A liquid crystal monomer performance prediction method and a mixed crystal performance prediction method based on GAT

Abstract

The application belongs to the technical field of liquid crystal material and data driving material design, and discloses a liquid crystal monomer performance prediction method and a mixed crystal performance prediction method based on GAT, which represents the liquid crystal monomer as a molecular graph, encodes the chemical bond connection relationship and chemical environment by using a graph attention network, and introduces the polarization rate, the dipole moment and the orbital energy level and other quantitative calculation physical information in the pre-training stage; a PIML molecular prediction model is established under a structure perception data division strategy, and accurate prediction of key properties such as birefringence, dielectric parameters and clearing point is realized. On the basis of constructing a mixed crystal prediction model based on KernelSVR prediction, forward mixed crystal performance prediction or reverse mixed crystal formula generation is realized, which significantly reduces the formula search cost and improves the research and development efficiency.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of liquid crystal materials and data-driven material design, and in particular to a method for predicting the performance of liquid crystal monomers and mixed crystals based on GAT. Background Technology

[0002] High birefringence liquid crystal materials typically rely on a high-rigidity conjugated framework and strong polar end groups to improve optical anisotropy and dielectric response, but this often comes with problems such as increased rotational viscosity, difficulty in temperature range control, and conflicts between multiple performance objectives. In actual research and development, it is difficult for a single liquid crystal monomer to simultaneously meet comprehensive indicators such as high birefringence, suitable dielectric parameters, sufficient clearing point, and low rotational viscosity. Furthermore, it is difficult for a single liquid crystal monomer to achieve compatibility among the aforementioned mutually restrictive indicators. Therefore, multi-component mixed crystals are commonly used, which introduce different functional monomers and adjust their ratios to achieve synergistic optimization of "contribution superposition and suppression compensation".

[0003] The combination space for multi-component mixed crystals is vast, with nonlinear coupling and compatibility constraints between components. Traditional data-driven approaches often employ the QSAR framework: obtaining molecular structures through quantitative calculations, computing numerous hand-crafted descriptors, selecting a small number of descriptors, and then using multiple linear regression to establish a birefringence prediction model. The dataset is then divided into training and validation sets at a fixed ratio to evaluate external prediction capabilities. While this approach can achieve rapid predictions within a certain range, it is heavily reliant on descriptor engineering and linear assumptions, and often struggles to naturally express the nonlinear contributions of molecular topology and local chemical environment to physical properties. Furthermore, it is difficult to directly generalize to the reverse formulation problem of "monomer structure-ratio-mixed crystal properties." On the other hand, reverse design / generative design patents for general materials have proposed learning the joint distribution of "structure-property" or generating structural samples that satisfy target properties in the latent space. However, when directly applied to liquid crystal mixed crystal systems, further solutions are needed: effective characterization of the liquid crystal molecular graph structure, physical consistency constraints related to key properties such as birefringence, and computationally expressible process constraints such as mixed crystal compatibility and phase stability. Summary of the Invention

[0004] The purpose of this invention is to address the technical deficiencies in the prior art by providing a method for predicting the performance of liquid crystal monomers and mixed crystals based on GAT.

[0005] The technical solution adopted to achieve the purpose of this invention is: A method for predicting the performance of liquid crystal monomers based on GAT includes the following steps: Step 1: Collect liquid crystal monomer molecules and their corresponding physical property parameters, perform preprocessing, and establish a monomer dataset; Step 2: Obtain the physical parameters of each liquid crystal monomer molecule in Step 1, and add the physical parameters as feature dimensions to the monomer dataset to establish a physical enhancement dataset. Step 3: Divide the physics augmentation dataset from Step 2 into a training set and a test set; Step 4: Convert the liquid crystal molecules into a molecular graph. Based on the molecular graph, construct a graph attention network (GAT) with embedded physical knowledge. Cascade the GAT and the regression predictor (MLP) to form a PIML molecular prediction model. The GAT uses the molecular graph as nodes and outputs molecular embeddings as feature vectors. Each molecule embedding is connected to one or more regression predictors (MLPs), and each regression predictor (MLP) is connected to the feature vector. As input, predict the physical properties of liquid crystal monomer molecules; Using the training set samples from step 3 combined with the loss function L total The PIML molecular prediction model was trained and its hyperparameters were optimized. The accuracy of the PIML molecular prediction model was verified using the validation set samples from step 3. Step 5: Input the molecular diagram of the liquid crystal monomer molecule to be verified into the PIML performance prediction model trained and verified in Step 4 to obtain the predicted values ​​of its physical property parameters.

