A method and system for predicting the products of cellulose pyrolysis reaction

Abstract

The present invention discloses a method and system for predicting the products of cellulose pyrolysis reaction. First, the present invention constructs the cellulose molecular structure, performs energy minimization and structure relaxation processing to obtain the initial cellulose molecular configuration; then performs ab initio molecular dynamics simulation on the initial cellulose molecular configuration for a preset time period to obtain the simulated cellulose molecular configuration and its atomic forces, constructs an initial training set, and uses an iterative method of parallel training, cross-validation, and expansion of the training set for the multi-layer neural network model to obtain the deep learning potential function for cellulose pyrolysis; uses the deep learning potential function for cellulose pyrolysis to perform molecular dynamics simulation to obtain the types and structural information of cellulose pyrolysis products. The present invention uses a deep learning algorithm for model training and uses the obtained deep learning potential function for cellulose pyrolysis to perform molecular dynamics simulation, realizing the prediction of cellulose pyrolysis products with low cost, high speed, and scalable application.

Description

Technical Field

[0001] The present invention relates to the technical fields of chemical reaction molecular dynamics and biomass pyrolysis technology, and particularly relates to a method and a system for predicting cellulose pyrolysis reaction products. Background Art

[0002] In recent years, agricultural and forestry biomass energy, as a renewable clean energy, has been continuously valued by all parties. Considering that cellulose is its main component, the research on the cellulose pyrolysis mechanism can reflect the overall pyrolysis law of biomass and is the basis for further comprehensive utilization of biomass components. Usually, the temperature range of cellulose in industrial pyrolysis is 500-600°C, and clean energy such as biogas and bio-oil and multifunctional biochar materials can be produced. However, affected by the heat transfer, mass transfer, temperature, and pressure parameters of the pyrolysis boiler equipment, the distribution of cellulose pyrolysis products is uneven and it is difficult to obtain efficient application. Therefore, predicting pyrolysis products and clarifying the reaction mechanism are important prerequisites for the popularization and application of cellulose pyrolysis.

[0003] Currently, the prediction of cellulose pyrolysis products is mainly based on two directions: experiment and simulation. Experiments are usually based on TGA-FTIR (Thermogravimetric analysis-Fourier Transform Infrared) for isothermal or non-isothermal pyrolysis experiments. The thermogravimetric curve can clarify the cellulose thermal decomposition behavior, and the FTIR characterization results can provide product functional group information. However, there are still certain limitations in experimental techniques. Due to the fast heating rate, short product residence time, and short half-life of reaction intermediates, only the final state of pyrolysis products can be observed. The biogas, bio-oil, and biochar obtained from cellulose pyrolysis compete with each other. Exploring the reaction path based on the reaction path of generating small molecule biogas can regulate products. The functional group structures of the obtained bio-oil and biochar are complex and their surface functional group characteristics will affect subsequent applications. Therefore, it is difficult to study the reaction path experimentally, the characterization has limitations, and the cost is high. Simulations are usually based on ab initio molecular dynamics (AIMD) or REAXFF (Reactive force field) reaction force fields. The AIMD calculation based on the first principle is accurate but takes a lot of time. The cellulose composed of 348 atoms requires 6-8 years to calculate and simulate 250 ps of pyrolysis based on a 28-core CPU. The REAXFF parameters vary in different models and chemical reactions. If used without parameter adjustment, the reaction mechanism cannot be accurately described, and the cost of parameter adjustment for complex systems is high and requires complete first principle calculations or experimental data. Therefore, it is difficult to predict the distribution and structure of cellulose pyrolysis products at low cost using existing methods. Summary of the Invention

[0004] The object of the present invention is to provide a method and system for predicting the products of cellulose pyrolysis reaction, so as to realize the prediction of the distribution and structure of cellulose pyrolysis products with low cost and high speed.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for predicting the products of cellulose pyrolysis reaction, the method comprising the following steps:

[0007] Construct a cellulose molecular structure and calculate the volume parameter of the simulation box;

[0008] Perform energy minimization and structural relaxation on the cellulose molecular structure and the simulation box to obtain an initial cellulose molecular configuration with optimized atomic spacing, bond angles and dihedral angles;

[0009] Perform ab initio molecular dynamics simulation on the initial cellulose molecular configuration for a preset period of time to obtain multiple simulated cellulose molecular configurations and the atomic forces of each cellulose molecular configuration, and construct an initial training set;

[0010] Based on the initial training set, use a multi-layer neural network model for parallel training, cross-validation, and iterative method of expanding the training set to obtain a cellulose pyrolysis deep learning potential function;

[0011] Use the cellulose pyrolysis deep learning potential function to perform molecular dynamics simulation to obtain the types and structural information of cellulose pyrolysis products.

[0012] Optionally, the formula for calculating the volume parameter of the simulation box is:

[0013]

[0014]

[0015] wherein, V represents the volume of the pyrolysis chamber, m represents the mass of the cellulose sample, mW represents the relative molecular mass of the cellulose sample; ρ1 represents the molecular density, Vs represents the volume parameter of the simulation box, N A represents Avogadro's constant.

