A high-performance water-based acrylic emulsion formula screening method

Abstract

The application provides a high-performance water-based acrylic emulsion formula screening method, comprising the following steps: constructing a multi-source heterogeneous iterable database and differentiating storage; extracting key features from the database and performing cross-scale feature conversion and screening to obtain important key features; taking the important key features as input and the target performance as output, establishing a machine learning model and optimizing; using the optimized model to perform high-throughput virtual screening on virtual formula and process combinations, screening out candidate formulas with synergistic optimization of flame retardancy, hardness and adhesion, preparing water-based acrylic emulsion samples, obtaining the flame retardancy, hardness and adhesion data of the samples, completing verification and analysis; according to the verification data, further updating the database and model based on an active learning strategy, and performing explainability analysis of the machine learning model to guide subsequent experiments; repeating the above steps until the target high-performance water-based acrylic emulsion formula is obtained.

Description

Technical Field

[0001] This invention relates to the field of materials science and technology, and in particular to a method for screening high-performance waterborne acrylic emulsion formulations. Background Technology

[0002] Waterborne acrylic emulsions are widely used in coatings, adhesives, and textile coatings due to their environmental friendliness, non-toxicity, and good weather resistance. However, traditional waterborne acrylic emulsions often struggle to achieve a balance between key properties such as flame retardancy, hardness, and adhesion. This is especially true in applications with extremely high safety requirements (such as coatings for electronic devices and interior trim for transportation vehicles), where balancing high flame retardancy (e.g., UL-94 V-0 rating) with high hardness and high adhesion becomes a technical bottleneck.

[0003] Currently, the development of high-performance waterborne acrylic emulsions mainly relies on trial-and-error experiments and experience adjustments. Researchers design formulations (such as monomer combinations and flame retardant selection) and processes (such as polymerization temperature and feeding methods) based on limited chemical knowledge and past experience, and then conduct a large number of repetitive experiments to verify performance. This method has significant drawbacks: (1) Long R&D cycle and high cost: Obtaining a formulation with excellent comprehensive performance requires a lot of time, raw materials and human resources. (2) Limited exploration space: The formulation and process parameters constitute a very high-dimensional complex space, and it is difficult to conduct a comprehensive and systematic exploration based on human experience. It is easy to get stuck in local optima and fail to discover new formulations with greater potential. (3) Vague understanding of mechanism: There are complex nonlinear interactions between the components, process parameters and final performance. Traditional methods are difficult to quantitatively reveal these inherent laws, making performance optimization lack clear theoretical guidance.

[0004] In recent years, although machine learning technology has shown great potential in materials science, its application in the design of waterborne acrylic emulsions still faces challenges: (1) Data bottleneck: High-quality, large-scale datasets are the basis for model effectiveness. Data in the field of waterborne acrylic emulsions are often scattered from different sources (experiments, literature, simulations), and are structurally heterogeneous and have inconsistent standards, lacking effective integration and management solutions. (2) Complexity of feature engineering: Emulsion performance is the result of the combined effects of multiple scale factors such as molecular structure, microstructure and macroscopic process. How to select the key feature subset most relevant to the target performance from massive parameters is the key to achieving accurate prediction. (3) Model applicability: Simple linear models or ordinary neural networks are difficult to capture the complex temporal dependencies and nonlinear relationships between formulation, process and performance. At the same time, the black-box nature of the model makes the decision-making process opaque, which is not conducive to guiding actual chemical design. Summary of the Invention

[0005] This invention addresses the challenge that, despite the great potential of machine learning technology in materials science in recent years, its application in the design of waterborne acrylic emulsions still faces difficulties, and provides a method for screening high-performance waterborne acrylic emulsion formulations.