[0006] In the above technical solution, in step 1, the experimental physical property parameters include unusual light n e Ordinary light o Birefringence Δn, dielectric constant ε along the long axis of the molecule Dielectric constant ε along the short axis of the molecule ⊥ Dielectric constant Δε, Clearing point C p elastic constant K 11 and K 33 .

[0007] In the above technical solution, the preprocessing step in step 1 includes data cleaning and standardization.

[0008] In the above technical solution, the physical parameters in step 2 include polarizability (α), dipole moment (µ), rotation tensor (Eg), or HOMO / LUMO energy levels.

[0009] In the above technical solution, in step 3, a structure-aware partitioning strategy based on Morgan molecular fingerprints is used to partition the physically enhanced dataset.

[0010] In the above technical solution, in step 4, L total = L data +λ· Lphys +μ· L eg ,in L total Let be the total loss function during the training phase. L data To monitor losses, L eg λ represents the weight regularization or sparse constraint, and μ represents the weight coefficients, respectively.

[0011] In the above technical solution, in step 4, the physical property parameter includes the cleaning point C. p Birefringence Δn and dielectric constant Δε.

[0012] Another aspect of the present invention includes a method for predicting the performance of mixed crystals based on the liquid crystal monomer performance prediction method, comprising the following steps: Step 6: Based on the predicted physical property parameters of a series of liquid crystal monomer molecules obtained in Steps 1 to 5, select N liquid crystal monomer molecules that meet the physical property parameter requirements as candidate monomers for mixed crystals. Step 7: Construct a mixed crystal formulation through orthogonal experiments and measure the physical properties of the mixed crystal; Step 8: Combine the monomers from each mixed crystal formulation in Step 7 and input them into the trained and validated GAT to obtain the feature vectors of the monomers. By the mass fraction of each monomer Further calculation of the overall eigenvector of the mixed crystal Based on the properties of mixed crystals and the properties of each monomer Construct a mixed crystal dataset, then divide it into training and validation sets; build a mixed crystal prediction model based on Kernel SVR prediction, with the input of the mixed crystal prediction model being... and The output is the mixed crystal properties. The mixed crystal prediction model is trained using the training set, and the error is evaluated using the validation set and experiments. Step 9, Predicting the performance of forward mixed crystals or generating reverse mixed crystal formulations: The feature vector of the candidate single entity obtained in step 6 and Input the trained and validated mixed crystal prediction model to obtain the physical properties of mixed crystals composed of candidate monomers; Alternatively, the desired mixed crystal properties and eigenvectors of candidate monomers can be used. Input the trained and validated mixed crystal prediction model to obtain the values ​​of each candidate monomer. .

[0013] In the above technical solution, in step 7, , For the first iComponent mass fraction.

[0014] In the above technical solution, step 9, the mixed crystal properties include birefringence Δn, dielectric constant Δε, and clearing point C. p Or rotational viscosity γ1.

[0015] Another aspect of the present invention includes an electronic device comprising one or more processors and a memory; the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the liquid crystal monomer performance prediction method or the mixed crystal performance prediction method.

[0016] Another aspect of the present invention includes a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the liquid crystal monomer performance prediction method or the mixed crystal performance prediction method.

[0017] Another aspect of the present invention includes a computer program product comprising computer-executable instructions which, when executed, are used in the liquid crystal monomer performance prediction method or the mixed crystal performance prediction method.

[0018] Compared with the prior art, the beneficial effects of the present invention are: 1) Improve the stability and generalization ability of predicting key performance of liquid crystal monomers under small sample conditions by embedding physical information and pre-training. 2) Through interpretability analysis, operable structure-performance rules are formed to support the rational design of high birefringence liquid crystal monomers; 3) A latent space weighted average method for mixed crystal characterization is proposed, which transforms the overall characteristics of multi-component formulations into machine-learnable order parameters, and combines this with Kernel SVR to achieve the characterization of Δn, Δε, and C. p Rapid prediction of γ1; 4) Under constraints, reverse proportional search and on-demand generation are achieved, significantly reducing the cost of formula search and improving R&D efficiency. Attached Figure Description

