An acrylic emulsion emulsifier screening method based on coarse-grained molecular dynamics model and machine learning acceleration

Abstract

A kind of acrylic emulsion emulsifier screening method based on coarse-grained molecular dynamics model and machine learning acceleration, comprising the following steps: S1: respectively constructing low-resolution coarse-grained model and high-resolution coarse-grained model of acrylate monomer-water-emulsifier ternary interface model, and by obtaining corresponding simulation trajectory;S2: extract low-resolution level performance index and high-resolution level performance index;S3: establish ladder coarse-grained comprehensive score model, calculate comprehensive score;S4: construct training data set;S5: screening high information content sample using active learning strategy, and iteratively expand training data set;S6: using the finally trained machine learning model to the candidate emulsifier library is carried out hierarchical virtual screening, output recommended emulsifier list.The present application realizes the efficient annotation of data by the hierarchical architecture of " low-resolution coarse-grained high-throughput preliminary screening→ high-resolution coarse-grained accurate rescreening", realizes the rapid screening of acrylic emulsion emulsifier in combination with active learning.

Description

Technical Field

[0001] This invention relates to the field of polymer emulsion polymerization technology, specifically to a method for screening emulsifiers for acrylic emulsion polymerization, particularly a method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration. Background Technology

[0002] Acrylic emulsions are high-molecular-weight aqueous dispersions prepared by emulsion polymerization of acrylate monomers (including methyl methacrylate (MMA), butyl acrylate (BA), methacrylic acid (MAA), etc.). They are widely used in water-based coatings, adhesives, textile coatings, and other fields. Emulsifiers are an indispensable key component in emulsion polymerization systems, directly affecting core qualities such as polymerization stability, latex particle size, film-forming properties, and coating water resistance.

[0003] Based on their mechanisms of action, emulsifiers for acrylic emulsions are mainly divided into two categories: traditional physically adsorbent emulsifiers (such as sodium dodecyl sulfate (SDS), sodium dodecylbenzene sulfonate (SDBS), and fatty alcohol polyoxyethylene ethers) and reactive emulsifiers (also known as polymerizable emulsifiers, whose molecular structure contains unsaturated double bonds that can participate in free radical polymerization, such as sodium allyl hydroxypropyl sulfonate (COPS-1) and 2-acrylamido-2-methylpropanesulfonic acid (AMPS)). Traditional physically adsorbent emulsifiers stabilize latex particles through physical adsorption, are low in cost, and are abundant in variety, but suffer from problems such as free emulsifier migration and decreased water resistance of the coating film. Reactive emulsifiers anchor to the surface of latex particles through covalent bonds, which can significantly reduce free emulsifier residue and improve the water resistance of the coating film, but their structural design is more complex and their cost is higher. In practical industrial applications, both types of emulsifiers are often used alone or in combination, and both require important screening.

[0004] Currently, emulsifier screening still heavily relies on an "experience-based trial-and-error" approach, which is time-consuming, costly, and difficult to scale. Although molecular dynamics (MD) simulations have been introduced into the study of surfactant interfacial behavior, existing technologies face the following key technical bottlenecks: First, existing coarse-grained force fields lack specific parameterization schemes for acrylate monomers and emulsifiers (especially reactive emulsifiers). Acrylate monomers have polar ester groups, and reactive emulsifiers contain polymerizable double bonds. Directly applying general coarse-grained force fields (such as Martini) will lead to systematic biases in the prediction of interfacial properties.

[0005] Second, existing MD screening methods struggle to balance accuracy and efficiency. All-atom MD simulations offer high accuracy but come with extremely high computational costs—a single simulation can take days to weeks, making it difficult to support rapid screening of large numbers of candidate molecules. While coarse-grained MD can significantly improve computational efficiency, single-resolution coarse-grained models are either overly simplified (with only a few beads per molecule), failing to distinguish subtle structural differences between homologous series, or have excessively high resolution, leading to a significant increase in computational costs.

[0006] There are no reports on how to achieve a "fast first, accurate later" stepwise screening within a single coarse-grained framework, particularly regarding reactive emulsifiers and acrylic monomers. Active learning sampling strategies have not yet been applied in multi-resolution coarse-grained simulation scenarios. Low-resolution and high-resolution coarse-grained simulations have different computational costs and information densities; therefore, there are no reports on how to design hierarchical sampling strategies to enable active learning to intelligently allocate annotation resources between the two resolutions, while simultaneously using low-resolution simulation results to guide the initial configuration construction of high-resolution simulations.

[0007] Therefore, developing a virtual screening method for acrylic emulsion emulsifiers (covering traditional physical adsorption emulsifiers and reactive emulsifiers) based on coarse-grained molecular dynamics, integrating a step-by-step resolution strategy, active learning, and machine learning is of significant technical and economic value for accelerating the development and selection of high-performance acrylic emulsion emulsifiers. Summary of the Invention

[0008] The technical problem this invention aims to solve is to provide a screening method for acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration. This method employs a hierarchical architecture of "low-resolution coarse-grained high-throughput primary screening → high-resolution coarse-grained precise secondary screening," combined with active learning to achieve efficient data annotation, forming an end-to-end rapid screening system capable of quickly screening acrylic emulsion emulsifiers. The technical solution adopted is as follows: A method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration is characterized by the following steps: S1: Construct low-resolution coarse-grained models and high-resolution coarse-grained models of the ternary interface model of acrylate monomer-water-emulsifier, respectively, and obtain the corresponding low-resolution coarse-grained simulation trajectory and high-resolution coarse-grained simulation trajectory through coarse-grained molecular dynamics simulation. S2: Extract low-resolution hierarchical performance indicators based on the low-resolution coarse-grained simulation trajectory, and extract high-resolution hierarchical performance indicators based on the high-resolution coarse-grained simulation trajectory. S3: Establish a step-by-step coarse-grained comprehensive scoring model, generate a low-resolution score based on the low-resolution hierarchical performance index for the first-level screening, generate a high-resolution score based on the high-resolution hierarchical performance index for the second-level screening, and calculate the comprehensive score. S4: Extract the structural descriptor of the emulsifier molecule, and use the low-resolution hierarchical performance index and high-resolution hierarchical performance index as coarse-grained performance labels to construct a training dataset; S5: Employ an active learning strategy to select high-information samples and iteratively expand the training dataset; S6: The final trained machine learning model is used to perform hierarchical virtual screening of the candidate emulsifier library. The model first screens the candidate emulsifier library with a low-resolution coarse-grained model and then screens the candidate emulsifier library with a high-resolution coarse-grained model. The model then outputs a recommended list of emulsifiers.

[0009] In the preferred embodiment, the above-mentioned method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration further includes step S7: emulsion polymerization experimental verification and closed-loop optimization.

[0010] In preferred step S7, the method for experimental verification and loop closure optimization of emulsion polymerization includes the following steps: S7-1 Select several emulsifiers (including traditional physical adsorption type emulsifiers and reactive emulsifiers) that rank high in the virtual screening comprehensive score, and conduct emulsion polymerization experiments according to the target acrylic emulsion formulation; S7-2 compares experimental data with predicted results, analyzes the sources of deviation, and optimizes coarse-grained force field parameters or machine learning models accordingly. S7-3 adds the experimental data to the training dataset and retrains the machine learning model.

[0011] By updating the machine learning model, the method can be continuously improved in subsequent screening, achieving closed-loop optimization of "virtual screening → experimental verification → model correction".

[0012] In a preferred embodiment, the emulsifier includes traditional physical adsorption emulsifiers and reactive emulsifiers; the reactive emulsifier contains at least one carbon-carbon double bond in its molecular structure that can participate in free radical polymerization.

[0013] In step S1 above, a ladder-style coarse-grained model library for acrylic emulsion emulsifiers is constructed, which includes two resolution levels. They share the same coarse-grained force field framework but have different levels of bead mapping fineness: the low-resolution coarse-grained model (CG-low) is used for high-throughput pre-screening and rapid sorting; the high-resolution coarse-grained model (CG-high) is used for more refined evaluation of the interfacial properties of the pre-screened preferred molecules.

