A method for constructing an ionic liquid machine learning potential function considering special hydrogen bonds

Abstract

The application provides a method for constructing an ionic liquid machine learning potential function considering special hydrogen bonds, which is used to solve the technical problem of the mismatch between efficiency and accuracy in the existing molecular simulation technology, especially the deficiency in high-precision analysis of special hydrogen bonds in ionic liquids. The steps of the application are as follows: using DFT to calculate ionic liquids to construct an initial data set; using a deep learning model of machine learning to fit the initial data set to generate a machine learning potential function model; using the potential function model to perform molecular dynamics simulation under constant temperature and constant volume ensemble, analyzing the simulation results, and combining the upper and lower limits of the force error to select new conformations; performing single-point calculation on the selected conformations, collecting and arranging the data of the obtained conformations to form a new data set with the existing data set, and training the deep potential model using the new data set until a deep potential model meeting the accuracy is obtained. The machine learning potential function model can well describe the special hydrogen bond interaction of ionic liquids, and provides a basis support for high-precision simulation and rational design of ionic liquids.

Description

Technical Field

[0001] This invention relates to the technical field of chemical thermodynamics, and in particular to a method for constructing a machine learning potential function for ionic liquids that takes into account specific hydrogen bonds. Background Technology

[0002] Ionic liquids, as green and environmentally friendly solvents, exhibit excellent properties such as low vapor pressure, high chemical stability, and a wide electrochemical window. Utilizing molecular simulation techniques to analyze the relationship between the structure and properties of ionic liquids at the molecular scale helps to understand their scientific essence and provides theoretical support for their engineering applications.

[0003] In the field of computational simulation of ionic liquids, the two main methods are quantum chemical calculations and molecular dynamics simulations. Quantum chemical calculations, based on the principles of quantum mechanics, have a significant advantage in studying electronic structure. This simulation method is highly accurate, but limited by computational resources, it is usually only suitable for simulations of small systems and short timescales. In contrast, classical force-field molecular dynamics simulations simulate the system by artificially fitting internal interaction potentials and simulating the system based on empirical force fields. This method is usually highly efficient and has a longer simulation time, making it suitable for larger systems, but its accuracy is often not guaranteed, and its computational precision is frequently questioned.

[0004] To address this challenge, this invention develops a novel potential function for ionic liquids based on quantum chemical calculations and classical force field simulations. This method utilizes machine learning to learn from quantum chemical calculation data, construct a relevant potential function model, and then perform molecular dynamics simulations. This approach aims to combine the precision of quantum chemical calculations with the efficiency of classical force field simulations, providing strong support for the high-precision structural property research and novel design of ionic liquids.

[0005] Patent application number 202310176655.3 discloses a method for encoding ionic liquid features based on graph neural networks, including the following steps: 1) collecting the names of smils and physicochemical properties of ionic liquids; 2) constructing molecular graphs of ionic liquid cations and anions, generating node and edge information; 3) fusing the encoded node and edge information using a graph neural network to generate molecular features of the ionic liquid; 4) training a model on the generated molecular features with the prediction of ionic liquid physicochemical properties as the target. This invention provides an ionic liquid feature graph encoding generation technology that integrates atomic and structural properties, achieving direct fusion of atomic and structural features of ionic liquid composition, resulting in simple feature composition. The feature training model built based on this invention can achieve better property prediction results with less computational resources. However, the physicochemical property prediction model of this invention is based on experimental or computational simulation data, which has limited sources and inconsistent quality. Summary of the Invention

[0006] To address the mismatch between efficiency and accuracy in existing molecular simulation techniques, particularly in the high-precision analysis of specific hydrogen bonds in ionic liquids, this invention proposes a method for constructing a machine learning potential function for ionic liquids that considers these specific hydrogen bonds. Using the trained potential function, molecular dynamics simulations with near-quantum chemical precision can be achieved, not only improving simulation accuracy but also providing new possibilities for related research and applications. This invention provides a more accurate potential function for molecular simulation of ionic liquids, thereby generating more accurate physicochemical property data, which can better guide data prediction and experimental design.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows: a method for constructing a machine learning potential function for ionic liquids considering special hydrogen bonds, comprising the following steps:

[0008] Step 1: Use DFT to calculate ionic liquids and construct an initial dataset;

[0009] Step 2: Use a deep learning model from machine learning to fit the initial dataset and generate a potential function model from machine learning.