[0019] Figure 1 Overall flowchart for pre-training and property prediction of Physical Augmented Graph Attention Network (GAT); Figure 2 The test set fitting results for the birefringence / dielectric constant / clearing point predictor are shown in the figure. Figure 3 This is a schematic diagram of the liquid crystal monomer structure and synthesis route; Figure 4 Contributes to the Shapley group; Figure 5For the hidden space weighted characterization of mixed crystals; Figure 6 Kernel SVR prediction framework diagram; Figure 7 A comparison graph of predicted and experimental values ​​for the target formulation; Figure 8 It is a deviation statistics chart. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] The detailed test methods for physical and photoelectric properties involved in the following embodiments are as follows: Clear the highlights (C) p ): Polarizing hot stage method: The liquid crystal sample is coated on a glass slide and placed in an orthogonal polarizing microscope hot stage, with a heating rate of 2℃ / min. The point at which the liquid crystal sample begins to turn black from a bright state is observed under a polarizing microscope; this is the clearing point. Alternatively, differential scanning calorimetry can be used: Under a N2 atmosphere, the heating rate is set to 2℃ / min. Birefringence (Δn): Using an Abbe refractometer, under constant temperature conditions of 25℃ and a light source of 589nm, the ordinary light (n) was measured. o ) and unusual light (n e The refractive index of ), birefringence Δn=n e -n o ; Dielectric constant (Δε): The test was conducted using an LCR meter under constant temperature conditions of 25℃. Δε=ε -ε ⊥ That is, the dielectric constant along the long axis of the molecule (ε) ) and the dielectric constant along the short axis of the molecule (ε) ⊥ The difference between ) elastic constant (K) 11 K 33 ): Under constant temperature conditions of 25℃, K was obtained by fitting the capacitance-voltage (CV) curve of the liquid crystal. 11 and K 33 ; Rotational viscosity (γ1): Under constant temperature conditions of 25℃, the transient current value I of the liquid crystal molecules deflecting with the electric field is measured by applying a voltage to the liquid crystal test cell. p The rotational viscosity γ1 was calculated.

[0022] For details on the descriptions and units of the above physical properties, please refer to Table 1.

[0023] Table 1. Relevant physical properties of liquid crystal materials

[0024] In the following embodiments, the structural abbreviations of each liquid crystal material are shown in Table 2.

[0025] Table 2 Abbreviations of Liquid Crystal Material Structures

[0026] In the following embodiments, the structures of each liquid crystal material are shown in Table 3.

[0027] Table 3 Examples of related liquid crystal material structures

[0028] Example 1 A method for predicting and reverse designing the properties of high birefringence liquid crystal materials based on GAT, such as Figure 1 As shown, it includes the following steps: Step 1, establishing and standardizing the single dataset, the specific steps are as follows: Step s11, data acquisition: acquire the structural information of liquid crystal monomer molecules and their corresponding physical property parameters, including but not limited to: n e n o , Δn, ε ε ⊥ , Δε, M p C p (or T) NI ), K 11 K 33 Indicators related to the availability of mixed crystals.

[0029] Step s12, data cleaning and standardization: remove incomplete or duplicate records, and remove samples with missing physical properties or incomparable test conditions; Step s13, Standardization: Standardize each dimension variable. The standardization calculation formula is z = (x - μ) / σ to eliminate dimensional differences and improve training stability. Here, z is the standardized dimensionless variable, x is the value of a certain original variable in a certain sample, μ is the mean of the variable in the training set, and σ is the standard deviation of the variable in the training set.

[0030] Step 2, Physical Information Embedding and Pre-training Target Definition, the specific steps are as follows: Step s21, Physical Feature Extraction: Calculate or obtain the physical parameters of each liquid crystal monomer molecule in the monomer dataset, and add them to the monomer dataset to obtain a physically enhanced dataset. Specifically: The physical parameters of each liquid crystal monomer molecule are extracted using quantum chemical calculations. These physical parameters include, but are not limited to, polarizability (α), dipole moment (µ), rotation tensor (Eg), and HOMO / LUMO energy levels.

[0031] For example, according to the Vuks equation, the birefringence Δn of a liquid crystal is positively correlated with the polarizability Δα of its molecules. Using the calculated Δα as a feature input or training target can force the neural network to "understand" the physical origin of birefringence, rather than simply memorizing the structure.

[0032] Step s22, Physical information embedding: The polarizability (α), dipole moment (µ), and rotation tensor (E) obtained in step s21 are embedded into the physical information. g Physical parameters such as HOMO / LUMO energy levels are embedded as physical information or pre-training supervision signals in the graph attention network (GAT).

[0033] Step 3, the partitioning strategy for the physics augmentation dataset, the specific steps are as follows: The physics augmentation dataset is randomly partitioned for the i-th time according to an 8:2 ratio of training set to test set, where i ∈ F, and F is the preset total number of partitions. To improve the generalization stability of the model, the partitioning of the physics augmentation dataset adopts a structure-aware partitioning strategy based on Morgan's molecular fingerprint (e.g., first stratifying / clustering according to structural similarity, and then sampling within each stratum at an 8:2 ratio).