[0014] In the preferred embodiment, in step S1 above, each emulsifier molecule in the low-resolution coarse-grained model is represented by 3 to 6 coarse-grained beads; in the high-resolution coarse-grained model, each emulsifier molecule is represented by 10 to 20 coarse-grained beads; the initial configuration of the high-resolution coarse-grained model is obtained by performing a bead splitting operation on the final configuration of the low-resolution simulation.

[0015] In the preferred embodiment, the coarse-grained model described in step S1 above performs force field parameter calibration for the polar ester groups of the acrylate monomer, including increasing the Lennard-Jones potential well depth between the ester group representing the bead and the water droplet.

[0016] In the preferred embodiment, the method for constructing a low-resolution coarse-grained model of the ternary interface model of acrylate monomer-water-emulsifier in step S1 includes the following steps: S1-1-1: Establish a low-resolution coarse-grained model of acrylate monomers and water; Acrylic ester monomer molecules (including one or more of methyl methacrylate (MMA), butyl acrylate (BA), methacrylic acid (MAA), and ethyl acrylate (EA)) are coarsened at a low resolution level according to a "4-to-1" or "5-to-1" mapping scheme, with each coarsened bead representing 4 to 5 heavy atoms. The ester functional groups and hydrophobic alkyl chains of acrylate monomers are mapped to polar beads and non-polar beads, respectively; the water molecule adopts an implicit water model, which does not explicitly include water molecules, but shields electrostatic interactions by setting a continuous dielectric constant (ε≈60–80) to maximize the calculation speed. S1-1-2: Establish a low-resolution coarse-grained model of the emulsifier; Based on the chemical structure of emulsifier molecules, a simplified functional module mapping strategy is adopted: hydrophilic head groups (such as sulfate -SO4) are mapped... - sulfonate-SO3 - Carboxylate group -COO - For example, a polyoxyethylene ether chain can be mapped to a hydrophilic bead with a specific charge and polarity; a hydrophobic tail chain (alkyl chain) can be mapped to 1 to 2 hydrophobic beads (depending on the tail chain length); for reactive emulsifiers, their polymerizable double bond functional groups (allyl, acrylamide, vinyl, methacrylate, etc.) can be mapped to a reactive bead. Overall, each emulsifier molecule is represented by 3 to 6 coarse-grained beads in the low-resolution coarse-grained model; S1-1-3: Constructing a low-resolution coarse-grained model of the ternary interface of acrylate monomer-water-emulsifier; Emulsifier molecules were arranged on both sides of the oil / water interface according to a preset interface area / emulsifier molecule ratio. Counterions were added to neutralize the net charge of the system. Subsequently, energy minimization, short-time equilibrium simulation (50-100 ns), and production simulation (1000-2000 ns) were performed, and the low-resolution coarse-grained simulation trajectory was recorded.

[0017] For low-resolution coarse-grained models, Martini 3 force field or its extended version calibrated with coarse-grained parameters is used; key parameters include: Lennard-Jones potential energy parameters between different bead types (hydrophilic, hydrophobic, polar, nonpolar, reactive) – interaction potential well depth ε and equilibrium distance σ; for reactive emulsifiers, the spring constant bonding parameters of reactive beads also need to be calibrated.

[0018] The aforementioned heavy atom refers to all atoms in a molecule except for hydrogen atoms (such as C, N, O, S, etc.). In coarse-grained mapping, hydrogen atoms are implicitly treated according to Martini force field convention because of their light mass and small contribution to interfacial behavior.

[0019] The aforementioned polar beads, nonpolar beads, hydrophilic beads, hydrophobic beads, and reactive beads are all specific subclasses of "coarse-grained beads." They are classified according to their physicochemical properties and functions: polarity into polar / nonpolar beads; hydrophilic / hydrophobic beads into hydrophilic / hydrophobic beads; and reactive beads into reactive beads (only used for reactive emulsifiers).

[0020] In the preferred embodiment, the method for constructing a high-resolution coarse-grained model of the acrylate monomer-water-emulsifier ternary interface model in step S1 includes the following steps: S1-2-1: Initial configuration of the high-resolution coarse-grained model; By refining the final configuration of the low-resolution simulation by bead splitting, the initial configuration of the high-resolution coarse-grained model is obtained: one coarse-grained bead in the low-resolution coarse-grained model is converted into multiple (e.g., 2 to 4) second coarse-grained beads in the high-resolution coarse-grained model according to a preset splitting rule. The initial position of the second coarse-grained beads is determined by geometric interpolation and local energy minimization. S1-2-2: After short-term rebalancing (20-50 ns) of the constructed high-resolution coarse-grained model, production simulation (1000-2000 ns) is performed, and the high-resolution coarse-grained simulation trajectory is recorded.

[0021] In step S1-2-1 above, the initial configuration of the high-resolution coarse-grained model is obtained by performing bead splitting refinement on the final configuration of the low-resolution simulation. This process avoids the time-consuming process of reconstructing the high-resolution coarse-grained model from a random state.

[0022] Increasing the number of beads on the basis of the low-resolution coarse-grained model can distinguish more detailed molecular structures. The splitting rules of the coarse-grained beads can be as follows: (1) For acrylate monomers, the original polar bead is split into two second coarse-grained beads, one of which represents the carbonyl group and oxygen atom of the ester group, and the other represents the alkoxy part of the ester group; the hydrophobic alkyl chain is refined at a density of one bead for every 2 carbon atoms; (2) For emulsifiers, the hydrophilic head group is split into 2 to 3 second coarse-grained beads to distinguish the charge distribution inside the head group; the hydrophobic tail chain is refined at a density of one second coarse-grained bead for every 3 to 4 carbon atoms; (3) For reactive emulsifiers, the polymerizable double bond is retained as an independent reactive bead with a clear direction; (4) Overall, each emulsifier molecule is represented by 10 to 20 second coarse-grained beads in the high-resolution coarse-grained model.

[0023] In step S2 above, both low-resolution and high-resolution hierarchical performance indicators typically include at least one of interfacial tension, interfacial formation energy, interfacial arrangement characteristics, and desorption free energy; for emulsifiers containing polymerizable double bonds, double bond exposure index and copolymerization tendency index are also included.

[0024] Preferably, in step S2 above, the method for extracting low-resolution hierarchical performance indicators and high-resolution hierarchical performance indicators includes the following steps: S2-1: Extract low-resolution hierarchical performance indicators; Low-resolution layer performance metrics include: (1) Low-resolution interfacial tension γ low The trajectory was calculated from a low-resolution coarse-grained simulation using stress tensor analysis. (2) Equilibrium time τ of emulsifier self-assembly at the interface eq (The time required from the start of the simulation to the steady state of interface coverage). (3) For reactive emulsifiers, calculate the low-resolution exposure index E of polymerizable double bonds. dblow Calculate using the following formula:

[0025] Where N frames d represents the total number of frames captured. min (t) represents the t-th frame (t≤N), d min (t) represents the centroid distance between the reactive bead containing the polymerizable double bond and the nearest monomer bead in frame t, and dc The contact characteristic distance is 0.6 nm. For traditional physically adsorbed emulsifiers, this indicator defaults to E. dblow Set to 0; (4) Critical micelle concentration (CMC) estimated based on low-resolution simulation low (Low-resolution simulations were performed at multiple gradient total emulsifier concentrations. The concentrations of unassociated (free) emulsifier molecules were statistically analyzed, and the inflection point where the free concentration-total concentration curve changed from a linear increase to a plateau (dynamic equilibrium) was taken as the CMC.) low ); S2-2: Extract high-resolution hierarchical performance indicators; High-resolution hierarchical performance metrics include: (1) High-resolution interfacial tension γ high This serves as a refinement of the low-resolution values. (2) Interface formation energy ∆G calculated by high-resolution coarse-grained simulation trajectory ads The umbrella sampling method was used, with the distance between the centroid of the emulsifier molecule and the center of the interface as the reaction coordinate, to calculate the change in free energy for migration from the aqueous phase to the interface and stable adsorption. (3) For reactive emulsifiers, calculate the high-resolution exposure index E of polymerizable double bonds. dbhigh and co-aggregation tendency index I copol (Estimated by the contact frequency and binding free energy between reactive beads and monomer beads); For traditional physical adsorption emulsifiers, these two indicators are not included in the scoring or are set to default values ​​(e.g., the default values ​​for high-resolution exposure index and copolymerization tendency index are both set to 0); The high-resolution exposure index E of the polymerizable double bond dbhigh Low-resolution exposure index E for polymerizable double bonds dblow The same formula is used for calculation, only the simulated trajectory is replaced with a high-resolution trajectory; The copolymerization tendency index I copol Defined as:

[0026] Where ΔG pair The binding free energy (kJ·mol) of reactive beads to monomer beads was estimated from high-resolution coarse-grained simulations. -1 ); k copol The scaling factor for the Sigmoid function (dimensionless, recommended range k). copol ∈[1, 5]), control I copol Sensitivity to input variables; α db High-resolution double bond exposure index Edbhigh Weighting coefficients (dimensionless, recommended range α) db (∈[0.3, 0.7]) (4) Interface arrangement characteristic parameters at high resolution: the average tilt angle θ between the molecular long axis and the interface normal (obtained by calculating the angle between the vector connecting the center of the hydrophilic head group and the end of the hydrophobic tail chain and the interface normal), and the standard deviation σ of the tilt angle distribution. θ ; (5) Interface desorption free energy barrier ∆G desorp The change in free energy of emulsifier migration from the interface to the aqueous phase was calculated using umbrella-shaped sampling.