[0010] Step 3: Perform molecular dynamics simulations under isothermal and constant-volume ensemble using the potential function model, analyze the simulation results, and screen out new conformations by combining the upper and lower limits of force error.

[0011] Step 4: Perform single-point calculations on the conformations selected in Step 3, collect and organize the data of the obtained conformations, add them to the initial dataset to form a new dataset, and use the new dataset to train the deep potential energy model to generate a deep potential energy model with the required accuracy.

[0012] Preferably, the method for calculating ionic liquids using DFT is as follows: the initial structure of the ionic liquid is optimized by calling VASP software, a series of structural groups with slight differences are generated by expansion and contraction perturbation, molecular dynamics simulation based on first principles is performed, and initial data for force field training is obtained.

[0013] Preferably, the first-principles molecular dynamics simulation has 50 steps and a step size of 0.5 femtoseconds.

[0014] The ionic liquid is an imidazole ionic liquid, and the anions include one or more of the following: bromide ion, chloride ion, iodide ion, nitrate ion, hexafluorophosphate ion, tetrafluoroborate ion, acetate ion, thiocyanate ion, trifluoromethanesulfonate ion, trifluoromethanesulfonylimide ion, and dicyandiamide ion.

[0015] The initial dataset includes 1000 frames of initial data, and the dpdata tool is used to convert the data obtained from first-principles molecular dynamics simulations into the format required for training the DeePMD-kit software.

[0016] Preferably, the method for fitting the initial dataset using a deep learning model of machine learning is as follows: four independent force field models are trained simultaneously using the DeePMD-kit software, with different random number seeds for the four force field models; a deep neural network is used with the structure and energy in the local coordinate system as input data, and the hidden layer calculates the energy of the system through a series of linear and nonlinear transformations and outputs the results.

[0017] Preferably, each machine learning training step is 400,000 steps; the input data is information on coordinates, forces, energy, and virials obtained from first-principles molecular dynamics simulations.

[0018] Preferably, in the force field model construction, the potential energy of the system is the sum of the energies of each atom, i.e., E = Σ i E i E i The energy of the i-th atom in the system is determined by the i-th atom at the cutoff radius R. i The local environment within the atom determines the potential energy; a function model from atomic coordinates to atomic potential energy is trained using a deep neural network; the coordinates of the i-th atom are mapped to a descriptor through a deep neural network, and then mapped to an energy value through another deep neural network.

[0019] The deep neural networks described are all three-layer DNN networks with residual modules added. The descriptor for the deep neural network is se_e2_a.

[0020] Preferably, the upper and lower limits of the force error are 0.1 and 0.3, respectively; in the molecular dynamics simulation, each step represents a frame configuration, and error detection is performed on each frame configuration: Wherein, model bias ε t R represents the maximum standard deviation of the atomic force in the computational model. t Indicates the current frame configuration. F represents the first-principles force on the i-th atom in the current frame structure. w,i (R t ) represents the ensemble average force of the current frame structure; candidate configurations are selected based on model deviations between the upper and lower limits.

[0021] Preferably, the single-point calculation utilizes the VASP software to collect and organize the candidate conformation data through the dpdata tool and add it to the initial dataset;

[0022] A new round of training is performed on the new dataset until a model with an accuracy exceeding 90% and a failure rate close to 0 is trained; the accuracy is calculated as the number of accurate configurations / the total number of configurations * 100%.