[0034] Step 4, construct the PIML molecular prediction model: Step s41: Construct a molecular graph G=(V,E) for each liquid crystal monomer molecule: nodes V represent atoms, and edges E represent chemical bonds; construct a graph attention network GAT with embedded physical knowledge based on the molecular graph, and encode the molecular graph into feature vectors using additive pooling or equivalent pooling. The GAT is trained using training set samples. The GAT uses the molecular graph G as nodes and p as edges, and outputs the molecular embedding h as the feature vector of each liquid crystal monomer molecule. .

[0035] Step s42: Connect one or more regression predictors (MLPs) to each molecular embedding h, cascade the GAT and regression predictors (MLPs) to form a PIML molecular prediction model, and achieve multi-task regression output. Each regression predictor (MLP) uses the feature vector... As input, predict the physical properties (Δn, Δε, or C) of the liquid crystal monomer molecules. p ).

[0036] Step s43, using L total Constraints are applied to each PIML molecular prediction model: Introducing physical consistency constraintsL phys To ensure that the prediction results of the regression predictor MLP satisfy known physical correlations or scaling rules, for example, by using physical enhancement features such as Δα as bridging, a consistency regularization is added to the prediction of Δn: L total = L data +λ· L phys +μ· L eg ,in L total Let be the total loss function during the training phase. L data For monitoring losses (such as MSE / Huber). L eg λ represents the weight regularization or sparse constraint, and μ represents the weight coefficients, respectively.

[0037] Interpretability analysis outputs structural contribution patterns: Interpretability analysis is performed on the regression predictor MLP to obtain the contribution patterns of functional groups / structural motifs to physical properties. Shapley analysis is performed on the MLP to obtain the contribution patterns of structural motifs: non-aromatic rings, long aliphatic carbon chains, and CF3 contribute negatively to refractive index and dielectric constant; alkynyl groups, isothiocyanates, benzene rings, halogen atoms, and nitrogen-carbon triple bonds contribute positively; furthermore, halogen atoms, nitrogen-carbon triple bonds, long aliphatic carbon chains, and vinyl groups contribute negatively to the clearing point, while non-aromatic rings, isothiocyanates, ester groups, aromaticity, and alkynyl groups contribute positively to the clearing point. Figure 4 The aforementioned criteria are used for subsequent candidate pool screening and can be examined.

[0038] Step s44, verify the accuracy of the PIML performance prediction model (at least covering Δn): the PIML performance prediction model achieves a correlation coefficient of 0.9989 for the birefringence test set, with a mean absolute error of 0.02; it also has acceptable prediction accuracy for dielectric parameters and cleaning points, such as... Figure 2 The accuracy verification demonstrates the reliability of the PIML performance prediction model, and then proceeds to step 5.

[0039] Step 5: Input the molecular diagram of the liquid crystal monomer molecule to be verified into the PIML performance prediction model to obtain the physical property parameters of the liquid crystal monomer molecule to be verified (cleaning point C). p Predicted values ​​of birefringence Δn and dielectric constant Δε.

[0040] Example 2 This embodiment, based on Embodiment 1, further predicts the performance of mixed crystals.

[0041] Based on steps 1 to 5 of the embodiment, step 6 is then performed: Step 6, monomer screening, general formula space constraint, and synthetic route, specifically includes the following steps: From the physical property parameters of each liquid crystal monomer molecule obtained in step 5, monomer structures that meet the target requirements are selected as candidate monomers, and the candidate monomers are further screened to obtain mixed crystal formulations.

[0042] Specifically, based on the principles of interpretability analysis, target performance constraints are constructed, and candidate monomers are screened in a target-oriented manner to obtain a candidate monomer library for mixed crystal formulations.

[0043] Step 7, in this embodiment, the high birefringence liquid crystal monomer is limited to a compound that conforms to at least one of the following general chemical formulas (I)-(III): Chemical formula (Ⅰ): Diphenylacetylene-based liquid crystal monomer

[0044] Where X1 and X2 are independently H, F, or Cl; R1 is C. n H （2n+1） Alkyl chain, n=1-12, preferably 3-10; R2 is -CN, -NCS, -OCH3, -OCF3, etc.

[0045] This general formula is used to balance high Δn with processability and to construct the structural space of monomers with low melting point / high solubility regulation.

