Intelligent generation method of low-temperature-resistant ionic liquid for lubrication demand

Abstract

The invention discloses an intelligent generation method of low-temperature-resistant ionic liquid for lubrication requirements. The method comprises the following steps: constructing a low-temperature-resistant ionic liquid data set, screening samples, and supplementing missing data; carrying out structure coding by adopting an SMILES format and extracting a molecular descriptor; dividing a training set, a verification set and a test set based on application scene stratified sampling; constructing a graph attention network model, extracting molecular structure characteristics through a multi-head attention mechanism, and respectively predicting tribological performance and freezing points by adopting double-branch output; training the model by using an AdamW optimizer and a transfer learning strategy; a candidate ionic liquid formula is generated through combination of a simulated annealing algorithm and molecular dynamics simulation, and the optimal ionic liquid formula is output through multi-objective optimization. According to the method, the technical problems of efficient screening and accurate performance prediction of the ionic liquid lubricant in the extremely low-temperature environment are solved.

Description

Technical Field

[0001] The present invention relates to the technical field of the intersection of lubricant material design and artificial intelligence, and specifically to a method for intelligently generating low-temperature-resistant ionic liquid lubricants for lubrication requirements under extreme working conditions. It belongs to the technical field of new functional material design, deep learning modeling and intelligent formula screening. Background Art

[0002] With the rapid development of high-end equipment manufacturing, aerospace, polar exploration, and other technologies, critical equipment often faces complex operating environments such as extreme low temperatures, placing higher demands on lubricant performance in these environments. Traditional solid lubricants or molecular liquid lubricants often exhibit problems such as solidification and poor fluidity at low temperatures, leading to lubrication failure and seriously affecting the stable operation and service life of the equipment.

[0003] Ionic liquids, room-temperature molten salts composed of anions and cations, are ideal candidates for extreme-condition lubricants due to their excellent lubrication properties, high stability, wide temperature range, and unique designability. Especially in low-temperature lubrication scenarios, ionic liquids exhibit outstanding properties such as low freezing point, controllable viscosity, and low wear rate, demonstrating significant performance advantages and broad application potential.

[0004] However, despite the significant potential of ionic liquids, rapid screening for ideal lubricants with a low freezing point, suitable viscosity, and excellent lubrication properties remains a major challenge in current research and development. Specifically, current research and development faces the following bottlenecks: 1) Reliance on iterative trial-and-error methods, resulting in low R&D efficiency and high costs: Currently, the development of ionic liquid lubricants for extreme low temperatures primarily relies on the traditional "Edison-style" trial-and-error approach. Researchers must invest significant time and resources to individually formulate ionic liquids of varying compositions and conduct tedious lubrication performance tests under various extreme low-temperature conditions. This process is not only time-consuming and resource-intensive, but also faces extremely limited experimental coverage with tens of thousands of potential combinations, making comprehensive and efficient screening difficult. 2) The complex structure-property relationship of ionic liquids limits the effectiveness of traditional prediction methods: Due to the highly tunable molecular structure of ionic liquids, their performance is significantly influenced by their microstructure and exhibits a high degree of nonlinearity. Current performance prediction methods based on rule-based derivation and molecular simulations struggle to fully and accurately describe their lubrication behavior at extreme low temperatures, lacking an efficient and systematic performance prediction and screening mechanism. 3) Lack of a systematic intelligent design and screening platform: Existing research mostly focuses on local performance prediction or experimental verification, and lacks an intelligent generation path that integrates molecular structure encoding, performance prediction and combinatorial optimization. It is difficult to support the targeted matching and precise recommendation of lubrication requirements under complex working conditions such as extremely low temperatures.

[0005] Therefore, an intelligent design method integrating molecular structure expression, deep learning modeling, performance prediction, and formulation optimization is urgently needed to rapidly screen and predict ionic liquid lubricant formulations with excellent performance under extreme low-temperature conditions. This method can realize the intelligent design and optimization of ionic liquid lubricants from the perspective of "data-structure-performance". This method can significantly improve the R&D efficiency and reduce R&D costs of ionic liquid lubricants, and promote their engineering applications in key fields such as aerospace, military industry, and extreme manufacturing. Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent generation method for low-temperature resistant ionic liquids based on deep learning and intelligent optimization algorithms to address the problems in the existing technology of low-temperature resistant ionic liquid lubricant development that rely on traditional experimental trial and error methods, resulting in low research and development efficiency and high costs, as well as inaccurate structure-performance relationship modeling and insufficient multi-objective optimization capabilities in performance prediction methods based on rule deduction or simple simulation. The method realizes the intelligence of the entire process from molecular structure and performance prediction to formula optimization, and efficiently screens ionic liquid formulas with low freezing point, suitable kinematic viscosity and excellent friction performance, significantly improving the research and development efficiency of lubricants in extreme low-temperature environments, and promoting the application of ionic liquid lubricants under extreme working conditions in key fields such as aerospace, artillery, etc.