[0023] Preferably, based on the deep potential energy model trained at the initial temperature, the temperature conditions in the input file are changed to explore datasets at different temperatures, and the new dataset is used to train the deep potential energy model.

[0024] When the system temperature is 300K, the temperature range involved is 250K to 700K.

[0025] Preferably, in molecular dynamics simulations, the vibrational density is calculated using a Fourier transform of the atomic velocity autocorrelation function:

[0026]

[0027] Where P(ω) is the vibrational density at frequency ω, N is the total number of atoms in the system, and v i (t) represents the velocity of the i-th atom in the system at time t. <v i (t)v i (0)> is the velocity autocorrelation function, d t Integral representing time.

[0028] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0029] (1) Machine learning models have the ability to learn complex data and exhibit highly nonlinear functional dependencies. This invention applies machine learning technology to the construction of ionic liquid potential functions. A deep learning model is established using DeepMD-kit and DPGEN tools on data from quantum chemical calculations to construct a potential function model that integrates structure-energy and structure-force. Applying this potential function model to molecular dynamics (MD) simulations effectively balances the efficiency and accuracy of the simulation, thus effectively overcoming the limitations of traditional quantum chemical calculations and empirical force field (EFF) simulations.

[0030] (2) By employing an iterative training method, this invention significantly improves the adaptability and accuracy of the potential function: through more than 30 iterations of training, covering a temperature range of 250K to 700K, this invention can accurately simulate different hydrogen bond configurations of ionic liquids. In the short-time molecular dynamics simulation process, force error is used as the main criterion in the early stage of training to screen different conformations and calculate their single-point energies, thereby effectively expanding the training dataset.

[0031] (3) A machine learning potential function verification criterion was proposed: First, 200 frames of data that were not used in the training were selected as the test set. Through single-point calculation based on density functional theory (DFT) and molecular dynamics simulation based on deep potential energy (DP), the force and energy errors of 10 ionic liquids such as [Emim]Br to the test dataset were calculated. The force and energy showed good consistency. Second, the distribution of key bond lengths and bond angles and the vibration distribution of ionic liquids during the simulation were analyzed and compared with empirical force fields and quantum chemical calculations. The results showed that the accuracy of the machine learning training force field was comparable to that of DFT. Finally, the special hydrogen bond interactions of ionic liquid anions and cations were analyzed. The potential function model of machine learning can describe the special hydrogen bond interactions of ionic liquids well.

[0032] This invention provides fundamental support for the high-precision simulation and rational design of ionic liquids, and is expected to play a significant role in related fields. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 This is the main flowchart of the present invention.

[0035] Figure 2 This is a sample iteration diagram of the machine learning potential function of the present invention at different temperatures.

[0036] Figure 3 The diagram shows a comparison of the force and energy errors of the quantum chemical calculation and machine learning potential functions of ionic liquids in this invention, where (a) is the force comparison and (b) is the energy comparison.

[0037] Figure 4 The distribution of CN bond length and NCN bond angle of the ionic liquid cation imidazole ring of the present invention is shown, wherein (a) is the bond length distribution and (b) is the bond angle distribution.

[0038] Figure 5 This is a schematic diagram of the vibrational dynamic density analysis of the ionic liquid of the present invention.

[0039] Figure 6 This is a schematic diagram of the hydrogen bond length distribution of the ionic liquid of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] like Figure 1 As shown, a method for constructing a machine learning potential function for ionic liquids considering special hydrogen bonds is presented, and its main steps are as follows:

[0042] Step 1: Use DFT to calculate ionic liquids and construct an initial dataset.

[0043] The initial structure of the ionic liquid was optimized using VASP software (https: / / www.vasp.at / ). Next, scaling perturbations were applied to generate a series of structural groups with slight differences. First-principles molecular dynamics (AIMD) simulations were then performed to obtain initial data for force field training. The AIMD simulation consisted of 50 steps with a step size of 0.5 femtoseconds (fs) to expand the data sample.