[0027] The above copolymerization tendency index I copol In the formula, k copol Control I copol Sensitivity to input variables; k copol The larger the value of k, the steeper the curve; recommended range k copol ∈ [1, 5], can be determined through cross-validation on the training set. α db Characterize the relative contribution of double bond exposure to copolymerization tendency.

[0028] Preferably, in step S3 above, the method for establishing a step-by-step coarse-grained comprehensive scoring model includes the following steps: S3-1 Construct a low-resolution scoring model and generate a low-resolution score based on the low-resolution hierarchical performance index. low Used for the first round of high-throughput filtration;

[0029] Where s(·) is the min-max normalization function, which normalizes all molecules in the candidate library across samples to [0,1], with the direction being "the larger the better"; δ reactive The reactive emulsifier is identified by a value of 1 (reactive type, 0 for traditional physical adsorption type); w1 represents the low-resolution interfacial tension s(γ). low The weights of ) are (recommended 0.2–0.4); w2 is the self-assembly equilibrium time s(τ). eq The weights of ) are (0.1–0.3); w3 is the critical micelle concentration s (CMC). low The weights of ) are (0.2–0.4); w4 is the low-resolution double bond exposure index s(E) dblow The weights of the emulsifiers are 0.1–0.3 (reactive type only); each weight can be automatically adjusted according to the emulsifier type or uniformly set to the default value. According to Score low The candidate pool is divided into three levels: "priority recommendation" (the highest 20-30%), "deferred" and "elimination" (the lowest 30-40%). S3-2 A high-resolution scoring model is constructed for molecules in the "Preferred Recommendation" level, and a high-resolution score is generated based on the high-resolution hierarchical performance index. high :

[0030] Where s(·) is the min-max normalization function, which normalizes all molecules in the candidate library across samples to [0,1], with the direction being "the larger the better"; w5 is the high-resolution interfacial tension s(γ). high The weight of ) (recommended 0.1–0.3); w6 is the interface formation free energy s(ΔG) ads The weights of ) are (0.2–0.4); w7 is the weight of the molecular tilt angle s(θ) (0.1–0.2); w8 is the weight of the desorption free energy s(ΔG). desorp The weights of ) are (0.1–0.3); w9 is the high-resolution double bond exposure index s(E) dbhigh The weights of ) (0.1–0.3, reactive only); w 10 The copolymerization tendency index s(I) copol The weights of the emulsifiers are 0.1–0.3 (reactive type only); each weight can be automatically adjusted according to the emulsifier type or uniformly set to the default value. S3-3 Construct a comprehensive scoring model and generate a comprehensive score. final ;

[0031] Where α is the weight of the low-resolution score in the final score (0 ≤ α ≤ 1, which can be dynamically adjusted according to the confidence level of the molecule in the low-resolution simulation); Score low norm Score high norm Score low Score high The value after global normalization ([0,1]).

[0032] Score low The score is obtained by weighted summation of four terms. high The result is obtained by weighted summation of 6 terms; although each term has been normalized to [0,1] using s(·), the total score range after weighted summation may differ. Direct linear combination can introduce bias due to inconsistent magnitudes, therefore, when substituting into the Score... final Previously, Score low With Score high Perform a global normalization on each (e.g., normalize min-max to [0,1]), and then combine them linearly.

[0033] The above step S3 is finally based on Score final Output the recommended sort list.

[0034] Preferably, in step S4 above, the method for constructing the training dataset includes the following steps: S4-1: Extract the structure descriptor vector x of the emulsifier molecule, including: (1) One-dimensional descriptors: molecular weight, number of rotatable bonds, number of hydrogen bond donors / acceptors; (2) Two-dimensional topological descriptor: topological polar surface area TPSA, ALOGP hydrophobicity parameter, molecular refractive index, HLB value (calculated by Griffin method); (3) Ion / charge descriptor: ion type (anionic / nonionic / cationic) and formal charge number of the hydrophilic head group; (4) Reactivity descriptor (only for reactive emulsifiers): polymerizable double bond type, double bond position index, steric hindrance parameter; the double bond position index is the number of chemical bonds between the polymerizable double bond and the hydrophilic head group, and the steric hindrance parameter is the number of substituents on the carbon of the double bond (monosubstituted, 1,1-disubstituted, etc.); for traditional physical adsorption emulsifiers, all of the above reactivity descriptors are set to 0; (5) 3D / Conformation Descriptor: Dipole moment and molecular size (radius of gyration) of the lowest energy conformation of the molecule obtained through fast force field optimization; This conformation optimization is only calculated using conventional molecular structure descriptors and does not involve molecular dynamics simulation; S4-2: Construct a performance label vector y from the performance metrics extracted in step S2, including low-resolution performance metrics (γ). low τ eq CMC low E dblow ) and high-resolution performance metrics (γ) high ΔG ads , θ, ΔG desorp E dbhigh I copol ); among which, the reaction-related indicators are only effective for reactive emulsifiers, and the corresponding values ​​for traditional physical adsorption emulsifiers are set to 0; Score low and Score high Calculated using the fixed scoring formula in step S3, it is not used as the prediction target of the machine learning model; S4-3: Construct the original training dataset D0 = {(x i , y i , t i )}, where t iHere, x is the emulsifier type label (traditional physical adsorption emulsifier / reactive emulsifier), and i is the sample index, representing the i-th sample in the training dataset (i = 1, 2, …, N; N is the total number of samples in the training set). i Let y be the structure descriptor vector of the i-th emulsifier molecule (extracted from S4-1). i For its corresponding performance label vector (calculated from S2 and S3), t i ∈ {0, 1} is its type label (0 represents traditional physical adsorption emulsifier, 1 represents reactive emulsifier).

[0035] Preferably, in step S5 above, the method of using an active learning strategy to select high-information samples and iteratively expand the training dataset includes the following steps: S5-1 Training the initial model: Select a small number (10-15) of representative emulsifiers with large structural differences from the candidate molecules, perform low-resolution coarse-grained simulation and obtain performance indicators as labels, which together with the descriptor vector x and type label t from step S4 constitute the initial training set D0. Two models are trained based on D0: M low M is used to predict low-resolution performance metrics. high Used to predict high-resolution performance metrics; both models use (x,t) as input features and can automatically distinguish emulsifier types; S5-2 Identify the molecules with the least model uncertainty: For unlabeled molecules in candidate pool U, use M... low Predict its low-resolution performance metrics, and provide the uncertainty U. low (x), representing the model's confidence level in the prediction; U low (x) is measured by the dispersion (variance) of the predicted values ​​of multiple model members. The more dispersed the predictions, the less confident the model is. S5-3 Molecules were selected and labeled in two stages: First-level sampling: Select U from U low (x) The largest k1 molecules (k1≥5) are the most uncertain group of molecules in the model. Low-resolution coarse-grained simulations are performed on them to obtain low-resolution performance labels. Second-level sampling: Among the molecules already labeled in the first level, scores are calculated using a low-resolution method. low Sort the data from high to low, take the first k2 (k2≤k1) and perform high-resolution coarse-grained simulation to obtain high-resolution performance labels; the initial configuration of the high-resolution simulation is obtained by bead splitting operation from the final configuration at the end of the low-resolution simulation. S5-4 Update the model: Incorporate the newly labeled samples from this round into the training dataset, and retrain or fine-tune M.low and M high ; S5-5 Termination Criteria: Repeat S5-2 to S5-4 until the model achieves the required accuracy in cross-validation (e.g., coefficient of determination R² ≥ 0.8 or mean absolute error MAE ≤ 3.5 kJ / mol), then output the final M value. low and M high .

