental learning mechanism of the model is triggered, and the Gaussian process regression part is locally retrained using the newly collected data.

[0057] The incremental learning process employs a Bayesian online update algorithm to ensure that the adjustment range of model parameters does not exceed twice the standard deviation of the statistical characteristics of the original training set, thereby avoiding drastic fluctuations in the model due to a small amount of new data. The online feedback correction module is connected to the central control system via industrial Ethernet to achieve real-time data transmission and command issuance.

[0058] Furthermore, this invention automatically switches between preset application windows and performance constraint thresholds for different application scenarios.

[0059] In practical implementation, the electronic and electrical appliance casing scenario emphasizes high fluidity and flame retardancy, with a melt flow rate window set at 15–25 g / 10 min; the automotive engine peripheral component scenario emphasizes high heat distortion temperature and mechanical strength, with the heat distortion temperature threshold raised to 240°C and the tensile strength threshold raised to 90 MPa; the aerospace interior component scenario requires the highest flame retardancy rating, with the vertical flammability rating constrained to V-0 (i.e., The central control system automatically identifies the application scenario based on order information and loads the corresponding set of constraints to ensure that the optimization results meet the needs of end users.

[0060] During the proportioning execution phase, the nanocomposite flame-retardant masterbatch addition ratio output by the optimization algorithm is a continuous value, which needs to be rounded to the nearest 0.5% integer multiple before being used as the final instruction value. This final instruction value is sent to the loss-in-weight feeding system via the Modbus TCP protocol. The latter adjusts the feeding rate in real time according to the throughput of the main mixing equipment to ensure that the addition ratio accuracy is controlled within ±0.2%. The loss-in-weight feeding system maintains bidirectional communication with the central control system and can report the deviation between the actual feeding amount and the set value in real time for the system to perform closed-loop correction.

[0061] In terms of system deployment, the hybrid machine learning model is deployed on edge computing nodes equipped with dedicated neural network acceleration chips (such as NVIDIA Jetson AGX Orin or Huawei Ascend Atlas 500), with inference latency of less than 50 milliseconds, meeting the requirements for real-time allocation decision-making.

[0062] The model input interface supports the OPCUA protocol and can be directly connected to Enterprise Resource Planning (ERP) and Manufacturing Execution System (MES) to achieve end-to-end automated generation from material performance requirements in customer orders to specific proportioning instructions, without manual intervention.

[0063] In terms of data management, during the implementation of this invention, all experimental data, model parameters, and optimization results are stored in a blockchain distributed ledger. Each data block includes a timestamp, operator identifier, equipment number, and hash checksum to ensure data immutability and end-to-end traceability. This mechanism guarantees the traceability and compliance of the entire formulation development process, making it particularly suitable for the aerospace and medical fields, where material consistency requirements are stringent.

[0064] In one specific embodiment, the target application scenario is automotive engine peripheral components, requiring a heat distortion temperature of not less than 240°C, a tensile strength of not less than 90 MPa, a limiting oxygen index of not less than 28%, a vertical combustion rating of not less than V-1, and a melt flow rate between 8 and 15 g / 10 min.

[0065] The engineering plastic matrix was selected as polyamide 66, and the auxiliary additives were antioxidant 1010 (0.3%), lubricant EBS (0.2%), nucleating agent talc (1.0%), and compatibilizer maleic anhydride grafted polypropylene (2.0%).

[0066] Using the method described in this embodiment, after prediction by a hybrid model and optimization by the improved NSGA-II, the optimal ratio scheme is obtained: The nanocomposite flame retardant masterbatch was added at a ratio of 18.5%, the processing temperature was 270℃, the screw speed was 320rpm, and the injection pressure was 110MPa. Samples were prepared according to this scheme.

[0067] The measured performance is as follows: limiting oxygen index 29.1%, vertical flammability rating V-0, tensile strength 92.3 MPa, flexural modulus 3250 MPa, heat distortion temperature 243℃, melt flow rate 12.7 g / 10 min, all of which meet the constraints.

[0068] Comparative Example 1: As a comparative example, the traditional orthogonal experimental design method was used ( The formulation was optimized. Polyamide 66 was selected as the matrix, and the auxiliary additives were fixed as above. The following four factors were investigated: the addition ratio of nanocomposite flame retardant masterbatch (15%, 20%, 25%), processing temperature (250℃, 270℃, 290℃), screw speed (250rpm, 350rpm, 450rpm), and injection pressure (90MPa, 110MPa, 130MPa).