[0006] One of the technical solutions of the present invention is to provide a high-performance waterborne acrylic emulsion formulation screening method, (1) integrating experimental, textual data and simulation data to construct a multi-source heterogeneous iterative database, and storing the database in a differentiated manner; (2) extracting key features from the database, and performing cross-scale (molecular, microscopic, macroscopic) feature transformation and screening on the key features to obtain important key features; (3) using the important key features as input and the target performance as output, establishing and optimizing a machine learning model to obtain an optimized machine learning model; (4) using the optimized machine learning model to perform high-throughput processing on virtual formulations and process combinations. Virtual screening, combined with multi-objective optimization algorithm to screen out candidate formulations with synergistic optimization of flame retardancy, hardness and adhesion, to obtain target candidate formulations; (5) Prepare waterborne acrylic emulsion samples using the target candidate formulations, perform performance tests on the samples to obtain the flame retardancy, hardness and adhesion data of the samples, and complete the experimental verification and performance analysis of the target candidate formulations; (6) Based on the verification data of the candidate formulations, based on the active learning strategy, further update the database and the optimized machine learning model, and perform interpretability analysis of the machine learning model to guide subsequent experiments; repeat the above steps until the target high-performance waterborne acrylic emulsion formulation is obtained.

[0007] The database integrates data from multiple sources and with different structures (experimental data, literature data, and simulation data), primarily focusing on flame retardant properties, hardness, and adhesion properties. The experimental data includes collected experimental data on acrylic emulsion formulations, the textual data consists of material parameters extracted from relevant literature, and the simulation data is generated through methods such as coarse-grained molecular dynamics (CG-MD) simulations to produce reliable data.

[0008] Furthermore, the features include component features and process features. The component features include monomer type and ratio, initiator type and dosage, functional monomers (including flame retardant monomers), chain transfer agents, and the type and content of fillers (such as nano aluminum hydroxide and expandable graphite). The process features include polymerization temperature, polymerization time, stirring rate, feeding method, emulsification process parameters, and curing conditions.

[0009] The differentiated storage of the database is specifically manifested in the classification and storage based on the functional type of water-based acrylic emulsions, including two categories: intrinsic flame retardant (achieved by introducing flame retardant monomers or structures) and composite flame retardant (achieved by adding physical fillers). Corresponding feature tags are established for each type of data. The feature tags include three tags: original database tags (storing unprocessed experimental, literature, and simulation data), feature database tags (storing cleaned and labeled feature vectors), and performance database tags (such as UL94 flame retardant rating and hardness rating), to achieve differentiated data storage and iteration.

[0010] Furthermore, the cross-scale feature screening includes a method that combines correlation analysis, principal component analysis, and domain knowledge-driven approaches to screen out the most important key features that are most relevant to the target performance from multi-scale features.

[0011] The cross-scale feature transformation includes dividing the key features into discrete data key features and continuous data key features, performing one-hot encoding on the discrete data key features (such as flame retardancy rating and adhesion rating), performing min-max normalization on the continuous data key features (such as monomer content and temperature), and transforming and expressing the cross-scale features using descriptors such as SMILS and Morgan fingerprints, converting the key features into machine-readable language, i.e., data standardization processing.

[0012] Furthermore, the machine learning model is established using methods including a long short-term memory network and a support vector machine. The long short-term memory network performs time-series data processing on the key features to obtain the hardness value, and the support vector machine module performs static feature processing on the key features to obtain the flame retardant rating and adhesion rating.

[0013] The Long Short-Term Memory (LSTM) network module is particularly suitable for handling the impact of time-series process parameters on final performance because it can effectively capture the sequential dependence and long-term correlation between formulation and process parameters. Meanwhile, the Support Vector Machine (SVM) module can simultaneously build regression and classification models.

[0014] The accuracy of the established machine learning model is initially assessed using the coefficient of determination and root mean square error. Five-fold cross-validation is then used to determine whether the machine learning model is overfitting. Simultaneously, hyperparameter tuning is performed to obtain an optimized machine learning model. Furthermore, based on an active learning strategy, the updated database is used to train and optimize the machine learning model.

[0015] The high-throughput virtual screening includes generating virtual formula-process combinations using an optimized machine learning model, thereby quickly predicting the flame retardancy, hardness, and adhesion of the virtual formula-process combinations, and using a multi-objective optimization algorithm to screen out the Pareto optimal solution set that performs well in terms of flame retardancy, hardness, and adhesion performance indicators.