[0036] Preferably, in step S6 above, the method for performing hierarchical virtual screening of the acrylic emulsion emulsifier library using the finally trained machine learning model includes the following steps: S6-1 First-level screening: using M low The model quickly predicts all unlabeled molecules in the candidate library and calculates the score. low Eliminate the bottom 30% to 40% of the applicants; S6-2 Second-stage screening: For the remaining molecules that passed the first-stage screening, M... high The model predicts high-resolution performance metrics and calculates the score. high and overall score final Output the final recommendation list.

[0037] This invention employs a hierarchical architecture of "low-resolution coarse-grained high-throughput primary screening → high-resolution coarse-grained precise secondary screening," combined with active learning to achieve efficient data annotation, forming an end-to-end rapid screening system capable of quickly screening acrylic emulsion emulsifiers. Compared with existing technologies, this invention has the following significant advantages: 1. Wide applicability, covering both traditional and reactive emulsifiers. This method, targeting the polar characteristics of acrylate monomer systems, constructs a step-by-step pure coarse-grained screening framework that combines low-resolution coarse-grained high-throughput pre-screening with high-resolution coarse-grained precise secondary screening. It is suitable for the unified screening of both traditional physical adsorption emulsifiers and reactive polymerizable emulsifiers.

[0038] 2. A coarse-grained model is used to achieve a step-by-step ultra-high-throughput screening. The low-resolution model uses a simplified bead mapping (3-6 beads per molecule) to quickly calculate indices such as interfacial tension and self-assembly equilibrium time, eliminating low-potential molecules at extremely low cost. The high-resolution model inherits the low-resolution equilibrium configuration through bead splitting operations, using a finer mapping (10-20 beads per molecule) to calculate interfacial formation free energy, desorption energy barrier, etc. For reactive emulsifiers, it also calculates the double bond exposure index and copolymerization tendency index. Completely independent of all-atom simulations, the time step can be increased to 20-40 fs (compared to 2 fs for all-atom simulations) due to a 5-20 fold reduction in the number of beads, thus improving the spatiotemporal scale by 2-3 orders of magnitude compared to all-atom simulations. Taking 500 candidate molecules as an example, the entire process takes approximately 2 months under parallel conditions with 8 GPUs. Simultaneously, this method uses coarse-grained MD simulation results to train two machine learning surrogate models (M... low M high The screening process is completed by using a substitute model instead of expensive full-library simulation, and finally the model is verified and fed back by emulsion polymerization experiments, forming a closed loop of "training-screening-verification".

[0039] 3. The specific force field adaptation for acrylate systems fills a technological gap. A systematic calibration of the coarse-grained force field was performed, pioneering a method for virtual screening of acrylic emulsion emulsifiers through coarse-grained molecular dynamics simulation. Attached Figure Description

[0040] Figure 1 This is a flowchart of a preferred embodiment of the present invention. Detailed Implementation

[0041] refer to Figure 1 Example 1: In this example, the acrylic emulsion emulsifier screening method based on coarse-grained molecular dynamics model and machine learning acceleration includes the following steps: S1: Construct low-resolution coarse-grained models and high-resolution coarse-grained models of the ternary interface model of acrylate monomer-water-emulsifier, respectively, and obtain the corresponding low-resolution coarse-grained simulation trajectory and high-resolution coarse-grained simulation trajectory through coarse-grained molecular dynamics simulation. The method for constructing a low-resolution coarse-grained model of the acrylate monomer-water-emulsifier ternary interface model in step S1 includes the following steps: S1-1-1: Establish a low-resolution coarse-grained model of acrylate monomers and water; Acrylic ester monomer molecules (including one or more of methyl methacrylate (MMA), butyl acrylate (BA), methacrylic acid (MAA), and ethyl acrylate (EA)) are coarsened at a low resolution level according to a "4-to-1" or "5-to-1" mapping scheme, with each coarsened bead representing 4 to 5 heavy atoms. The ester functional groups and hydrophobic alkyl chains of acrylate monomers are mapped to polar beads and nonpolar beads, respectively; water molecules are modeled using an implicit water model, which does not explicitly include water molecules, but instead shields electrostatic interactions by setting a continuous dielectric constant (ε≈60–80, for example 70) to maximize computational speed. S1-1-2: Establish a low-resolution coarse-grained model of the emulsifier; Based on the chemical structure of emulsifier molecules, a simplified functional module mapping strategy is adopted: hydrophilic head groups (such as sulfate -SO4) are mapped... - sulfonate-SO3 - Carboxylate group -COO - For example, a polyoxyethylene ether chain can be mapped to a hydrophilic bead with a specific charge and polarity; a hydrophobic tail chain (alkyl chain) can be mapped to 1 to 2 hydrophobic beads (depending on the tail chain length); for reactive emulsifiers, their polymerizable double bond functional groups (allyl, acrylamide, vinyl, methacrylate, etc.) can be mapped to a reactive bead. Overall, each emulsifier molecule is represented by 3 to 6 coarse-grained beads in the low-resolution coarse-grained model; For low-resolution coarse-grained models, Martini 3 force field or its extended version calibrated with coarse-grained parameters is used; key parameters include: Lennard-Jones potential energy parameters between different bead types (hydrophilic, hydrophobic, polar, nonpolar, reactive) – interaction potential well depth ε and equilibrium distance σ; for reactive emulsifiers, the spring constant bonding parameters of reactive beads also need to be calibrated. S1-1-3: Constructing a low-resolution coarse-grained model of the ternary interface of acrylate monomer-water-emulsifier; Emulsifier molecules are arranged on both sides of the oil / water interface according to a preset interface area / emulsifier molecule ratio. Counterions are added to neutralize the net charge of the system. Then, energy minimization, short-time equilibrium simulation (50-100 ns, e.g., 80 ns) and production simulation (1000-2000 ns, e.g., 1600 ns) are performed, and low-resolution coarse-grained simulation trajectories are recorded. The method for constructing a high-resolution coarse-grained model of the acrylate monomer-water-emulsifier ternary interface model in step S1 includes the following steps: S1-2-1: Initial configuration of the high-resolution coarse-grained model; By refining the final configuration of the low-resolution simulation by bead splitting, the initial configuration of the high-resolution coarse-grained model is obtained: one coarse-grained bead in the low-resolution coarse-grained model is converted into multiple (e.g., 2 to 4) second coarse-grained beads in the high-resolution coarse-grained model according to a preset splitting rule. The initial position of the second coarse-grained beads is determined by geometric interpolation and local energy minimization. The splitting rules for coarse-grained beads are as follows: (1) For acrylate monomers, the original polar bead is split into two second coarse-grained beads, one of which represents the carbonyl group and oxygen atom of the ester group, and the other represents the alkoxy group of the ester group; the hydrophobic alkyl chain is refined at a density of one bead for every 2 carbon atoms; (2) For emulsifiers, the hydrophilic head group is split into 2 to 3 second coarse-grained beads to distinguish the charge distribution inside the head group; the hydrophobic tail chain is refined at a density of one second coarse-grained bead for every 3 to 4 carbon atoms; (3) For reactive emulsifiers, the polymerizable double bond is retained as an independent reactive bead with a clear direction; (4) Overall, each emulsifier molecule is represented by 10 to 20 second coarse-grained beads in the high-resolution coarse-grained model; S1-2-2: After short-term rebalancing (20-50 ns, e.g., 30 ns) of the constructed high-resolution coarse-grained model, conduct production simulation (1000-2000 ns, e.g., 1500 ns) and record the high-resolution coarse-grained simulation trajectory. S2: Extract low-resolution hierarchical performance indicators based on the low-resolution coarse-grained simulation trajectory, and extract high-resolution hierarchical performance indicators based on the high-resolution coarse-grained simulation trajectory. In step S2, the method for extracting low-resolution hierarchical performance metrics and high-resolution hierarchical performance metrics includes the following steps: S2-1: Extract low-resolution hierarchical performance indicators; Low-resolution layer performance metrics include: (1) Low-resolution interfacial tension γ low The trajectory was calculated from a low-resolution coarse-grained simulation using stress tensor analysis. (2) Equilibrium time τ of emulsifier self-assembly at the interface eq (The time required from the start of the simulation to the steady state of interface coverage). (3) For reactive emulsifiers, calculate the low-resolution exposure index E of polymerizable double bonds. dblow Calculate using the following formula:

[0042] Where d min (t) represents the centroid distance between the reactive bead containing the polymerizable double bond and the nearest monomer bead in frame t, and d cThe contact characteristic distance is 0.6 nm. For traditional physically adsorbed emulsifiers, this indicator defaults to E. dblow Set to 0; (4) Critical micelle concentration (CMC) estimated based on low-resolution simulation low (Low-resolution simulations were performed at multiple gradient total emulsifier concentrations. The concentrations of unassociated (free) emulsifier molecules were statistically analyzed, and the inflection point where the free concentration-total concentration curve changed from a linear increase to a plateau (dynamic equilibrium) was taken as the CMC.) low ); S2-2: Extract high-resolution hierarchical performance indicators; High-resolution hierarchical performance metrics include: (1) High-resolution interfacial tension γ high This serves as a refinement of the low-resolution values. (2) Interface formation energy ∆G calculated by high-resolution coarse-grained simulation trajectory ads The umbrella sampling method was used, with the distance between the centroid of the emulsifier molecule and the center of the interface as the reaction coordinate, to calculate the change in free energy for migration from the aqueous phase to the interface and stable adsorption. (3) For reactive emulsifiers, calculate the high-resolution exposure index E of polymerizable double bonds. dbhigh and co-aggregation tendency index I copol (Estimated by the contact frequency and binding free energy between reactive beads and monomer beads); For traditional physical adsorption emulsifiers, these two indicators are not included in the scoring or are set to default values ​​(e.g., the default values ​​for high-resolution exposure index and copolymerization tendency index are both set to 0); The high-resolution exposure index E of the polymerizable double bond dbhigh Low-resolution exposure index E for polymerizable double bonds dblow The same formula is used for calculation, only the simulated trajectory is replaced with a high-resolution trajectory; The copolymerization tendency index I copol Defined as:

[0043] Where ΔG pair The binding free energy (kJ·mol⁻¹) of reactive beads to monomer beads was estimated from high-resolution coarse-grained simulations. copol The scaling factor for the Sigmoid function (dimensionless, recommended range k). copol ∈[1, 5], which can be determined through cross-validation on the training set), control I copol Sensitivity to input variables, k copol The larger the value of α, the steeper the curve. dbHigh-resolution double bond exposure index E dbhigh Weighting coefficients (dimensionless, recommended range α) db ∈[0.3, 0.7]), for example α db =0.5, representing the relative contribution of double bond exposure to copolymerization tendency; (4) Interface arrangement characteristic parameters at high resolution: the average tilt angle θ between the molecular long axis and the interface normal (obtained by calculating the angle between the vector connecting the center of the hydrophilic head group and the end of the hydrophobic tail chain and the interface normal), and the standard deviation σ of the tilt angle distribution. θ ; (5) Interface desorption free energy barrier ∆G desorp : Calculate the free energy change of emulsifier migrating from the interface to the aqueous phase using umbrella-shaped sampling; S3: Establish a step-by-step coarse-grained comprehensive scoring model, generate a low-resolution score based on the low-resolution hierarchical performance index for the first-level screening, generate a high-resolution score based on the high-resolution hierarchical performance index for the second-level screening, and calculate the comprehensive score. In step S3, the method for establishing the step-by-step coarse-grained comprehensive scoring model includes the following steps: S3-1 Construct a low-resolution scoring model and generate a low-resolution score based on the low-resolution hierarchical performance index. low Used for the first round of high-throughput filtration;

[0044] Where s(·) is the min-max normalization function, which normalizes all molecules in the candidate library across samples to [0,1], with the direction being "the larger the better"; δ reactive The reactive emulsifier is identified by a value of 1 (reactive type, 0 for traditional physical adsorption type); w1 represents the low-resolution interfacial tension s(γ). low The weights of ) are (recommended 0.2–0.4, e.g. 0.3); w2 is the self-assembly equilibrium time s(τ) eq The weights of ) are (0.1–0.3, e.g. 0.2); w3 is the critical micelle concentration s (CMC). low The weights of ) (0.2–0.4, e.g. 0.3); w4 is the low-resolution double bond exposure index s(E) dblow The weights of the emulsifiers are set (0.1–0.3, e.g., 0.2 for reactive emulsifiers only); each weight can be automatically adjusted according to the emulsifier type or uniformly set to the default value. According to Score low The candidate pool is divided into three levels: "priority recommendation" (the highest 20-30%, for example, the highest 25%), "deferred" and "elimination" (the lowest 30-40%, for example, the lowest 30%). S3-2 A high-resolution scoring model is constructed for molecules in the "Preferred Recommendation" level, and a high-resolution score is generated based on the high-resolution hierarchical performance index. high :

[0045] Where s(·) is the min-max normalization function, which normalizes all molecules in the candidate library across samples to [0,1], with the direction being "the larger the better"; w5 is the high-resolution interfacial tension s(γ). high The weight of ) (recommended 0.1–0.3, e.g. 0.2); w6 is the interface formation free energy s(ΔG) ads The weights of w7 and w8 are: w7 = w8 + w8 + w9 ... desorp The weights of ) (0.1–0.3, e.g. 0.2); w9 is the high-resolution double bond exposure index s(E) dbhigh The weight of ) (0.1–0.3, e.g. 0.1, reactive only); w 10 The copolymerization tendency index s(I) copol The weights of the emulsifiers are set (0.1–0.3, e.g., 0.1 for reactive emulsifiers only); each weight can be automatically adjusted according to the emulsifier type or uniformly set to the default value. S3-3 Construct a comprehensive scoring model and generate a comprehensive score. final ;