[0044] The ionic liquid is an imidazole-based ionic liquid, wherein the anions include: bromide ions (Br₂). - ), chloride ions (Cl) - ), iodide ions (I) - ), nitrate ions (NO3) - ), hexafluorophosphate ions (PF6) - ), tetrafluoroborate ion (BF4) - Acetate ions (AC) - ), thiocyanate ion (SCN) - ), trifluoromethanesulfonate ion (OTf) - ), trifluoromethanesulfonylimide ion (TFSI) - ) and dicyandiamide ions (DCA) - (e.g., to study the universality of machine learning potential functions for different ionic liquids.)

[0045] The initial dataset was constructed based on AIMD simulation, collecting 1000 frames of data. The dpdata tool was then used to convert the data obtained from AIMD into the format required for training the DeePMD-kit software.

[0046] Step 2: Use a deep learning model of machine learning to fit the initial dataset obtained in Step 1, thereby generating a machine learning potential function model (also known as a deep potential energy model).

[0047] Four independent force field models were trained simultaneously using the DeePMD-kit software (https: / / docs.deepmodeling.com / ), the only difference between the four models being their random number seeds. Deep neural networks (DNNs) were used, with information such as structure and energy in the local coordinate system (coordinates, forces, energy, and virial information obtained from AIMD simulations) as input data. The hidden layers calculated the system's energy through a series of linear and nonlinear transformations and output the results. Each machine learning training session was set to 400,000 steps.

[0048] In the force field model, the potential energy of the system is defined as the sum of the energies of each atom, i.e., E = Σ i E i E i This represents the energy of the i-th atom in the system, which is determined by the i-th atom at the cutoff radius R. i The potential energy of an atom is determined by its local environment. To simplify calculations, we can approximate it as the potential energy of each atom being only related to itself and its cutoff radius R. i The coordinates of neighboring atoms within the potential energy range are relevant. Therefore, a DNN network is considered to train a function model from atomic coordinates to their atomic potential energy. However, this method cannot accurately represent the translation, rotation, and permutation symmetries in the potential energy surface. Therefore, the coordinates of atom i are first mapped to a descriptor through a DNN network, and then mapped to the energy value through another DNN network. The mapping neural networks used are simple three-layer DNN networks with residual modules added. The descriptor selected in this invention is se_e2_a. The concept of the local environment of atoms ensures that the potential function based on the neural network can be extrapolated from a small system to a large system.

[0049] Step 3: Use the potential function model from Step 2 to perform molecular dynamics (MD) simulations under an isothermal constant volume (NVT) ensemble. By analyzing the simulation results and considering the upper and lower limits of force error, new conformations are selected.

[0050] The upper and lower limits of force error are set to 0.1 and 0.3, respectively. In the MD simulation, each step represents a frame configuration. Here, an error criterion needs to be introduced to detect errors in each frame configuration. Where, ε t R represents the maximum standard deviation of the atomic force in the computational model, serving as the model bias. t Indicates the current frame configuration. F represents the first-principles force on the i-th atom in this frame structure. w,i (R tThe value represents the ensemble average force of the frame structure. Through this step, candidate configurations with model biases between the upper and lower limits are selected and proceed to the next labeling stage. A bias below the set lower limit means the trained force field can accurately predict this configuration without requiring new training; a bias above the set lower limit indicates the possible existence of non-physical configurations, and therefore these are excluded.

[0051] Step 4: Use VASP software to perform single-point calculations on the candidate conformations selected in Step 3, and use the dpdata tool to collect and organize the data and add it to the initial dataset to form a new dataset. Use the new dataset to train the deep potential energy model to obtain a deep potential energy model with the required accuracy.

[0052] Iteration: A new round of training is performed on a new dataset. This iterative training process continues until a model with an accuracy exceeding 90% and a failure rate close to 0 is trained. Accuracy = (Number of accurate configurations / Total number of configurations) * 100%. The toolkit DPGEN (https: / / docs.deepmodeling.com / ) collects and organizes the data from each iteration during the training process.