[0046] Chemical formula (II): Triphenylacetylene-based liquid crystal monomer

[0047] Where X1 and X2 are independently H, F, or Cl; R1 is C. n H 2n+1 Alkyl chain, n=1-12, preferably 3-10; R2 is -CN, -NCS, -OCH3, -OCF3, etc.

[0048] This general formula is used to improve rigidity and conjugate length, and to construct the main structural space of a "high birefringence / high-definition bright spot contribution type" single entity.

[0049] Chemical formula (Ⅲ): terphenyl liquid crystal monomer

[0050] Where X1 and X2 are independently H, F, or Cl; R1 is C. n H 2n+1 Alkyl chain, n=1-12, preferably 3-10; R2 is -CN, -NCS, -OCH3, -OCF3, etc.

[0051] This general formula is used to construct the structural space of low rotational viscosity-controlled monomers and to form a formulation synergy with high birefringence components.

[0052] Step s71, Selecting monomers and grouping into three functional categories: Using high birefringence liquid crystal B as the matrix, low melting point liquid crystal A and low rotational viscosity liquid crystal C are selected as the mixing components. Finally, based on experiments, a total of eight monomers in three categories are determined: 2 types of type B, 3 types of type A, and 3 types of type C.

[0053] Step s72, taking isothiocyanate-based liquid crystal monomers as an example, the liquid crystal monomers are synthesized as follows: General synthetic route for component A (preferred embodiment applicable to chemical formula (I))

[0054] (1) Acetylene bridging skeleton construction: Under an inert atmosphere, halogenated aromatic compounds and terminal alkyne aromatic compounds are added to an organic solvent and an organic base system, and a palladium catalyst and a copper co-catalyst are added to carry out a coupling reaction to obtain acetylene bridging intermediate M1.

[0055] The halogenated group is preferably Br or I; the solvent is preferably THF, DMF or a mixture thereof; the organic base is preferably triethylamine or diisopropylamine; the catalytic system can be Pd(PPh3)4 / CuI or an equivalent system.

[0056] (2) Functional group transformation yields aromatic amine intermediates: The precursor functional group (e.g., nitro) of intermediate M1 is reduced to obtain amino-containing aromatic amine intermediate M2.

[0057] (3) Amino-terminal isothiocyanate esterification: The aromatic amine intermediate M2 is reacted with phosgene (CSCl2) or an equivalent isothiocyanate esterifying agent in the presence of a basic acid scavenger to obtain the target isothiocyanate monomer of chemical formula (Ⅰ).

[0058] The acid scavenger can be a carbonate or an organic base; after the reaction, it can be purified by extraction, washing, drying and column chromatography / recrystallization.

[0059] The route and Figure 3 The middle component A corresponds to this.

[0060] General synthetic route for component B (preferred embodiment applicable to chemical formula (II))

[0061] (1) Introduction of acetylene unit and construction of coupling sites: Sonogashira coupling was performed under palladium / copper co-catalysis to obtain acetylene-bridged intermediate P1 with halogenated or borateable sites.

[0062] (2) Boration yields an arylboronic acid (ester) intermediate (for subsequent Suzuki assembly of the third aromatic ring): Lithylating and borylation of P1 yields P2, an arylboronic acid / boronic acid ester intermediate with an acetylene skeleton.

[0063] (3) Suzuki coupling to construct a triarylaceyne skeleton and introduce an amino precursor: P2 was coupled with a halogenated aromatic compound containing an amino group or a compound that can be converted to an amino group via Suzuki coupling to obtain a triarylaceyne skeleton intermediate P3; when the target is chemical formula (III), a fluorinated aromatic ring substrate was selected to form a polyfluorinated substitution mode.

[0064] (4) Amino-terminal isothiocyanate esterification: P3 was esterified with isothiocyanate to obtain the target monomer of chemical formula (II), and then purified to obtain the final product.

[0065] The route and Figure 3 The middle component B corresponds to this.

[0066] General synthetic route for component C (preferred embodiment applicable to chemical formula (III))

[0067] (1) Suzuki coupling to construct a terphenyl framework: Arylboronic acid (or borate ester) is coupled with a halogenated aromatic compound via Suzuki coupling in the presence of a palladium catalyst and a base to obtain biphenyl / terphenyl skeleton intermediate T1.

[0068] (2) Optional borylation intermediates are constructed for ring expansion: When it is necessary to further introduce a third aromatic ring, the intermediate can be lithiated-boronized to generate boric acid intermediate T2, and then subjected to secondary Suzuki coupling with a halogenated aromatic compound to obtain intermediate T3 containing an amino group or capable of being converted to an amino group.