[0069] Nine sets of experiments were conducted, and the optimal performance was selected based on the following parameters: additive ratio 20%, processing temperature 270℃, screw speed 350rpm, and injection pressure 110MPa. The measured performance was: limiting oxygen index 28.5%, vertical flammability rating V-1, tensile strength 88.7MPa, flexural modulus 3100MPa, heat deflection temperature 238℃, and melt flow rate 13.2g / 10min. However, the tensile strength and heat deflection temperature did not meet the performance requirements for automotive engine peripheral components (90MPa and 240℃, respectively).

[0070] The performance data of Example 1 and Comparative Example 1 are summarized in Table 1 below: Table 1:

[0071] As shown in Table 1 above, the method proposed in this invention achieves a higher level of comprehensive performance while satisfying all performance constraints, especially in key indicators such as tensile strength and heat distortion temperature, which are significantly superior to traditional orthogonal experimental methods. This is attributed to the fact that the hybrid machine learning model constructed in this invention can effectively capture the complex coupling effects in the high-dimensional nonlinear parameter space, while the improved multi-objective optimization algorithm can efficiently search the global Pareto optimal solution set under multiple constraints.

[0072] Furthermore, the online feedback correction module played a crucial role in the continuous production process.

[0073] During a continuous production run of 50 batches, starting from batch 32, a change in raw material batches caused a drift in the material's rheological properties, resulting in the measured melt flow rate (MFR) being more than 5% lower than the model's predicted value for three consecutive batches (predicted 13.5 g / 10 min, measured 12.5-12.7 g / 10 min). The system automatically triggered an incremental learning mechanism, using newly acquired data including process parameters and measured MFR to locally retrain the Gaussian process regression model. After retraining, the model's prediction error for the melt flow rate of subsequent batches significantly decreased, and from batch 40 onwards, the deviation between the predicted and measured values ​​recovered to within 2%, effectively ensuring the stability of the processing technology.

[0074] This invention, by constructing a data-driven systematic optimization framework, achieves precise, efficient, and adaptive control of the proportion of nanocomposite flame-retardant masterbatch in engineering plastics, resulting in significant technological advancements and beneficial effects, specifically reflected in the following aspects: This invention fundamentally changes the traditional material development approach that relies on trial and error. By constructing a multi-dimensional input feature set including the type of engineering plastic matrix, the proportion of nanocomposite flame retardant masterbatch, auxiliary additives, and process parameters, and establishing a nonlinear mapping relationship between this feature set and six key performance indicators such as limiting oxygen index and vertical burning rating, it achieves, for the first time, a quantitative modeling of multi-scale, multi-physics coupling effects in a high-dimensional parameter space. This overcomes the inherent defect of traditional methods, which face difficulties in finding the global optimal solution due to their inability to analyze the interactions between complex variables.

[0075] The structural design of the hybrid machine learning model in this invention brings a substantial improvement in prediction accuracy and generalization ability. The hybrid model, composed of a cascaded multilayer perceptron and Gaussian process regression, combines the learning ability of deep neural networks for complex nonlinear patterns with the quantification ability of Gaussian processes for prediction uncertainty. Through a residual connection mechanism, the multilayer perceptron captures the main trend, while the Gaussian process regression fits the local residuals, enabling the model to obtain high-precision mapping relationships even with limited experimental data, providing a reliable surrogate model for subsequent optimization.

[0076] The improved non-dominated sorting genetic algorithm and online feedback correction mechanism of this invention jointly ensure the dynamic optimality and engineering practicality of the optimized solution. The improved non-dominated sorting genetic algorithm introduces dynamic crowding distance and a convergence criterion based on performance stability, significantly improving the Pareto front search efficiency and solution set quality under multi-objective and multi-constraint conditions. The online feedback correction step triggers incremental learning based on Bayesian updates by real-time monitoring of key process responses such as melt flow rate, enabling the hybrid machine learning model to adaptively compensate for raw material fluctuations and process drift, ensuring the performance consistency and stability of the optimized formula in large-scale continuous production.

[0077] This invention achieves closed-loop control from intelligent optimization to precise execution. The continuously optimized proportions are quantified and executed through a high-precision loss-in-weight feeding system. A hybrid machine learning model is deployed on low-latency edge computing nodes, supporting seamless integration with the manufacturing execution system, enabling end-to-end automated decision-making from performance requirements to production instructions. Combined with the tamper-proof and traceable data throughout the entire process guaranteed by blockchain technology, the system provides a complete technical foundation for compliant production and quality control of high-end engineering plastics.