[0007] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] 1) Data preprocessing and enhanced coding: SQL queries were used to select ionic liquids with a freezing point below -30°C, a kinematic viscosity of 100–500 cP at -40°C, a four-ball test friction coefficient ≤ 0.15, and a wear volume ≤ 0.8 mm from the ionic liquid database. 3 The initial data set was established by using samples of the original data. Entries with missing lubrication performance data were supplemented through literature search or experimental measurement. The structure was encoded using the SMILES format, and descriptors such as the number of hydrogen bond donors / acceptors, van der Waals surface area, and molecular polar surface area were extracted using molecular modeling tools and unified into standardized structured data input.

[0009] The data sources used include the ILThermo database, ASTM standard experimental documents and SCI journal documents, and the source type and confidence level of each data are marked to enhance data traceability and weight distinction

[0010] 2) Dataset division: Stratified sampling is performed based on application scenario labels in a ratio of 7:1.5:1.5, and the training set, validation set, and test set are divided to ensure a balanced distribution of samples for various working conditions.

[0011] 3) Graph Neural Network Model Construction: A graph attention network (GAT) was constructed based on the PyTorch Geometric framework, with 8 attention heads configured in each layer. The molecular graph structure converted from SMILES was used as the input node. The node features included the atom type and its corresponding molecular descriptor. The model adopted a dual-branch output structure: the first branch predicted the friction coefficient, wear rate, and viscosity, and the second branch predicted the freezing point, which were used for the prediction of tribological properties and freezing point, respectively.

[0012] 4) Model Construction and Training: Based on the dataset partitioned in step 3), a graph attention neural network (GAT) was constructed using the PyTorch Geometric framework. Each layer of the model was configured with eight attention heads. The input nodes contained atomic types and molecular descriptors, and a dual-branch output structure was used to predict the friction coefficient, wear rate, viscosity, and freezing point. During training, the AdamW optimizer and cosine annealing strategy were used to dynamically adjust the learning rate. A multi-task weighted loss function was introduced to dynamically adjust the weights of each task. The network was initialized by loading pre-trained parameters from the materials domain.

[0013] A multi-task weighted loss function is introduced in the model training process, and the corresponding loss weights are dynamically adjusted according to the performance in the validation set. The specific form of the loss function is as follows:

[0014] L total =∑(w k ×L k );

[0015] Where k = 1 to 4, L k represents the loss value of the kth prediction task (including friction coefficient, wear rate, viscosity, and freezing point), w k Indicates the weight coefficient of the task.

[0016] Weight coefficient w k Dynamic calculation according to the following formula:

[0017] w k =exp(η×Acc k ) / ∑[exp(η×Acc j )];

[0018] Where j = 1 to 4, Acc k represents the normalized accuracy evaluation value of the k-th task on the validation set, η is the smoothing factor, and its value range is 0.5 to 1.0.

[0019] This strategy dynamically adjusts the attention weights during model training based on the performance of each task in the validation set, thereby improving the overall model's multi-task collaborative learning and generalization capabilities. Pre-trained models in the materials field are also loaded to enhance the model's initial expressiveness. The model's generalization capabilities in tribological performance and freezing point prediction tasks are evaluated on the test set, providing accurate prediction support for subsequent ionic liquid formulation screening.

[0020] 5) Formulation generation and screening: After the user inputs the target temperature and performance parameters, the model inference stage is entered after standardization. The simulated annealing algorithm is combined with the GAT prediction results to search for candidate ionic liquid formula combinations in the database, limiting the cations to imidazole, pyridine or quaternary phosphonium, and excluding those containing Cl-, Br - Ionic liquid formulations with halogen anions.

[0021] 6) Stability Verification and Optimization: Molecular dynamics simulations were performed on the selected candidate ionic liquid formulas to evaluate their structural stability at -40°C. 10 ns simulations were performed using the AMBER force field. Multi-objective optimization scoring was performed using the NSGA-II algorithm. The four optimization objectives of minimizing the friction coefficient, minimizing the wear rate, approaching the target viscosity, and minimizing the freezing point were comprehensively considered. Candidate ionic liquid formulas with a score higher than 8 out of 10 were selected.