[0016] Furthermore, an optimized Long Short-Term Memory (LSTM) network module is used to perform high-throughput virtual screening of virtual formulation and process combinations to obtain selected combinations. These combinations are then input into a trained machine learning model to quickly predict the flame retardancy, hardness, and adhesion performance of the selected combinations. A multi-objective optimization algorithm (NSGA-II) is employed to select the Pareto optimal solution set that performs well across multiple performance metrics, outputting candidate formulations on the Pareto optimal front. The selected target candidate formulations are then subjected to actual preparation and performance testing to obtain their true flame retardancy, hardness, and adhesion data. This data is used as new samples to update the database, and the updated database is used to retrain the machine learning model, enabling continuous iterative optimization.

[0017] Furthermore, interpretability analysis is performed on the machine learning model using SHAP, Pearson correlation matrix, or feature removal. Specifically, SHAP, Pearson correlation matrix, and feature removal are used to analyze the machine learning model, quantify the contribution of each input feature (such as the content of a certain monomer or a certain process temperature) to the predicted performance (such as flame retardancy), and rank the importance of key features, thereby revealing the key driving factors affecting the design of high-performance waterborne acrylic emulsions.

[0018] Furthermore, the performance includes flame retardancy reaching UL94 V-0, hardness ≥ 3H, and adhesion ≥ 5B, and the candidate formulation is a formulation that simultaneously meets the above three performance targets.

[0019] Compared with the prior art, the present invention has the following advantages: 1) This invention is a high-performance waterborne acrylic emulsion development method based on machine learning and active learning framework, which focuses on solving the problem of synergistic optimization of flame retardancy, hardness and adhesion; it can integrate multi-source data, intelligently screen features, establish high-precision prediction models, and realize a new method for efficient screening and iterative optimization of massive candidate formulations, so as to break through the limitations of traditional R&D model and quickly and accurately design waterborne acrylic emulsions with excellent comprehensive performance.

[0020] 2) This invention enables the rapid design and development of high-performance waterborne acrylic emulsions by integrating multi-source heterogeneous data, intelligent analysis and prediction.

[0021] 3) This invention uses high-throughput virtual screening to evaluate tens of thousands or even millions of formulation possibilities in a short time, shortening the traditional R&D cycle of several months to several years to several weeks.

[0022] 4) This invention can efficiently screen and discover high-performance formulation combinations, and achieve synergistic optimization of multiple high-performance targets such as flame retardancy, hardness, and adhesion.

[0023] 5) The interpretability analysis of the machine learning model in this invention can quantitatively reveal the degree of influence of each factor on performance, transforming R&D from "experience-driven" to "data and mechanism-driven", which has important scientific guiding significance.

[0024] 6) The experimental verification data in this invention is fed back into the database, enabling the model to continuously learn and iteratively optimize, forming an intelligent design system that gets smarter the more it is used. Attached Figure Description

[0025] Figure 1 This is a schematic diagram illustrating the principle of the high-performance waterborne acrylic emulsion development method of the present invention. Detailed Implementation

[0026] The following examples are provided to further illustrate the present invention and are intended to explain the invention, not to limit its scope. Unless otherwise specified, all figures are expressed in parts by weight and weight percentages.

[0027] Unless otherwise specified, the raw materials used in this invention are all conventional commercially available products; unless otherwise specified, the methods used in this invention are all conventional methods in the field.

[0028] The embodiments of the present invention will be further described below with reference to several examples.

[0029] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0030] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0031] Example 1: Model Construction and Screening Based on Intrinsically Flame-Retardant Waterborne Acrylic Emulsions This embodiment details the specific process of designing an intrinsically flame-retardant high-performance waterborne acrylic emulsion using the method of the present invention.

[0032] S1: Construction of Multi-Source Heterogeneous Iterable Databases A. Data source integration: (1) Experimental data: Data from 125 sets of acrylic emulsion polymerization experiments were collected in the laboratory. Formulation data included: the ratio of methyl methacrylate (MMA), butyl acrylate (BA), and acrylic acid (AA); the amount of flame-retardant monomer vinyltriisopropoxysilane (VTIPS) (0-15 wt%); the amount of initiator (ammonium persulfate APS); and the amount of emulsifier (sodium dodecyl sulfate SDS). Process data included: polymerization temperature (75-85 ℃), stirring speed (200-400 rpm), and dropping time.