[0053] Variable Temperature: Based on the machine learning potential function model trained at the initial temperature, the temperature conditions in the input file are further changed to explore the dataset at different temperatures, thereby improving the reliability and stability of the potential function.

[0054] In this invention, the different temperatures refer to temperatures fluctuating within a certain range around the target temperature. For example, when the temperature of the system under study is 300K, the temperature range involved is defined as 250K to 700K.

[0055] In the embodiments of this invention, an application of machine learning potential functions for ionic liquids is demonstrated to analyze the unique hydrogen bond properties of ionic liquids. This invention enables in-depth analysis of the unique hydrogen bond properties of ionic liquid systems while maintaining accuracy and efficiency, providing fundamental support for establishing structure-property relationships for ionic liquid systems. This is of great significance for promoting the functional development and practical applications of ionic liquids.

[0056] The specific implementation method of the present invention will be described in detail below with reference to specific embodiments.

[0057] Example 1

[0058] Taking 1-ethyl-3-methylimidazolium bromide ionic liquid ([Emim]Br) as an example, the training situation of each iteration in the machine learning potential function training process is given, as well as the changes in the number of detected conformations and the success rate. Figure 2As shown, the [Emim]Br potential function model underwent 39 iterations during training, covering a temperature range from 250K to 700K. By the end of the 12th iteration, the success rate of the conformations screened at 300K reached 99%, indicating relatively successful training at 300K. However, in actual molecular dynamics simulations, it was found that the trained potential function was completely inadequate for nanosecond (ns) level simulations. Due to the strong vibrational characteristics of ions, the system temperature easily rises above 400K, or even higher. When the system temperature reaches 400K, the trained potential function does not include conformations at this temperature, leading to significant structural deformation and ultimately simulation interruption. This highlights the importance of sufficient training across different temperature ranges to ensure the potential function model remains stable and accurate under a wider range of application conditions.

[0059] After a series of tests, it was found that when the training temperature of the potential function is increased to above 700K, the nanosecond-level simulation requirements of the NVT ensemble at 300K can be met. Meanwhile, it was observed that at 250K, 300K, and 500K, the success rate of conformations gradually approaches 100% with increasing iterations, while at 700K it only reaches 90%, but the failure rate remains below 0.5%. This illustrates the difficulty of training at high temperatures, but high-temperature training is an indispensable and crucial step in the training of the ionic liquid potential function.

[0060] Example 2

[0061] The machine learning potential function of 1-ethyl-3-methylimidazolium chloride ionic liquid ([Emim]Cl) was trained using the same steps as in Example 1. After 38 iterations of training, covering a temperature range of 250-700 K, a training success rate of 95% was achieved, while maintaining a failure rate of less than 0.5%.

[0062] Example 3

[0063] Analysis of the force and energy errors in the test dataset in Example 1 shows the prediction results as follows: Figure 3 As shown, by comparing the molecular dynamics simulations based on the machine learning potential function model and the calculation results of DFT, the two show good consistency, with an accuracy greater than 95%, which verifies the rationality of the trained model. Figure 3 ).

[0064] Example 4

[0065] The machine learning potential function model trained in Example 1 was used as the force field parameter for LAMMPS (https: / / www.lammps.org / ) for molecular dynamics simulation. The results show that the simulation can almost achieve the accuracy of quantum chemical calculations, and the simulation time can reach 2 ns. In terms of efficiency, the method of this invention is comparable to empirical force field simulation. Based on the above molecular simulation results, the evolution process of ionic liquid structure was further analyzed: the CN bond length and NCN bond angle on the imidazole ring were selected for analysis, and the bond length and bond angle distributions in deep potential energy (DP) molecular dynamics simulation, DFT molecular dynamics simulation, and empirical force field (EFF) based molecular dynamics simulation were obtained. The differences between the three were compared, and the results are as follows. Figure 4 As shown. The results show that the bond length and bond angle distribution simulated by the machine learning potential function trained in this invention perfectly matches the first-principles calculations, and there are significant differences from the empirical force field simulation results.