[0022] 7) Result output and visualization: Output the preferred ionic liquid formula with the highest score after multi-objective optimization, and display its SMILES encoding and performance prediction value through a visual interface, supporting CSV / JSON export format and API interface call.

[0023] Compared with the prior art, the present invention has the following significant advantages:

[0024] (1) Achieve efficient screening of low-temperature resistant lubricants: Build a graph neural network model to replace traditional trial-and-error experiments, improving the efficiency of formula generation by more than an order of magnitude;

[0025] (2) Strong structure-performance modeling capability: Based on GAT and multi-task prediction network, it can model the complex nonlinear relationship between molecular structure and low-temperature friction performance, with higher prediction accuracy than traditional linear modeling or regression models;

[0026] (3) Significantly improved intelligent optimization capabilities: The introduction of simulated annealing and NSGA-II optimization algorithms, combined with user-entered customized goals, enables intelligent recipe matching for extreme environments;

[0027] (4) The results are highly interpretable: the SMILES structure and predicted performance corresponding to the formula are output, and the formula screening path and performance comparison are displayed in a visual interface to enhance the feasibility of engineering implementation;

[0028] (5) Good scalability: The method framework is universal and can be adapted to more complex material screening tasks by expanding descriptors, improving model structure, or integrating other lubrication performance indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is an overall flow chart of the method for intelligently generating a low-temperature resistant ionic liquid lubricant for lubrication requirements under extreme working conditions.

[0030] Figure 2 It is a schematic diagram of the intelligent design and optimization visualization interface of "data-structure-performance" built based on intelligent generation methods. DETAILED DESCRIPTION

[0031] In order to better understand the technical content of the present invention, specific embodiments are given and described below with reference to the accompanying drawings.

[0032] Various aspects of the present invention are described herein with reference to the accompanying drawings, in which a number of illustrative embodiments are shown. The embodiments of the present invention are not necessarily intended to encompass all aspects of the present invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, may be implemented in any of a number of ways, as the concepts and embodiments disclosed herein are not limited to any particular implementation. In addition, some aspects disclosed herein may be used alone or in any appropriate combination with other aspects disclosed herein.

[0033] The present invention provides a method for intelligently generating a low-temperature resistant ionic liquid lubricant based on deep learning, comprising the following steps:

[0034] S1: Data Preparation and Enhancement

[0035] By querying the ionic liquid database (such as v3.2, ILThermo) screened samples that met the following criteria:

[0036] -Freezing point <-30℃

[0037] --Kinematic viscosity at 40°C: 100-500 cP

[0038] -Friction coefficient of four-ball test ≤ 0.15

[0039] - Wear volume ≤ 0.8mm 3

[0040] For missing data entries, we supplemented the data with SCI journal literature data and experimental measurement data according to ASTM D2270. We used the SMILES format for molecular structure encoding and the RDKit 2022.09 toolkit to extract molecular descriptors, including:

[0041] Number of hydrogen bond donors (HBD)

[0042] Number of hydrogen bond acceptors (HBA)

[0043] Van der Waals surface area (SMR_VSA)

[0044] Molecularly polar surface (TPSA)

[0045] S2: Building a graph neural network model

[0046] Build a Graph Attention Network (GAT) model based on the PyTorch Geometric 1.12 framework: - Each layer is configured with 8 attention heads to aggregate node information - Input layer: atom type + molecular descriptor (128-dimensional features)

[0047] -Dual-branch output architecture:

[0048] Branch 1: Predict friction coefficient, wear rate, and viscosity (3 output nodes)

[0049] Branch 2: Predict freezing point (1 output node)

[0050] -Implementing a dynamic weighted multi-task learning strategy

[0051] S3: Model training and optimization

[0052] AdamW optimizer is used for training, key parameters are:

[0053] - Initial learning rate: 0.001

[0054] -Batch size: 32

[0055] - Training epochs: 100

[0056] The learning rate is adjusted using a cosine annealing strategy:

[0057] η t =η min +1 / 2*(η max -η min )*(1+cos(π*T cur / T max ))

[0058] Load the Materials Project pre-trained model weights to initialize the GAT layer parameters.