[0033] (2) Text data: 80 sets of data were extracted from 30 relevant academic papers and patents, including the ratio of the monomers, process conditions, and corresponding performance data: limiting oxygen index (LOI), UL-94 flame retardant rating, pencil hardness, and cross-cut adhesion rating.

[0034] (3) Simulation data: For some representative formulations (such as different VTIPS contents), coarse-grained molecular dynamics (CG-MD) was used to simulate and calculate parameters such as the initial thermal decomposition temperature and char formation tendency under high temperature field, generating 15 sets of simulation data.

[0035] B. Differentiated Database Architecture: The collected raw data (a total of 220 sets) was stored in the original database.

[0036] Feature vectors are extracted and constructed from the data, including: macroscopic parameters such as mole fraction of each monomer, VTIPS content, initiator concentration, polymerization temperature, and stirring rate; molecular descriptors obtained through calculation, such as the molecular weight of each monomer; and thermal decomposition temperature extracted from simulation data. These feature vectors are stored in a feature database. Corresponding performance labels (LOI, UL-94 rating, pencil hardness, adhesion rating) are stored in a performance database. Differential backups are performed on the database to ensure data security and traceability. In this embodiment, such data is labeled as "intrinsically flame retardant".

[0037] S2: Cross-scale Feature Filtering and Transformation By combining correlation analysis (calculating the Pearson correlation coefficient between each feature and performance index) and principal component analysis (PCA) for dimensionality reduction, and by introducing domain knowledge (such as the knowledge that phosphorus-containing and silicon monomers help with flame retardancy), 15 key feature subsets were selected from the initial 50-plus features.

[0038] Key features include: VTIPS content, MMA / BA ratio, APS concentration, polymerization temperature, and molecular chain rigidity descriptors calculated by Morgan fingerprinting. These features cover information from the molecular scale (monomer structure) to the macroscopic scale (process parameters).

[0039] Discrete data such as UL-94 ratings (NR, V-2, V-1, V-0) and adhesion ratings (0-5B) are encoded using one-hot encoding. Continuous data such as monomer content, temperature, LOI value, and hardness value are normalized using Min-Max and scaled to the [0, 1] range.

[0040] S3: Machine Learning Model Building and Optimization Using the 15 key features obtained from S2 as input and the three properties of LOI value, hardness, and adhesion as output, a dual-branch LSTM model and an SVM model are established.

[0041] Branch 1 (Time-series branch): Processes simulated process time-series data (e.g., discretizing temperature control curves into time series). Branch 2 (Static branch): Processes static formulation characteristics (e.g., monomer ratio, VTIPS content). The outputs of the two branches are merged and connected to a fully connected network, ultimately predicting three performance metrics simultaneously.

[0042] Using R 2 Using RMSE as the evaluation metric, hyperparameters were tuned through 5-fold cross-validation.

[0043] An active learning framework was implemented: After the initial model was trained on 200 datasets, the remaining 20 unlabeled datasets were ranked by prediction uncertainty, and the samples with the greatest uncertainty were prioritized for experimental validation. The new experimental data was then fed back into the database to retrain the model. The final model achieved an average R-value of [missing information - likely a specific value or measure] for the three performance predictions. 2 It reached 0.91.

[0044] S4: Candidate Sample Screening Based on High Throughput Using optimized LSTM and SVM models, approximately 50,000 formulation-process combinations conforming to chemical rules (such as VTIPS content varying between 5-20% and temperature varying between 70-90 °C) are generated in the virtual space.

[0045] Quickly predict the LOI, hardness, and adhesion performance of each virtual sample.

[0046] The multi-objective optimization algorithm NSGA-II was used to screen samples with the objectives of maximizing LOI (>30%), maximizing hardness (≥3H), and maximizing adhesion (≥5B) simultaneously, and output the Pareto optimal solution set, which includes 15 candidate samples with excellent overall performance.