[0069] (3) Isothiocyanate esterification yields -NCS end groups: The amino-containing intermediate T3 was converted into the target chemical formula (III) monomer under isothiocyanate esterification conditions, and the final product was obtained by purification.

[0070] The route and Figure 3 The middle component C corresponds to this.

[0071] Note: The above component AC route corresponds one-to-one with general formulas (Ⅰ)-(Ⅲ), which enables the candidate structures screened by machine learning to have a clear chemical synthesis path, thereby meeting the complete disclosure requirements of "algorithm-structure-synthesis-verification".

[0072] Step s73, Orthogonal Experiments and Mixed Crystal Datasets: Construction of Mixed Crystal Experimental Data Figure 3The three types of high birefringence monomers shown are used as a component library. Through a series of experiments, the component ratios are changed to efficiently explore the multidimensional formulation space.

[0073] Step s74, Mixed crystal property testing indicators: Measurements should include at least Δn, Δε, and C. p The mixed crystal properties of rotational viscosity γ1 are used to form a mixed crystal dataset.

[0074] Step 8, Characterization of mixed crystal formulation and mixed crystal prediction model based on Kernel SVR prediction, specifically includes the following steps: Step s81, weighted average representation of the mixed crystal formulation feature vector: The molecular diagrams of the three types of liquid crystal molecules A, B, and C are input into the trained and validated GAT model, and their respective feature vectors are obtained through the readout layer. According to the formula composition (mass fraction) Constructing the overall feature vector of the mixed crystal:

[0075] in, This represents the overall eigenvector of the mixed crystal. For the first i The mass fraction of each liquid crystal monomer, i.e., the mass fraction of each component. The eigenvectors of liquid crystal monomers; the overall eigenvectors of mixed crystals. Machine learning "order parameter" for mixed crystal formulation.

[0076] Construction of mixed crystal formulations, mixed crystal properties and The resulting mixed dataset is further divided into a training set and a validation set.

[0077] Step s82, construct a mixed crystal prediction model based on Kernel SVR prediction: train the mixed crystal prediction model based on the training set obtained in step 7, and... and The inputs to the mixed crystal prediction model predict the mixed crystal Δn, Δε, and C respectively. p With γ1.

[0078] SVR uses an RBF kernel; the penalty parameter and kernel parameter are searched, and the preferred value range is [2]. -10 ,2 10 ].

[0079] Step s82, error constraint The birefringence Δn, dielectric constant Δε, and clearing point C obtained by the mixed crystal prediction model were evaluated using the validation set. p The average deviation between the predicted and experimental values ​​(obtained experimentally) of rotational viscosity γ1 is controlled within 5%, which meets the engineering screening accuracy requirements.

[0080] Step 9: Forward Mixed Crystal Performance Prediction or Reverse Mixed Crystal Formulation Generation Forward mixed crystal performance prediction: Utilizing the properties of each monomer molecule in the mixed crystal to be predicted and The inputs are fed into the mixed crystal prediction model to predict Δn, Δε, and C of the mixed crystal, respectively. p With γ1.

[0081] Reverse mixed crystal formulation generation: This involves combining the monomer molecules in the desired mixed crystal... And the desired mixed crystal Δn, Δε, C p The γ1 is input into the mixed crystal prediction model to obtain the values ​​of each monomer molecule. The specific steps are as follows: Step s91, Low-dimensional variables and constraints: Since the mixed crystal is composed of three types of monomers and the mass fractions satisfy the constraints, the formulation ratio can be fully expressed by two independent degrees of freedom; two reduction parameters are introduced to reduce the high-dimensional ratio problem to low-dimensional searchable variables, and in each search, the following conditions are met: mass fraction of each component ≥0 and the sum is 1.

[0082] Step s92, Objective function: For any given ratio, use the mixed crystal prediction model from step 8 to predict the birefringence Δn, dielectric constant Δε, and clearing point C. p And rotational viscosity γ1; based on the target performance setpoint, the MAE between the predicted value and the target value is calculated as the objective function, and the performance-to-formulation inverse derivation is realized in the latent space through a constrained optimization algorithm to screen the comprehensive optimal formulation, such as Figures 5-6 As stated above.

[0083] Step s93, Target Parameter "Change Variable" and Experimental Design: This method uses the "target performance parameter set" as an explicit input variable: by changing (Δn, Δε, C) p The weights of γ1 and γ2 in the objective function can be used to form different experimental tasks, such as prioritizing high Δn, prioritizing low γ1, and prioritizing wide temperature range; and the same "prediction-search-preparation-test" closed-loop experiment is performed on each set of objective combinations.