[0078] This invention provides not only an optimization algorithm, but also a systematic solution that integrates materials science principles, data-driven modeling, multi-objective decision-making, and industrial automation. It effectively solves the core problem of synergistic optimization of multiple properties in nanocomposite flame retardant systems, and significantly improves the efficiency of formulation development, the comprehensive performance of materials, and the robustness of the production process.

[0079] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing the proportion of nanocomposite flame-retardant masterbatch in engineering plastics based on machine learning, characterized in that, Includes the following steps: Establish a multi-dimensional input feature set, which includes the type of engineering plastic matrix, the proportion of nanocomposite flame retardant masterbatch, the type and content of auxiliary additives, processing temperature, screw speed and injection pressure; Based on the multidimensional input feature set, composite material samples are prepared through a standardized sample preparation process, and six performance indicators of the composite material samples—limiting oxygen index, vertical burning rating, tensile strength, flexural modulus, heat distortion temperature, and melt flow rate—are measured as output response variables. The multidimensional input feature set and the output response variable constitute a training dataset, and a hybrid machine learning model is trained using the training dataset. The hybrid machine learning model is used to establish a mapping relationship between the multidimensional input feature set and the output response variable. Based on the hybrid machine learning model, a multi-objective optimization function is constructed, and an improved non-dominated sorting genetic algorithm is used to globally optimize the proportioning parameters to obtain the optimal proportioning scheme that meets the preset performance constraints.

2. The method for optimizing the proportion of nanocomposite flame retardant masterbatch in engineering plastics based on machine learning according to claim 1, characterized in that, The preparation method of the nanocomposite flame retardant masterbatch includes: Layered double hydroxides, organically modified montmorillonite, and surface-functionalized carbon nanotubes were compounded at a mass ratio of 1:0.8-1.2:0.3-0.5 to obtain nano flame-retardant fillers. The nano-flame retardant filler and carrier resin are melt-blended at 220-280°C using a twin-screw extruder to obtain the nano-composite flame retardant masterbatch. The layered double hydroxide has an average particle size of 50-100 nanometers; the interlayer spacing of the organically modified montmorillonite is not less than 2.5 nanometers; the aspect ratio of the surface-functionalized carbon nanotubes is not less than 200, and the surface is grafted with phosphorus- or nitrogen-containing functional groups through an acylation reaction, the functional groups being selected from amino groups. ), phosphate ester group ( R is a C1-C4 alkyl group or a melamine group, the grafting density is 5-15 wt%, and the carbon nanotubes are dispersed in formamide at a concentration of 1 mg / mL without precipitation after standing for 24 hours. The carrier resin is one of polyphenylene sulfide, polyamide 66, or polycarbonate. The nano flame-retardant filler has a mass fraction of 15%-30% in the nanocomposite flame-retardant masterbatch, and the carrier resin has a mass fraction of 70%-85%.

3. The method for optimizing the proportion of nanocomposite flame-retardant masterbatch in engineering plastics based on machine learning according to claim 1, characterized in that, The multidimensional input feature set is a vector of fixed dimensions; The types of engineering plastic matrices are represented by three-dimensional unique thermal encoding vectors; The proportion of nanocomposite flame retardant masterbatch added is input as a scalar value. The characteristic representation method of the auxiliary agent is as follows: A set of J possible adjuvant components is predefined. For each adjuvant component, two sub-features and a binary indicator variable are used for representation. and a continuous variable This indicates its mass percentage content; the multidimensional input feature set contains all J types of excipients. right; The processing temperature, screw speed and injection pressure are input as scalar values; All input features are Z-score normalized before entering the hybrid machine learning model.

4. The method for optimizing the proportion of nanocomposite flame retardant masterbatch in engineering plastics based on machine learning according to claim 1, characterized in that, The process of constructing the training dataset includes two stages: initial dataset construction and active learning augmentation. The initial dataset was generated by sampling within the feasible region of the multidimensional input feature set using an optimal Latin hypercube design, with a sample size of [number missing]. satisfy ; Preliminary experiments have verified that this sample size range can provide a reliable initial data foundation for subsequent model training while controlling experimental costs. For each set of sampling points in the initial dataset, a composite material sample is prepared using a standardized sample preparation process and tested three times. The average value of the six performance indicators is taken as the output response variable of the sample. The active learning augmentation phase, based on the Gaussian process regression component of the trained hybrid machine learning model, samples new experimental points in regions of high prediction uncertainty or near the Pareto front in the feature space. The number of new sample points M in each iteration is typically set to the initial sample size. 5%-10% to balance model update efficiency and experimental costs; After the newly added experimental points are verified by experiments, their data are added to the training dataset to iteratively update the model.