[0047] S5: Validation of target candidate samples One candidate sample with the best predictive performance was selected from the Pareto frontier for actual preparation. The key formulation was: MMA:BA:AA = 50:45:5 (weight ratio), and VTIPS was used at 12 wt%.

[0048] Actual test results show: LOI = 32.5%, UL-94 rating of V-0, pencil hardness of 4H, and adhesion of 5B. The results demonstrate that the model predictions are accurate, and a high-performance intrinsic flame-retardant emulsion has been successfully obtained.

[0049] S6: Model Iteration and Interpretability Analysis The validation data from S5 is used as a new sample and fed back into the database of S1 for updating.

[0050] The interpretability of the model was analyzed using the SHAP method. The analysis showed that VTIPS content was the most significant positive characteristic, playing a decisive role in improving LOI and flame retardant rating; the MMA / BA ratio was the most critical characteristic affecting hardness; while AA content and emulsifier dosage had significant effects on adhesion. These analytical results are consistent with chemical principles, enhancing the model's credibility and pointing the way for subsequent optimization.

[0051] Example 2: Rapid Development Based on Composite Flame-Retardant Waterborne Acrylic Emulsion This embodiment focuses on demonstrating the efficiency of the method of the present invention in developing additive flame retardant emulsions (composite flame retardant type), especially its advantages in handling multiple filler combinations.

[0052] S1: Construction of Multi-Source Heterogeneous Iterable Databases This embodiment constructs a sub-database focusing on composite flame-retardant emulsions. 180 sets of data were collected, with a focus on formulations containing fillers such as nano-aluminum hydroxide (nano-ATH), expandable graphite (EG), and nano-montmorillonite (MMT).

[0053] The data storage architecture is the same as in Example 1, but the data in this example is labeled as "composite flame retardant".

[0054] S2: Cross-scale Feature Filtering and Transformation Key characteristics identified included: nano-ATH addition amount, EG addition amount, MMT addition amount, filler surface modifier type (unique thermal code), monomer composition of the core emulsion (MMA / BA / AA ratio), and curing temperature. Data standardization was the same as in Example 1.

[0055] S3: Machine Learning Model Building and Optimization The same dual-branch LSTM and SVM models as in Example 1 were used for training and optimization. Due to the more complex filler parameters, the model stabilized after 5 iterations of active learning, ultimately achieving an accuracy of 95% in predicting flame retardancy ratings and a high accuracy in predicting hardness and adhesion R. 2 All exceeded 0.88.

[0056] S4: Candidate Sample Screening Based on High Throughput For the "composite flame retardant" database, 80,000 virtual samples were generated, with variables including different combinations and amounts of three fillers.

[0057] The optimized model was used for prediction, and NSGA-II was used to screen candidate samples with UL-94 V-0 rating, hardness ≥3H, and adhesion ≥5B.

[0058] S5: Validation of target candidate samples One sample with a novel packing combination was selected from the results for validation: nano-ATH: 10 phr, EG: 5 phr, MMT: 3 phr.

[0059] Actual preparation and testing showed that the performance of this formulation fully met the predicted targets: UL-94 V-0 rating, hardness of 3H, and adhesion of 5B. This method avoids the tedious process of traditionally requiring numerous trials of different filler ratios, and quickly identifies the optimal combination.

[0060] S6: Model Iteration and Interpretability Analysis Validation data was fed back into the database. Pearson correlation matrix analysis showed a strong positive correlation between the amount of nano-ATH added and the flame retardant performance (LOI), while the amount of EG added was crucial for achieving a UL-94 V-0 rating. The addition of MMT contributed positively to maintaining hardness, but excessive amounts impaired adhesion. These patterns provide clear quantitative guidance for filler formulation.

[0061] The above embodiments describe in detail the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, shall still fall within the scope of protection of the present invention if they do not exceed the scope covered by the specification.

[0062] Although embodiments of the invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the scope of the invention is not limited to the contents disclosed in the embodiments.