[0059] S4: Intelligent recipe generation and optimization

[0060] Receive user input of extreme temperature and target performance parameters:

[0061] {

[0062] "environment_temp":-50,

[0063] "target_viscosity":300,

[0064] "max_friction":0.10,

[0065] "min_melting_point":-60

[0066] }

[0067] Use simulated annealing algorithm to search for candidate combinations:

[0068] - Cation-limited: imidazole ([BMIM]+) or pyridine ([BPy]+)

[0069] -Anion exclusion: containing Cl-, Br - Other corrosive ingredients

[0070] Molecular dynamics simulations were performed using the AMBER force field:

[0071] 1. Energy minimization (5,000 steps)

[0072] 2. NVT ensemble equilibrium (-40°C, 100ps)

[0073] 3. NPT ensemble production (10 ns, time step 2 fs)

[0074] Multi-objective optimization using the NSGA-II algorithm:

[0075] -Optimization goals: reduce friction coefficient, reduce wear rate, bring viscosity close to target value, and reduce freezing point -Target weights: friction coefficient 0.4, wear rate 0.3, viscosity 0.2, freezing point 0.1

[0076] S5: Result output and verification

[0077] The optimal formula and predicted performance are displayed through a visual interface:

[0078] -3D rendering of molecular structure (implemented by RDKit)

[0079] -Performance comparison radar chart (generated by Plotly 5.15)

[0080] Supports data export in CSV / JSON format and provides RESTful API interface.

[0081] Example 1: Development of Ultra-Low-Temperature Lubricants for Aircraft Engines

[0082] This embodiment applies the method of the present invention to develop an aviation engine lubricant suitable for a -50°C environment.

[0083] Step S1:

[0084] Database source: v3.2 (12,340 records)

[0085] SQL filter code:

[0086] SELECT * FROM ionic_liquids

[0087] WHERE melting_point<-30

[0088] AND viscosity_40c BETWEEN 100AND 500

[0089] AND friction_coef <= 0.15;

[0090] ·632 valid records were obtained

[0091] Supplement 15 sets of ASTM D4172-22 standard experimental data

[0092] Molecular descriptor extraction: use RDKit to calculate HBD, HBA and other parameters

[0093] Step S2:

[0094] GAT model architecture:

[0095] class DualGAT(torch.nn.Module):

[0096] def__init__(self):

[0097] self.gat1=GATConv(128,64,heads=8)

[0098] self.branch1=Linear(64*8,3)

[0099] self.branch2=Linear(64*8,1)

[0100] Step S3:

[0101] Dynamic weight parameter: η = 0.85

[0102] Validation set performance:

[0103] Freezing point prediction mean absolute error (MAE): 3.2℃

[0104] Friction coefficient prediction mean absolute error (MAE): 0.008

[0105] Step S4:

[0106] Simulated annealing parameters: initial temperature 1000, cooling rate 0.95, iteration 5000 times

[0107] Molecular dynamics simulation: -50°C environment, 10ns trajectory

[0108] Step S5:

[0109] Optimal formula: [BMIM][TFSI] (1-butyl-3-methylimidazolium bis(trifluoromethanesulfonyl imide)

[0110] Comparison between prediction and actual measurement:

[0111] Freezing point: predicted value -62℃, measured value -61.2℃, error 1.3%

[0112] Friction coefficient: predicted value 0.085, measured value 0.089, error 4.7%

[0113] The visual interface outputs three-dimensional molecular structure and performance curves.

[0114] Example 2: Lubricant for Polar Wind Power Equipment (-70°C Environment)

[0115] Special parameter adjustments:

[0116] Freezing point threshold: <-65℃

[0117] Viscosity range: 150-400 cP (-70°C)

[0118] Molecular dynamics temperature: -70℃

[0119] Optimization results:

[0120] Optimal formula: [P 6，6，6，14 [BMB] (Trihexyltetradecylphosphonium bis(2,4,4-trimethylpentyl)phosphinate) Performance Verification:

[0121] Freezing point: predicted value -75℃, measured value -73.5℃

[0122] Viscosity at -70°C: predicted value 325cP, measured value 338cP

[0123] The results of this example show that the method of the present invention can efficiently generate ionic liquid lubricant formulations that meet the requirements of extreme low temperature environments, with a prediction accuracy error of <5% and a development cycle shortened by more than 85%.