[0066] Example 5

[0067] MD simulations were performed using the potential function model trained in Example 1 to analyze the vibrational spectrum information of the ionic liquid. Vibrational density of motion (VDOS) is a measurement of the phonon or vibrational distribution, providing spectral energy information and detecting dominant vibrational modes. In MD simulations, VDOS can be calculated using the Fourier transform of the atomic velocity autocorrelation function, as shown in the following formula:

[0068]

[0069] Where P(ω) is the vibrational spectrum at frequency ω, N is the total number of atoms in the system, and v i (t) represents the velocity of the i-th atom in the system at time t. <v i (t)v i (0)> is the velocity autocorrelation function, d t Integral representing time.

[0070] Taking [Emim]Br as an example, the VDOS of this ionic liquid was calculated using three simulation methods: DP, DFT, and EFF. The DP and EFF simulations used a 200 ps trajectory after optimization and equilibrium, taking one frame of the trajectory every 10 fs, and using the velocity of each atom in that frame structure as input data. The DFT simulation used a 20 ps trajectory, taking one frame every 1 ps, and similarly calculated the atom velocities as input. The results are as follows. Figure 5 As shown. By Figure 5 It can be seen that at a frequency of 700cm -1 1080cm -1 and 3215cm -1At this frequency, the EFF value is almost zero, indicating that no vibrational modes were captured. Conversely, the DP and EFF values ​​show distinct peaks at these three frequencies. Gaussian analysis of the specific vibrational modes at these frequencies revealed a peak at 700 cm⁻¹. -1 and 1080cm -1 The rotation is due to the imidazole ring itself. This can also be observed at a frequency of 3200 cm⁻¹. -1 In the vicinity, simulations under DP clearly reveal the vibrational modes induced by hydrogen bonding interactions between cations and anions in ionic liquids, which is impossible to achieve with traditional empirical force fields, marking a significant advantage of machine learning potential functions in this regard. Furthermore, the machine learning potential function can describe the torsional and stretching vibrations on the cationic imidazole ring in greater detail, almost reproducing all the precise vibrational characteristics obtained from first-principles calculations, thus reflecting the high-precision performance of the machine learning potential function trained in this invention in simulating the vibrational spectrum information of ionic liquids.

[0071] Example 6

[0072] The specific hydrogen bond interactions between the ionic liquid cations and anions were analyzed based on the DP, DFT, and EFF force field simulation results from Example 5. The bond length distribution results are as follows: Figure 6 As shown. By Figure 6 It can be seen that the peak value of the bond length distribution based on DP and DFT simulations is approximately The results are consistent, while the EFF simulation results deviate significantly from the DFT calculation results. This confirms the accuracy of the machine learning potential function model in predicting the specific hydrogen bonding interactions between cations and anions in ionic liquids.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a machine learning potential function for ionic liquids considering specific hydrogen bonds, characterized in that, The steps are as follows: Step 1: Use DFT to calculate ionic liquids and construct an initial dataset; Step 2: Use a deep learning model from machine learning to fit the initial dataset and generate a potential function model from machine learning. Step 3: Perform molecular dynamics simulations under isothermal and constant-volume ensemble using the potential function model, analyze the simulation results, and screen out new configurations by combining the upper and lower limits of force error. Step 4: Perform single-point calculations on the configurations selected in Step 3, collect and organize the data of the obtained configurations, add them to the initial dataset to form a new dataset, and use the new dataset to train the deep potential energy model to generate a deep potential energy model with the required accuracy. The method for fitting the initial dataset using a deep learning model for machine learning is as follows: four independent force field models are trained simultaneously using the DeePMD-kit software, with different random number seeds for the four force field models; a deep neural network is used with the structure and energy in the local coordinate system as input data, and the hidden layer calculates the energy of the system through a series of linear and nonlinear transformations and outputs the results. In the force field model, the potential energy of the system is the sum of the energies of each atom, i.e. , of which E i The energy of the i-th atom in the system is determined by the i-th atom at the cutoff radius R. i The local environment within the atom determines the potential energy; a function model from atomic coordinates to atomic potential energy is trained using a deep neural network; the coordinates of the i-th atom are mapped to a descriptor through a deep neural network, and then mapped to an energy value through another deep neural network; In molecular dynamics simulations, each step represents a configuration frame, and error detection is performed on each configuration frame: Among them, model bias This represents the maximum standard deviation of the atomic force in the computational model; Indicates the current frame configuration. This represents the first-principles force on the i-th atom in the current frame structure. The ensemble average force represents the current frame structure; candidate configurations are selected based on model biases between the upper and lower limits. In molecular dynamics simulations, vibrational density is calculated using the Fourier transform of the atomic velocity autocorrelation function: ； Where P(ω) is the vibrational density at frequency ω, and N is the total number of atoms in the system. Let be the velocity of the i-th atom in the system at time t, < > represents the velocity autocorrelation function. Integral representing time.