5. The method for optimizing the proportion of nanocomposite flame-retardant masterbatch in engineering plastics based on machine learning according to claim 1, characterized in that, The hybrid machine learning model is composed of a cascaded multilayer perceptron and a Gaussian process regression model. The multilayer perceptron has three hidden layers, with the number of neurons in each hidden layer being 128, 64, and 32 respectively. The activation function is a modified linear unit, and the output layer uses a linear activation function. The covariance function of the Gaussian process regression model is the Matrn kernel function, and its hyperparameters are determined by maximizing the marginal likelihood function. The output of the multilayer perceptron is used as the input of the Gaussian process regression model. The two are coupled through a residual connection mechanism. The Gaussian process regression model fits the residual between the predicted value and the true value of the multilayer perceptron. The final prediction result is the sum of the output of the multilayer perceptron and the output of the Gaussian process regression model.

6. The method for optimizing the proportion of nanocomposite flame retardant masterbatch in engineering plastics based on machine learning according to claim 5, characterized in that, The initial noise variance of the Gaussian process regression model is set to 0.01, and it decreases dynamically with the number of incremental learning iterations, with a decay coefficient of 0.

95. The weights of the multilayer perceptron are initialized using a uniform Xavier distribution, and the bias term is initialized to zero. The loss function of the hybrid machine learning model is a weighted sum of mean squared error and mean absolute error, with a weight ratio of 2:

1.

7. The method for optimizing the proportion of nanocomposite flame-retardant masterbatch in engineering plastics based on machine learning according to claim 1, characterized in that, The multi-objective optimization function is defined as a six-dimensional vector function. Where x is the matching parameter vector, This is a predicted value for the limiting oxygen index. The numerical rating corresponds to the vertical flammability rating, where V-0 corresponds to 3, V-1 to 2, V-2 to 1, and Fail to 0. This is the predicted value of tensile strength. This is the predicted value of the flexural modulus. This is the predicted value for heat distortion temperature. This is a predicted value for the melt flow rate; The optimization objective is to maximize At the same time Controls are placed within the preset application window.

8. The method for optimizing the proportion of nanocomposite flame retardant masterbatch in engineering plastics based on machine learning according to claim 7, characterized in that, The improved non-dominated sorting genetic algorithm (NSGA-II) introduces a dynamic crowding distance calculation mechanism and an elite retention strategy on the basis of the traditional NSGA-II algorithm; The population size is set at 200, the maximum number of generations is 100, the crossover probability is 0.9, and the mutation probability is 0.

1. In each generation of non-dominated solution set, a performance constraint threshold is set according to the application scenario requirements, and only individuals that meet all constraints are retained to enter the next generation. The performance constraint thresholds include: limiting oxygen index not less than 28%, vertical burning rating not less than V-1, tensile strength not less than 80 MPa, flexural modulus not less than 3000 MPa, heat distortion temperature not less than 220°C, and melt flow rate between 8 and 25 g / 10 min.

9. The method for optimizing the proportion of nanocomposite flame retardant masterbatch in engineering plastics based on machine learning according to claim 8, characterized in that, The improved non-dominated sorting genetic algorithm (NSGA-II) introduces a performance stability index in the convergence determination, which is defined as the sum of the standard deviations of the objective function values ​​in the most recent ten generations of non-dominated solution sets; When the performance stability index is less than the preset tolerance of 0.005 for five consecutive generations, the algorithm is determined to have converged and the evolution is terminated.

10. The method for optimizing the proportion of nanocomposite flame-retardant masterbatch in engineering plastics based on machine learning according to claim 1, characterized in that, It also includes an online feedback correction step: The measured values ​​of the melt flow rate of the product are collected in real time by melt flow rate meters and process sensors deployed on the production line. Processing temperature and injection pressure ; The collected process parameters are input into the hybrid machine learning model to obtain the corresponding predicted melt flow rate. ; Calculate K consecutive production batches ( The mean relative deviation of melt flow rate : ； when When this is triggered, the model's incremental learning mechanism is activated, incorporating recently collected data... The data pairs are added to the training dataset to perform local retraining on the Gaussian process regression part of the hybrid machine learning model.

Images