Claims

1. A method for intelligently producing low-temperature resistant ionic liquids for lubrication requirements under extreme working conditions, characterized in that: The following steps are involved: 1) Data preprocessing: SQL queries were used to filter samples from the ionic liquid database to establish an initial data set. The samples met the following requirements: freezing point below -30°C, kinematic viscosity at -40°C between 100 and 500 cP, four-ball test friction coefficient ≤ 0.15, and wear volume ≤ 0.8 mm. 3 ; 2) Data enhancement: missing entries were supplemented in the initial dataset, and the ionic liquid structures in the initial dataset were encoded using the SMILES format. Molecular descriptors including the number of hydrogen bond donors, the number of hydrogen bond acceptors, the van der Waals surface area, and the molecular polarity were extracted. 3) Dataset division: Stratified sampling is performed based on application scenario labels, and the initial dataset after structural encoding is divided into training set, validation set, and test set; 4) Model Construction: Based on the dataset partitioned in step 3), a graph attention neural network model was constructed using the PyTorch Geometric framework. Each layer had eight attention heads and used SMILES-encoded molecular structure graphs as input nodes. Node features included atom types and molecular descriptors. The model adopts a dual-branch output structure, which is used to predict tribological properties and freezing point respectively; 5) Model training: Use the training set divided in step 3) to train the graph attention neural network model constructed in step 4), using the AdamW optimizer and combining the cosine annealing strategy to dynamically adjust the learning rate to improve convergence stability; A multi-task weighted loss function was introduced during training, and the weights of each task were dynamically adjusted based on the performance of the validation set. The model's generalization ability in tribological performance and freezing point prediction tasks was evaluated on the test set, providing accurate prediction support for subsequent ionic liquid formulation screening. 6) Formula generation: The trained graph attention neural network model accepts the target temperature and lubrication performance requirements input by the user, performs standardization and batch inference; combines the simulated annealing algorithm to search for candidate ionic liquid formulas in the structure database, restricting the cation category to imidazole, pyridine or quaternary phosphonium, and excluding those containing Cl-, Br - Combination of halogen anions; 7) Optimization and Output: For the candidate ionic liquid formulas obtained in step 6), their structural stability at -40°C is evaluated using molecular dynamics simulations. A non-dominated sorting genetic algorithm is used for multi-objective comprehensive optimization. The preferred ionic liquid formulas with high scores are output, and their SMILES encoding and predicted performance data are displayed through a visual interface.

2. The method according to claim 1, wherein: In steps 1) and 2), the data sources of the initial data set include professional databases, SCI-indexed literature, and ASTM standard experimental data. Each data item is annotated with the source type and confidence level.

3. The method according to claim 1, wherein: In step 2), the molecular descriptor also includes molecular topological polar surface area and lipophilicity parameters, which are calculated using the RDKit toolkit.

4. The method according to claim 1, wherein: In step 3), stratified sampling is performed based on the application scenario labels in a ratio of 7:1.5:1.5, and the initial dataset after structure encoding is divided into training set, validation set and test set.

5. The method according to claim 1, wherein: In step 4), the performance indicators of the dual-branch output structure include the first branch predicting the friction coefficient, wear rate and viscosity, the second branch predicting the freezing point, and the branches output the results.

6. The method according to claim 1, wherein: In step 5), a multi-task weighted loss function is introduced in the model training process. The specific form of the loss function is as follows: L total =∑(w k ×L k ); Where k = 1 to 4, L k represents the loss value of the k-th prediction task, w k Indicates the weight coefficient of the task.

7. The method according to claim 6, characterized in that: The weight coefficient w k Dynamic calculation according to the following formula: w k =exp(η×Acc k ) / ∑[exp(η×Acc j )]; Where j = 1 to 4, Acc k represents the normalized accuracy evaluation value of the k-th task on the validation set, η is the smoothing factor, and its value range is 0.5 to 1.

0.

8. The method according to claim 1, wherein: In step 6), the structural stability of the candidate ionic liquid formulation is evaluated by molecular dynamics simulation. The molecular dynamics simulation is performed at -40°C and an AMBER force field is used to simulate a 10 ns trajectory.

9. The method according to claim 1, wherein: In step 7), based on the performance indicators obtained from the dual-branch output structure, a non-dominated sorting genetic algorithm is used to perform a multi-objective comprehensive scoring of the candidate ionic liquid formulas. The multi-objective optimization scoring includes four optimization goals: minimizing the friction coefficient, minimizing the wear rate, approaching the viscosity to the target value, and minimizing the freezing point.

10. The method according to claim 1, wherein: In step 7), the preferred ionic liquid formula refers to a candidate ionic liquid formula having a score higher than eight points after a multi-objective optimization scoring is performed on the candidate ionic liquid formula using a non-dominated sorting genetic algorithm on a ten-point scale.

Images