[0084] Examples 3, 4, and 5 below, by setting multiple sets of target performance combinations and experimentally measuring the actual performance, show that the average deviation between the prediction and the experiment is less than 5%, as shown in the following figures. Figures 7-8 As stated above.

[0085] Example 3 Component A represents a low-melting-point / high-solubility liquid crystal monomer; component B represents a high-birefringence / high-brightness liquid crystal monomer; component C represents a low-rotational-viscosity liquid crystal monomer. Target values: Δn = 0.4300, Δε ≈ 29.0, C p=150.0, γ1=380.0, with fixed SVR model parameters, and the formula ratio w i Using these variables, gradient descent or a genetic algorithm is used to find the optimal solution. System output: Recommended formulation is monomer A (15%) + monomer B (62%) + monomer C (23%).

[0086] The target parameters of the liquid crystal composition, including its component structure, mass percentage, and performance data, are shown in Table 4 below.

[0087] Table 4

[0088] For example, the experimental and predicted values ​​are compared as follows: Δn experimental 0.4335 / predicted 0.4302, Δε experimental 28.8 / predicted 29.0, C p Experimental result 164.5 / predicted 152.9, γ1 experimental result 404.4 / predicted 375.5, overall average deviation less than 5%.

[0089] This result demonstrates that, under target-driven inverse search, the generated formulation can maintain high birefringence while taking into account dielectric and temperature constraints, and the prediction-experiment agreement meets engineering-usable accuracy.

[0090] Example 4 Target Combination 2 (High Δn Priority): Increase Δn and increase the weight α_Δn of the Δn term, while limiting γ1 to no higher than a preset lower limit, C p Not lower than the preset lower limit; optimize the output according to constraints to predict the optimal formula; prepare and test Δn, Δε, and C. p γ1, record the error and write it back to the database.

[0091] The component structure, mass percentage, and performance data of the high Δn-preferred liquid crystal composition of this embodiment are shown in Table 5 below.

[0092] Table 5

[0093] Example 5 Target combination 3 (low γ1 priority): reduce γ1 and increase the weight α_γ1 of the γ1 term, while setting a minimum threshold for Δn; repeat the above closed loop.

[0094] The only changes in the above "change variables" are the target parameter set and weight vector; the rest of the data representation, predictor, and optimizer remain unchanged.

[0095] The component structure, mass percentage, and performance data of the low γ1-preferred liquid crystal composition of this embodiment are shown in Table 6 below.

[0096] Table 6

[0097] Example 6 A liquid crystal material performance prediction and reverse mixed crystal design system for implementing the method of the present invention includes a processor, a memory, and program instructions stored in the memory and executable by the processor. When the program instructions are executed, they are used to implement a liquid crystal monomer performance prediction method based on GAT or a mixed crystal performance prediction method based on the liquid crystal monomer performance prediction method.

[0098] Preferably, the functional modules of the system include at least: a data processing module for performing data cleaning, dimensional unification, and standardization; a molecular graph construction module for converting monomer structures into molecular graphs and generating node / edge features; a physical information acquisition module for reading or calculating physical parameters such as polarizability and dipole moment and completing embedding; a graph neural network training and inference module for training a GAT encoder and outputting monomer latent space vectors; and a mixed crystal characterization and SVR prediction module for constructing h mix And predict Δn, Δε, C p (or T) NI The module includes γ1 and γ2; an inverse optimization module for performing constrained optimization search with MAE as the objective function under mass fraction constraints and outputting candidate formulations; and a result output module for outputting monomer screening results, formulation ratios and corresponding predicted values.

[0099] Example 7 A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a method for predicting the performance of liquid crystal monomers based on GAT or a method for predicting the performance of mixed crystals based on the liquid crystal monomer performance prediction method.