2. The method for constructing the machine learning potential function of ionic liquids considering special hydrogen bonds according to claim 1, characterized in that, The method for calculating ionic liquids using DFT is as follows: the initial structure of the ionic liquid is optimized by calling VASP software, a series of structural groups with slight differences are generated by scaling perturbation, molecular dynamics simulation based on first principles is performed, and initial data for force field training is obtained.

3. The method for constructing the machine learning potential function of ionic liquids considering special hydrogen bonds according to claim 2, characterized in that, The first-principles molecular dynamics simulation consists of 50 steps with a step size of 0.5 femtoseconds. The ionic liquid is an imidazole ionic liquid, and the anions include one or more of the following: bromide ion, chloride ion, iodide ion, nitrate ion, hexafluorophosphate ion, tetrafluoroborate ion, acetate ion, thiocyanate ion, trifluoromethanesulfonate ion, trifluoromethanesulfonylimide ion, and dicyandiamide ion. The initial dataset includes 1000 frames of initial data, and the dpdata tool is used to convert the data obtained from first-principles molecular dynamics simulations into the format required for training the DeePMD-kit software.

4. The method for constructing the machine learning potential function of ionic liquids considering special hydrogen bonds according to claim 1 or 3, characterized in that, Each machine learning training session consists of 400,000 steps; the input data consists of coordinates, forces, energy, and virial information obtained from first-principles molecular dynamics simulations.

5. The method for constructing the machine learning potential function of ionic liquids considering special hydrogen bonds according to claim 4, characterized in that, The deep neural networks described are all three-layer DNN networks with residual modules added. The descriptor for the deep neural network is se_e2_a.

6. The method for constructing the machine learning potential function of ionic liquids considering special hydrogen bonds according to claim 1 or 5, characterized in that, The upper and lower limits of the force error are 0.3 and 0.1, respectively.

7. The method for constructing the machine learning potential function of ionic liquids considering special hydrogen bonds according to claim 6, characterized in that, The single-point calculation calls the VASP software, and the candidate configuration data collected and organized by the dpdata tool is added to the initial dataset. A new round of training is performed on the new dataset until a model with an accuracy of over 90% is trained; the accuracy is the number of accurate configurations / total number of configurations * 100%.

8. The method for constructing the machine learning potential function of ionic liquids considering special hydrogen bonds according to claim 7, characterized in that, Based on the deep potential energy model trained at the initial temperature, the temperature conditions in the input file are changed to explore datasets at different temperatures, and the new dataset is used to train the deep potential energy model. When the system temperature is 300 K, the temperature range involved is 250 K to 700 K.

Images