Claims

1. A method for screening high-performance waterborne acrylic emulsion formulations, characterized in that, Includes the following steps: (1) Integrate experimental data, text data and simulation data to construct a multi-source heterogeneous iterable database, and store the database in a differentiated manner; (2) Extract key features from the database, and perform cross-scale feature transformation and screening on the key features to obtain important key features; (3) Use the important key features as input and the target performance as output to establish and optimize a machine learning model to obtain an optimized machine learning model; (4) Use the optimized machine learning model to perform high-throughput virtual screening of virtual formula and process combination, and combine multi-objective optimization algorithm to screen out candidate formulas with synergistic optimization of flame retardancy, hardness and adhesion to obtain target candidate formulas; (5) Use the target candidate formula to prepare waterborne acrylic emulsion samples, perform performance tests on the samples to obtain the flame retardancy, hardness and adhesion data of the samples, and complete the experimental verification and performance analysis of the target candidate formulas; (6) Based on the verification data of the candidate formulation, the database and the optimized machine learning model are further updated based on the active learning strategy, and the interpretability analysis of the machine learning model is performed to guide subsequent experiments; repeat the above steps until the target high-performance waterborne acrylic emulsion formulation is obtained.

2. The method according to claim 1, characterized in that, The experimental data includes collected experimental data on acrylic emulsion formulations; the text data extracts material parameters from relevant literature; and the simulation data is generated through methods such as coarse-grained molecular dynamics simulation to produce relevant and reliable simulation data. The key features include component characteristics and process characteristics. The component characteristics include monomer type and ratio, initiator type and dosage, functional monomers, chain transfer agents, and filler type and content. The process characteristics include polymerization temperature, polymerization time, stirring rate, feeding method, emulsification process parameters, and curing conditions.

3. The method according to claim 2, characterized in that, Differentiated storage of the database includes classifying and storing data according to the functional type of water-based acrylic emulsions, including at least two categories: intrinsic flame retardant and composite flame retardant. Corresponding feature tags are established for each category of data. The feature tags include original database tags, feature database tags, and performance database tags, thereby enabling the storage and iteration of differentiated data.

4. The method according to claim 3, characterized in that, The cross-scale feature transformation includes dividing the key features into discrete data key features and continuous data key features, performing one-hot encoding on the discrete data key features, performing min-max normalization on the continuous data key features, and transforming and expressing the cross-scale features using descriptors such as SMILES and Morgan fingerprints, converting the key features into machine-readable language, that is, standardizing the data.

5. The method according to claim 4, characterized in that, The cross-scale feature screening method employs a combination of correlation analysis, principal component analysis, and domain knowledge-driven approaches to select the most important key features that are most relevant to the target performance from multi-scale features.

6. The method according to claim 5, characterized in that, The machine learning model is established using methods including a long short-term memory network and a support vector machine. The long short-term memory network performs time-series data processing on the key features to obtain the hardness value, and the support vector machine module performs static feature processing on the key features to obtain the flame retardancy rating and adhesion rating.

7. The method according to claim 6, characterized in that, The accuracy of the established machine learning model is initially assessed using the coefficient of determination and root mean square error. Five-fold cross-validation is then used to determine whether the machine learning model is overfitting. Simultaneously, hyperparameter tuning is performed to obtain an optimized machine learning model. Furthermore, based on an active learning strategy, the updated database is used to train and optimize the machine learning model.

8. The method according to claim 7, characterized in that, The high-throughput virtual screening includes generating virtual formula-process combinations using an optimized machine learning model, thereby quickly predicting the flame retardancy, hardness, and adhesion of the virtual formula-process combinations, and using a multi-objective optimization algorithm to screen out the Pareto optimal solution set that performs well in terms of flame retardancy, hardness, and adhesion performance indicators.

9. The method according to claim 8, characterized in that, The machine learning model was analyzed using SHAP, Pearson correlation matrix, or feature removal method to quantify the contribution of each input feature to the prediction performance and to rank the importance of key features, thereby revealing the key driving factors affecting the design of high-performance waterborne acrylic emulsions.

10. The method according to any one of claims 1 to 9, characterized in that, The target performance includes flame retardancy reaching UL94 V-0, hardness ≥ 3H, and adhesion ≥ 5B. The candidate formulation is a formulation that simultaneously meets the above three performance targets.

Images