[0046] Where α is the weight of the low-resolution score in the final score (0 ≤ α ≤ 1, which can be dynamically adjusted according to the confidence level of the molecule in the low-resolution simulation); Score low norm Score high norm Score low Score high The value after global normalization ([0,1]); Step S3 ends with pressing Score final Output the recommended sort list; S4: Extract the structural descriptor of the emulsifier molecule, and use the low-resolution hierarchical performance index and high-resolution hierarchical performance index as coarse-grained performance labels to construct a training dataset; In step S4, the method for constructing the training dataset includes the following steps: S4-1: Extract the structure descriptor vector x of the emulsifier molecule, including: (1) One-dimensional descriptors: molecular weight, number of rotatable bonds, number of hydrogen bond donors / acceptors; (2) Two-dimensional topological descriptor: topological polar surface area TPSA, ALOGP hydrophobicity parameter, molecular refractive index, HLB value (calculated by Griffin method); (3) Ion / charge descriptor: ion type (anionic / nonionic / cationic) and formal charge number of the hydrophilic head group; (4) Reactivity descriptor (only for reactive emulsifiers): polymerizable double bond type, double bond position index, steric hindrance parameter; the double bond position index is the number of chemical bonds between the polymerizable double bond and the hydrophilic head group, and the steric hindrance parameter is the number of substituents on the carbon of the double bond (monosubstituted, 1,1-disubstituted, etc.); for traditional physical adsorption emulsifiers, all of the above reactivity descriptors are set to 0; (5) 3D / Conformation Descriptor: Dipole moment and molecular size (radius of gyration) of the lowest energy conformation of the molecule obtained through fast force field optimization; This conformation optimization is only calculated using conventional molecular structure descriptors and does not involve molecular dynamics simulation; S4-2: Construct a performance label vector y from the performance metrics extracted in step S2, including low-resolution performance metrics (γ). low τ eq CMC low E dblow ) and high-resolution performance metrics (γ) high ΔG ads , θ, ΔG desorp E dbhigh I copol ); among which, the reaction-related indicators are only effective for reactive emulsifiers, and the corresponding values ​​for traditional physical adsorption emulsifiers are set to 0; Score low and Score high Calculated using the fixed scoring formula in step S3, it is not used as the prediction target of the machine learning model; S4-3: Construct the original training dataset D0 = {(x i , y i , t i )}, where t i Here, x is the emulsifier type label (traditional physical adsorption emulsifier / reactive emulsifier), and i is the sample index, representing the i-th sample in the training dataset (i = 1, 2, …, N; N is the total number of samples in the training set). i Let y be the structure descriptor vector of the i-th emulsifier molecule (extracted from S4-1). i For its corresponding performance label vector (calculated from S2 and S3), t i ∈ {0, 1} is its type label (0 represents traditional physical adsorption emulsifier, 1 represents reactive emulsifier); S5: Employ an active learning strategy to select high-information samples and iteratively expand the training dataset; In step S5, the method of using an active learning strategy to select high-information samples and iteratively expand the training dataset includes the following steps: S5-1 Training the initial model: Select a small number (10-15, for example 15) of representative emulsifiers with large structural differences from the candidate molecules, perform low-resolution coarse-grained simulation and obtain performance indicators as labels, which together with the descriptor vector x and type label t from step S4 constitute the initial training set D0. Two models are trained based on D0: M low M is used to predict low-resolution performance metrics. high Used to predict high-resolution performance metrics; both models use (x,t) as input features and can automatically distinguish emulsifier types; S5-2 Identify the molecules with the least model uncertainty: For unlabeled molecules in candidate pool U, use M... low Predict its low-resolution performance metrics, and provide the uncertainty U. low (x), representing the model's confidence level in the prediction; U low (x) is measured by the dispersion (variance) of the predicted values ​​of multiple model members. The more dispersed the predictions, the less confident the model is. S5-3 Molecules were selected and labeled in two stages: First-level sampling: Select U from U low (x) The largest k1 molecules (k1≥5, e.g. k1=15) are the most uncertain group of molecules in the model. Perform low-resolution coarse-grained simulations on them to obtain low-resolution performance labels. Second-level sampling: Among the molecules already labeled in the first level, scores are calculated using a low-resolution method. low Sort the data from high to low, take the first k2 (k2≤k1, for example k2=5) and perform high-resolution coarse-grained simulation to obtain high-resolution performance labels; the initial configuration of the high-resolution simulation is obtained by bead splitting operation from the final configuration at the end of the low-resolution simulation. S5-4 Update the model: Incorporate the newly labeled samples from this round into the training dataset, and retrain or fine-tune M. low and M high ; S5-5 Termination Criteria: Repeat S5-2 to S5-4 until the model achieves the required accuracy in cross-validation (e.g., coefficient of determination R² ≥ 0.8 or mean absolute error MAE ≤ 3.5 kJ / mol), then output the final M value. low and M high ; S6: The candidate emulsifier library is subjected to hierarchical virtual screening using the finally trained machine learning model. The screening is carried out through the first level of low-resolution coarse-grained model and the second level of high-resolution coarse-grained model, and a recommended list of emulsifiers is output. In step S6, the method for performing hierarchical virtual screening of the acrylic emulsion emulsifier library using the finally trained machine learning model includes the following steps: S6-1 First-level screening: using M low The model quickly predicts all unlabeled molecules in the candidate library and calculates the score. low Eliminate the bottom 30% to 40% (e.g., 30%) of the applicants; S6-2 Second-stage screening: For the remaining molecules that passed the first-stage screening, M... high The model predicts high-resolution performance metrics and calculates the score. high and overall score final Output the final recommendation list.

[0047] The above method also includes step S7: experimental verification and closed-loop optimization of emulsion polymerization.

[0048] In step S7, the method for experimental verification and loop closure optimization of emulsion polymerization includes the following steps: S7-1 Select several emulsifiers (including traditional physical adsorption type emulsifiers and reactive emulsifiers) that rank high in the virtual screening comprehensive score, and conduct emulsion polymerization experiments according to the target acrylic emulsion formulation; S7-2 compares experimental data with predicted results, analyzes the sources of deviation, and optimizes coarse-grained force field parameters or machine learning models accordingly. S7-3 adds the experimental data to the training dataset and retrains the machine learning model.

[0049] The following is a specific example of using the finally trained machine learning model to perform hierarchical virtual screening of a candidate emulsifier library, which includes traditional physical adsorption emulsifiers and reactive emulsifiers.

[0050] I. Specific data includes: (1) Establish a virtual library containing 500 candidate emulsifiers, of which 300 are traditional physical adsorption emulsifiers (SDS, SDBS, AES, AEO series, etc.) and 200 are reactive emulsifiers (COPS-1, SE-10N, AMPS, etc.). The acrylate monomer system is set as MMA:BA:MAA=45:50:5; (2) First-stage screening of the low-resolution coarse-grained model: Martini 3 force field was used, with each molecule represented by 4-5 beads, and reactive emulsifiers included an additional reactive bead. Batch runs of 200 ns low-resolution simulations were performed to calculate the score. low The bottom 150 are eliminated, and the remaining 350 advance to the next round. (3) Second-level screening of high-resolution coarse-grained models: High-resolution coarse-grained models were constructed for 350 molecules (12-15 beads per molecule), and simulations were run for 300 ns. General indicators (γ) were extracted. high ΔG ads , θ, ΔG desorp ) and response-specific indicators (E dbhigh I copol Calculate the score. high and overall score final The top 10 outputs are ranked (a mixed ranking of traditional physical adsorption emulsifiers and reactive emulsifiers).

[0051] Active learning iteration results: Initially, 10 molecules (containing two types) were randomly selected for full-process annotation. After 8 rounds of iteration (k1=15, k2=5 in each round), M... high The model applies to the traditional physical adsorption emulsifier ΔG ads The prediction error (MAE) was 3.2 kJ / mol for reactive emulsifier I. copol The predicted R² is 0.86.

[0052] II. Experimental Verification and Data Analysis: To verify the effectiveness of the virtual filtering method, the Score output by the virtual filtering method is... final The ranking results were systematically compared with the experimental performance of emulsion polymerization. The experimental design is as follows: From the recommendation group (Score) final Six emulsifiers were selected for emulsion polymerization experiments: two from the top five (one reactive and one conventional), two from the middle group (randomly selected from the 40%–60% ranking range), and two from the elimination group (randomly selected from the lowest 35%, one of which was an extreme case of demulsification). The SDS+COPS-1 compound system with known good performance was used as a positive control. Each emulsifier group was subjected to emulsion polymerization experiments with the same formulation (MMA:BA:MAA=45:50:5, emulsifier dosage 3wt%), and three replicates were performed for each group.

[0053] There is a clear physical causal relationship between virtual screening metrics and experimentally measurable performance: γ low / γ high (Interfacial tension) is negatively correlated with latex particle size; ΔGads (Interfacial formation free energy) is negatively correlated with emulsion storage stability; ΔG desorp The desorption free energy barrier is positively correlated with the centrifugal stability of the emulsion; E dbhigh The (high-resolution double bond exposure index) is negatively correlated with the amount of free emulsifier residue; I copol (Copolymerization tendency index) is positively correlated with the water resistance of the coating film; τ eq (Self-assembly equilibrium time) is negatively correlated with particle size distribution (PDI); CMC low The critical micelle concentration (CMC) is related to polymerization stability; the lower the CMC, the less emulsifier is needed to reach steady-state micelleification, and the more stable the polymerization process. dblow The low-resolution double bond exposure index is positively correlated with the copolymerization anchoring efficiency of reactive emulsifiers; θ (molecular tilt angle) is negatively correlated with the orderliness of the interfacial adsorption layer, and the smaller θ is, the denser the molecular arrangement and the more stable the interface.

[0054] Table 1 Summary of Emulsion Polymerization Experiment Results 1 Recommendation Group Reactive 82 0.07 98.2 0.0 No layering 0.12 2 Recommendation Group Traditional 105 0.12 96.8 0.2 No layering 1.1 3 intermediate group Reactive 142 0.17 94.0 0.5 Micro-stratification 0.6 4 intermediate group Traditional 178 0.22 91.5 1.2 Micro-stratification 2.2 5 Elimination Group Reactive 310 0.38 84.0 4.0 Layering 1.7 6 Elimination Group Traditional — — 78.5 8.0 Breast breaking — Table 2. Coating performance verification results (focusing on reactive emulsifiers) Ⅰ Recommendation Group Reactive 6.2 0 No foaming Ⅱ Recommendation Group Reactive 5.8 0 No foaming Ⅲ intermediate group Reactive 9.5 1 Microbubbling Ⅳ intermediate group Reactive 10.8 1 Microbubbling Ⅴ Elimination Group Reactive 16.0 2 foaming Ⅵ Elimination Group Reactive 16.5 2 foaming Note: The samples in Table 2 are six separately selected reactive emulsifiers, numbered (Ⅰ~Ⅵ) and do not correspond to those in Table 1. They are used only for specific verification of the coating performance of reactive emulsifiers.