[0100] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the performance of liquid crystal monomers based on GAT, characterized in that, Includes the following steps: Step 1: Collect liquid crystal monomer molecules and their corresponding physical property parameters, perform preprocessing, and establish a monomer dataset; Step 2: Obtain the physical parameters of each liquid crystal monomer molecule in Step 1, and add the physical parameters as feature dimensions to the monomer dataset to establish a physical enhancement dataset. Step 3: Divide the physics augmentation dataset from Step 2 into a training set and a test set; Step 4: Convert the liquid crystal molecules into a molecular graph. Based on the molecular graph, construct a graph attention network (GAT) with embedded physical knowledge. Cascade the GAT and the regression predictor (MLP) to form a PIML molecular prediction model. The GAT uses the molecular graph as nodes and outputs molecular embeddings as feature vectors. Each molecule embedding is connected to one or more regression predictors (MLPs), and each regression predictor (MLP) is connected to the feature vector. As input, predict the physical properties of liquid crystal monomer molecules; Using the training set samples from step 3 combined with the loss function L total The PIML molecular prediction model was trained and its hyperparameters were optimized. The accuracy of the PIML molecular prediction model was verified using the validation set samples from step 3. Step 5: Input the molecular diagram of the liquid crystal monomer molecule to be verified into the PIML performance prediction model trained and verified in Step 4 to obtain the predicted values ​​of its physical property parameters.

2. The method for predicting the performance of liquid crystal monomers based on GAT as described in claim 1, characterized in that, In step 1, the experimental physical property parameters include unusual light n e Ordinary light o Birefringence Δn, dielectric constant ε along the long axis of the molecule Dielectric constant ε along the short axis of the molecule ⊥ Dielectric constant Δε, Clearing point C p elastic constant K 11 and K 33 ; The preprocessing steps in step 1 include data cleaning and standardization.

3. The method for predicting the performance of liquid crystal monomers based on GAT as described in claim 1, characterized in that, The physical parameters in step 2 include polarizability α, dipole moment µ, rotation tensor Eg, or HOMO / LUMO energy levels.

4. The method for predicting the performance of liquid crystal monomers based on GAT as described in claim 1, characterized in that, In step 3, a structure-aware partitioning strategy based on Morgan molecular fingerprints is used to partition the physical augmentation dataset.

5. The method for predicting the performance of liquid crystal monomers based on GAT as described in claim 1, characterized in that, In step 4 L total = L data +λ· L phys +μ· L eg ,in L total Let be the total loss function during the training phase. L data To monitor losses, L eg λ represents the weight regularization or sparse constraint, and μ represents the weight coefficients, respectively. In step 4, the physical property parameters include the cleaning point C. p Birefringence Δn and dielectric constant Δε.

6. A method for predicting the performance of mixed crystals based on the liquid crystal monomer performance prediction method, characterized in that, Includes the following steps: Step 6: Based on steps 1 to 5 of the liquid crystal monomer performance prediction method as described in any one of claims 1-5, a series of physical property parameters of liquid crystal monomer molecules are predicted, and N liquid crystal monomer molecules that meet the physical property parameter requirements are selected as candidate monomers for mixed crystals. Step 7: Construct a mixed crystal formulation through orthogonal experiments and measure the physical properties of the mixed crystal; Step 8: Combine the monomers from each mixed crystal formulation in Step 7 and input them into the trained and validated GAT to obtain the feature vectors of the monomers. By the mass fraction of each monomer Further calculation of the overall eigenvector of the mixed crystal Based on the properties of mixed crystals and the properties of each monomer Construct a mixed crystal dataset, then divide it into training and validation sets; build a mixed crystal prediction model based on KernelSVR prediction, with the input of the mixed crystal prediction model being... and The output is the mixed crystal properties. The mixed crystal prediction model is trained using the training set, and the error is evaluated using the validation set and experiments. Step 9, Predicting the performance of forward mixed crystals or generating the reverse mixed crystal formula: The feature vector of the candidate single entity obtained in step 6 and Input the trained and validated mixed crystal prediction model to obtain the physical properties of mixed crystals composed of candidate monomers; Alternatively, the desired mixed crystal properties and eigenvectors of candidate monomers can be used. Input the trained and validated mixed crystal prediction model to obtain the values ​​of each candidate monomer. .

7. The method for predicting the performance of mixed crystals as described in claim 6, characterized in that, In step 7 ; In step 9, the mixed crystal properties include birefringence Δn, dielectric constant Δε, and clearing point C. p Or rotational viscosity γ1.

8. An electronic device, characterized in that, The electronic device includes one or more processors and a memory; the memory is used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the liquid crystal monomer performance prediction method as described in claim 1 or the mixed crystal performance prediction method as described in claim 6.

9. A computer-readable storage medium, characterized in that, It stores computer-executable instructions, which, when executed, are used to implement the liquid crystal monomer performance prediction method as described in claim 1 or the mixed crystal performance prediction method as described in claim 6.

10. A computer program product, characterized in that, The aforementioned computer program product includes computer-executable instructions that, when executed, implement the liquid crystal monomer performance prediction method as described in claim 1 or the mixed crystal performance prediction method as described in claim 6.

Images