[0055] Experimental results showed that both emulsifiers in the recommended group successfully prepared stable emulsions with average particle sizes of 82–105 nm (PDI 0.07–0.12) and monomer conversion rates of 96.8%–98.2%. After 30 days of storage, neither emulsifier showed stratification, gelation, or demulsification. In the eliminated group, one emulsifier (number 6) completely demulsified, while the other had a particle size of 310 nm (PDI 0.38) and a gelation rate of 4.0%. The two emulsifiers in the intermediate group had all indicators between the recommended and eliminated groups. Regarding the residual amount of free emulsifier, reactive emulsifiers (0.12%–1.7%) were generally lower than traditional emulsifiers (1.1%–2.2%), and the residual amount of reactive emulsifiers in the recommended group (0.12%) was only about one-tenth of that of traditional emulsifiers, confirming the mechanism of "double bond copolymerization anchoring → reduction of free emulsifier." The water absorption rate of the coating film of the recommended reactive emulsifier group (5.8%–6.2%) was significantly lower than that of the intermediate reactive group (9.5%–10.8%) and the eliminated reactive group (16.0%–16.5%), and was also lower than that of the virtual screening group I. copol The order was completely consistent, verifying the mapping relationship of "copolymerization tendency index → ​​coating water resistance".

[0056] The above results show that the experimental results of the recommended group, intermediate group and elimination group are significantly different, and the experimental indicators (particle size, monomer conversion rate, gelation rate, free emulsifier residue and coating water absorption rate) are consistent with the virtual score ranking direction, which proves that the virtual score has a continuous gradient discrimination ability rather than a random hit, and verifies the effectiveness and physical mechanism of the screening method.

[0057] The experimental data is fed back to step S7, and M is retrained after correcting the force field parameters or adding new indices. low / M high The model achieves a closed loop of "virtual screening → experimental verification → model correction". This is not only a means of verifying effectiveness, but also a mechanism for continuous optimization of the method.

[0058] Estimated total time: Taking the 500 candidate emulsifier library from Example 1 as an example, on a medium-sized computing cluster with 8 GPUs: low-resolution initial screening takes approximately 14 days, high-resolution secondary screening takes approximately 27 days, active learning iteration takes approximately 9 days, final prediction takes approximately 0.5 days, experimental verification takes approximately 11 days, and the entire process takes approximately 61 days (approximately 2 months). The actual computational speed of coarse-grained MD on GPUs is significantly higher than the simulation time: low-resolution CG systems have a scale of approximately 1000–2000 beads, with a computational speed of approximately 500–2000 ns / day / GPU (referencing GROMACS benchmark tests); high-resolution CG systems have a scale of approximately 3000–5000 beads, with a computational speed of approximately 200–800 ns / day / GPU. The core advantage of coarse-grained MD compared to all-atom MD is that the number of beads is reduced by 5–20 times, and the time step can be increased to 20–40 fs (compared to 2 fs for all-atom MD), thus improving the actual computational speed by approximately 2–3 orders of magnitude.

[0059] Traditional trial-and-error methods require each emulsifier to go through synthesis / procurement (1–2 weeks) → pilot polymerization (2–3 days) → performance testing (3–5 days), with a cycle of about 3–4 weeks per emulsifier; evaluating 500 candidates would take about 1 year (32 parallel groups).

[0060] The method of this invention requires only 2 months (8 GPUs) with the same candidate scale, and the final experimental verification quantity is reduced from 500 to 6 (a reduction of 98.8%), which fully demonstrates the technical advantage of coarse-grained MD model in improving the computation speed by 2-3 orders of magnitude compared with full atom simulation.

Claims

1. A method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration, characterized in that... Includes the following steps: S1: Construct low-resolution coarse-grained models and high-resolution coarse-grained models of the ternary interface model of acrylate monomer-water-emulsifier, respectively, and obtain the corresponding low-resolution coarse-grained simulation trajectory and high-resolution coarse-grained simulation trajectory through coarse-grained molecular dynamics simulation. S2: Extract low-resolution hierarchical performance indicators based on the low-resolution coarse-grained simulation trajectory, and extract high-resolution hierarchical performance indicators based on the high-resolution coarse-grained simulation trajectory. S3: Establish a step-by-step coarse-grained comprehensive scoring model, generate a low-resolution score based on the low-resolution hierarchical performance index for the first-level screening, generate a high-resolution score based on the high-resolution hierarchical performance index for the second-level screening, and calculate the comprehensive score. S4: Extract the structural descriptor of the emulsifier molecule, and use the low-resolution hierarchical performance index and high-resolution hierarchical performance index as coarse-grained performance labels to construct a training dataset; S5: Employ an active learning strategy to select high-information samples and iteratively expand the training dataset; S6: The final trained machine learning model is used to perform hierarchical virtual screening of the candidate emulsifier library. The model first screens the candidate emulsifier library with a low-resolution coarse-grained model and then screens the candidate emulsifier library with a high-resolution coarse-grained model. The model then outputs a recommended list of emulsifiers.

2. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 1, characterized in that: The emulsifier includes traditional physical adsorption emulsifiers and reactive emulsifiers; the molecular structure of the reactive emulsifier contains at least one carbon-carbon double bond that can participate in free radical polymerization.

3. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 1, characterized in that: In step S1, in the low-resolution coarse-grained model, each emulsifier molecule is represented by 3 to 6 coarse-grained beads; in the high-resolution coarse-grained model, each emulsifier molecule is represented by 10 to 20 coarse-grained beads; the initial configuration of the high-resolution coarse-grained model is obtained by performing a bead splitting operation on the final configuration of the low-resolution simulation.

4. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 1, characterized in that: The method for constructing a low-resolution coarse-grained model of the acrylate monomer-water-emulsifier ternary interface model in step S1 includes the following steps: S1-1-1: Establish a low-resolution coarse-grained model of acrylate monomers and water; Acrylic ester monomer molecules are coarsened at a low resolution level using a "4-to-1" or "5-to-1" mapping scheme, with each coarsened bead representing 4 to 5 heavy atoms. The ester functional groups and hydrophobic alkyl chains of acrylate monomers are mapped as polar beads and nonpolar beads, respectively; the water molecule adopts an implicit water model. S1-1-2: Establish a low-resolution coarse-grained model of the emulsifier; Overall, each emulsifier molecule is represented by 3 to 6 coarse-grained beads in the low-resolution coarse-grained model; S1-1-3: Constructing a low-resolution coarse-grained model of the ternary interface of acrylate monomer-water-emulsifier; Emulsifier molecules are arranged on both sides of the oil / water interface according to a preset interface area / emulsifier molecule ratio. Counterions are added to neutralize the net charge of the system. Then, energy minimization, short-time equilibrium simulation and production simulation are performed, and the low-resolution coarse-grained simulation trajectory is recorded. The method for constructing a high-resolution coarse-grained model of the acrylate monomer-water-emulsifier ternary interface model in step S1 includes the following steps: S1-2-1: Initial configuration of the high-resolution coarse-grained model; By refining the final configuration of the low-resolution simulation by bead splitting, the initial configuration of the high-resolution coarse-grained model is obtained: a coarse-grained bead in the low-resolution coarse-grained model is converted into multiple second coarse-grained beads in the high-resolution coarse-grained model according to a preset splitting rule. The initial position of the second coarse-grained beads is determined by geometric interpolation and local energy minimization. S1-2-2: After a short-term rebalancing of the constructed high-resolution coarse-grained model, production simulation is performed, and the high-resolution coarse-grained simulation trajectory is recorded.

5. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 1, characterized in that: In step S2, the method for extracting low-resolution hierarchical performance metrics and high-resolution hierarchical performance metrics includes the following steps: S2-1: Extract low-resolution hierarchical performance indicators; Low-resolution layer performance metrics include: (1) Low-resolution interfacial tension γ low ; (2) Equilibrium time τ of emulsifier self-assembly at the interface eq ; (3) For reactive emulsifiers, calculate the low-resolution exposure index E of polymerizable double bonds. dblow Calculate using the following formula: ， Where N frames d represents the total number of frames captured. min (t) represents the t-th frame (t≤N), d min (t) represents the centroid distance between the reactive bead containing the polymerizable double bond and the nearest monomer bead in frame t, and d c The contact characteristic distance is 0.6 nm. For traditional physically adsorbed emulsifiers, this indicator defaults to E. dblow Set to 0; (4) Critical micelle concentration (CMC) estimated based on low-resolution simulation low ; S2-2: Extract high-resolution hierarchical performance indicators; High-resolution hierarchical performance metrics include: (1) High-resolution interfacial tension γ high ; (2) Interface formation energy ∆G calculated by high-resolution coarse-grained simulation trajectory ads ; (3) For reactive emulsifiers, calculate the high-resolution exposure index E of polymerizable double bonds. dbhigh and co-aggregation tendency index I copol ; The high-resolution exposure index E of the polymerizable double bond dbhigh Low-resolution exposure index E for polymerizable double bonds dblow The same formula is used for calculation, only the simulated trajectory is replaced with a high-resolution trajectory; The copolymerization tendency index I copol Defined as: ， Where ΔG pair The binding free energy of reactive beads and monomer beads estimated from high-resolution coarse-grained simulations; k copol α is the scaling factor for the sigmoid function. db High-resolution double bond exposure index E dbhigh Weighting coefficients; (4) Interface arrangement characteristic parameters at high resolution: the average tilt angle θ between the molecular long axis and the interface normal, and the standard deviation σ of the tilt angle distribution. θ ; (5) Interface desorption free energy barrier ∆G desorp .

6. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 5, characterized in that: In step S3, the method for establishing a step-by-step coarse-grained comprehensive scoring model includes the following steps: S3-1 Construct a low-resolution scoring model and generate a low-resolution score based on the low-resolution hierarchical performance index. low Used for the first round of high-throughput filtration; ， Where s(·) is the min-max normalization function, which normalizes all molecules in the candidate library across samples to [0,1], with the direction being "the larger the better"; δ reactive This is an identifier for reactive emulsifiers; w1 represents the low-resolution interfacial tension s(γ). low The weights of ) are: w2 is the self-assembly equilibrium time s(τ). eq The weights of ) are: w3 is the critical micelle concentration s (CMC). low The weights of ) are: w4 is the low-resolution double bond exposure index s(E) dblow The weight of ); According to Score low The candidate pool is divided into three levels: "priority recommendation", "deferred" and "elimination". S3-2 Constructs a high-resolution scoring model for molecules in the "Preferred Recommendation" level, and generates a high-resolution score based on the high-resolution hierarchical performance index. high : ， Where s(·) is the min-max normalization function, which normalizes all molecules in the candidate library across samples to [0,1], with the direction being "the larger the better"; w5 is the high-resolution interfacial tension s(γ). high The weight of ); w6 is the interface formation free energy s(ΔG) ads The weight of ) is ; w7 is the weight of the molecular tilt angle s(θ); w8 is the desorption free energy s(ΔG) desorp The weight of ); w9 represents the high-resolution double bond exposure index s(E). dbhigh The weight of ); w 10 The copolymerization tendency index s(I) copol The weight of ); S3-3 Construct a comprehensive scoring model and generate a comprehensive score. final ; ， Where α is the weight of the low-resolution score in the final score; Score low norm Score high norm Score low Score high The value after global normalization.

7. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 5, characterized in that: In step S4, the method for constructing the training dataset includes the following steps: S4-1: Extract the structure descriptor vector x of the emulsifier molecule, including: (1) One-dimensional descriptors: molecular weight, number of rotatable bonds, number of hydrogen bond donors / acceptors; (2) Two-dimensional topological descriptors: topological polar surface area TPSA, ALOGP hydrophobicity parameter, molecular refractive index, HLB value; (3) Ion / charge descriptor: ion type and formal charge number of the hydrophilic head group; (4) Reactivity descriptor: can aggregate double bond type, double bond position index, steric hindrance parameter; (5) 3D / conformation descriptor: dipole moment and molecular size in the lowest energy conformation of the molecule obtained through fast force field optimization; S4-2: Construct a performance label vector y from the performance metrics extracted in step S2, including the low-resolution performance metric γ. low τ eq CMC low E dblow and high-resolution performance metrics γ high ΔG ads , θ, ΔG desorp E dbhigh I copol Among them, the reaction-related indicators are only effective for reactive emulsifiers; the corresponding values ​​for traditional physical adsorption emulsifiers are set to 0. low and Score high Calculated using the fixed scoring formula in step S3; S4-3: Construct the original training dataset D0 = {(x i , y i , t i )}, where t i Here, is the emulsifier type label, and is the sample index, representing the i-th sample in the training dataset (i = 1, 2, …, N; N is the total number of samples in the training set). x i Let y be the structure descriptor vector of the i-th emulsifier molecule (extracted from S4-1). i For its corresponding performance label vector (calculated from S2 and S3), t i ∈ {0, 1} is its type label (0 represents traditional physical adsorption emulsifier, 1 represents reactive emulsifier).

8. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 1, characterized in that: In step S5, the method of using an active learning strategy to select high-information samples and iteratively expand the training dataset includes the following steps: S5-1 Training the initial model: Select a small number of representative emulsifiers with large structural differences from the candidate molecules, perform low-resolution coarse-grained simulation and obtain performance indicators as labels, which together with the descriptor vector x and type label t from step S4 constitute the initial training set D0. Two models are trained based on D0: M low M is used to predict low-resolution performance metrics. high Used to predict high-resolution performance metrics; both models use (x,t) as input features and can automatically distinguish emulsifier types; S5-2 Identify the molecules with the least model uncertainty: For unlabeled molecules in candidate pool U, use M... low Predict its low-resolution performance metrics, and provide the uncertainty U. low (x), representing the model's confidence level in the prediction; U low (x) The more dispersed the predictions of multiple model members, the less certain the model is; S5-3 Molecules were selected and labeled in two stages: First-level sampling: Select U from U low (x) The largest k1 molecules (k1≥5) are the most uncertain group of molecules in the model. Low-resolution coarse-grained simulations are performed on them to obtain low-resolution performance labels. Second-level sampling: Among the molecules already labeled in the first level, scores are calculated using a low-resolution method. low Sort the data from high to low, take the first k2 (k2≤k1) and perform high-resolution coarse-grained simulation to obtain high-resolution performance labels; the initial configuration of the high-resolution simulation is obtained by bead splitting operation from the final configuration at the end of the low-resolution simulation. S5-4 Update the model: Incorporate the newly labeled samples from this round into the training dataset, and retrain or fine-tune M. low and M high ; S5-5 Termination condition: Repeat S5-2 to S5-4 until the model achieves the required accuracy in cross-validation, then output the final M value. low and M high .

9. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 1, characterized in that: In step S6, the method for hierarchical virtual screening of the acrylic emulsion emulsifier library using the finally trained machine learning model includes the following steps: S6-1 First-level screening: using M low The model quickly predicts all unlabeled molecules in the candidate library and calculates the score. low Eliminate the bottom 30% to 40% of the applicants; S6-2 Second-stage screening: For the remaining molecules that passed the first-stage screening, M... high The model predicts high-resolution performance metrics and calculates the score. high and overall score final Output the final recommendation list.

10. The method for screening acrylic emulsion emulsifiers based on coarse-grained molecular dynamics models and machine learning acceleration according to claim 1, characterized in that... It also includes step S7: experimental verification and closed-loop optimization of emulsion polymerization; In step S7, the method for experimental verification and loop closure optimization of emulsion polymerization includes the following steps: S7-1 Select several emulsifiers that rank high in the virtual screening comprehensive score and conduct emulsion polymerization experiments according to the target acrylic emulsion formulation; S7-2 compares experimental data with predicted results, analyzes the sources of deviation, and optimizes coarse-grained force field parameters or machine learning models accordingly. S7-3 adds the experimental data to the training dataset and retrains the machine learning model.

Images