[0016] Optionally, the performing energy minimization and structural relaxation on the cellulose molecular structure and the simulation box to obtain an initial cellulose molecular configuration with optimized atomic spacing, bond angles and dihedral angles specifically includes:

[0017] Use one or more of the stochastic gradient descent algorithm, the steepest gradient descent algorithm, and the conjugate gradient descent algorithm to perform energy minimization on the cellulose molecular structure in the simulation box to obtain a cellulose molecular configuration after energy minimization;

[0018] Perform structural relaxation on the cellulose molecular configuration after energy minimization to obtain the initial cellulose molecular configuration with optimized atomic distances, bond angles, and dihedral angles after structural relaxation processing.

[0019] Optionally, the iterative method of parallel training, cross-validation, and augmented training set using a multi-layer neural network model based on the initial training set to obtain the deep learning potential function for cellulose pyrolysis specifically includes:

[0020] Set the value of the iteration number k to 1;

[0021] Set the initial training set as the training set for the k-th iteration;

[0022] Select a preset number of training subsets from the training set for the k-th iteration;

[0023] Input the cellulose molecular configurations in each training subset into the multi-layer neural network model corresponding to each training subset one by one to obtain the system energy and atomic forces of each cellulose molecular configuration;

[0024] Calculate the loss function values of the multi-layer neural network models corresponding to each training subset respectively according to the system energy and atomic forces of each cellulose molecular configuration;

[0025] When the loss function value is not less than the loss function threshold, update the parameters of the multi-layer neural network model corresponding to each training subset, and return to the step "Input the cellulose molecular configurations in each training subset into the multi-layer neural network model corresponding to each training subset one by one to obtain the system energy and atomic forces of each cellulose molecular configuration";

[0026] When the loss function value is less than the loss function threshold, output the multi-layer neural network model corresponding to each training subset as the deep learning potential function for cellulose pyrolysis for the k-th iteration corresponding to each training subset.

[0027] Optionally, after the step of "When the loss function value is less than the loss function threshold, output the multi-layer neural network model corresponding to each training subset as the deep learning potential function for cellulose pyrolysis for the k-th iteration corresponding to each training subset", it further includes:

[0028] Calculate the model deviation of each atom of each cellulose molecular configuration according to the atomic forces of each cellulose molecular configuration in each training subset, and obtain the maximum model deviation of each cellulose molecular configuration in each training subset; the maximum model deviation is the maximum value of the model deviations of all atoms of the cellulose molecular configuration;

[0029] Select the cellulose molecular configuration at the moment when the maximum model deviation is within the preset range and the atomic forces of the cellulose molecular configuration as training samples, expand the training set of the k-th iteration to obtain the training set of the (k + 1)-th iteration, increase the value of k by 1, and return to the step of "selecting a preset number of training subsets from the training set of the k-th iteration" until the maximum model deviation of the cellulose molecular configurations at different moments of the pre-proportion is less than the deviation threshold.

[0030] Optionally, the calculation formula of the loss function value is:

[0031]

[0032] where L(p ε , p f ) represents the loss function value, N is the number of atoms in the cellulose molecule, ΔE is the root mean square error of the system energy corresponding to each training subset; ΔF i is the root mean square error of the atomic force of the i-th atom of the cellulose molecular configuration in each training subset, p ε and p f respectively represent the pre-exponential factors related to time of the energy term and the force term.

[0033] Optionally, the calculation formula of the model deviation is:

[0034]

[0035] where is the average force of the i-th atom calculated from the output of the deep learning cellulose pyrolysis deep learning potential function, ε is the maximum standard deviation of the atomic force, and d is a constant.

[0036] Optionally, use the deep learning potential function of cellulose pyrolysis to perform molecular dynamics simulation to obtain the types and structural information of cellulose pyrolysis products, specifically including:

[0037] Use the deep learning potential function of cellulose pyrolysis to perform molecular dynamics simulation to obtain the atomic motion trajectory of the cellulose molecular configuration changing with the molecular dynamics simulation step size t;

[0038] Use the open-source Openbabel software to capture the molecular configuration and chemical formula of the obtained products from the atomic motion trajectory;

[0039] Based on the molecular configuration and chemical formula of the obtained products, obtain the types, structures and quantities of cellulose pyrolysis products.

[0040] A system for predicting cellulose pyrolysis reaction products, the system includes:

[0041] A molecular structure construction and volume parameter calculation module for constructing a cellulose molecular structure and calculating the volume parameters of a simulation box;

[0042] An energy minimization and structural relaxation processing module for performing energy minimization processing and structural relaxation processing on the cellulose molecular structure and the simulation box to obtain an initial cellulose molecular configuration with optimized atomic spacing, bond angles, and dihedral angles;

[0043] An ab initio molecular dynamics simulation module for performing ab initio molecular dynamics simulation on the initial cellulose molecular configuration for a preset time period to obtain multiple simulated cellulose molecular configurations and the atomic forces of each cellulose molecular configuration, and constructing an initial training set;

[0044] A model training module for obtaining a deep learning potential function for cellulose pyrolysis by using an iterative method of parallel training, cross-validation, and expanding the training set based on the initial training set with a multi-layer neural network model;

[0045] A deep learning potential function simulation module for cellulose pyrolysis for performing molecular dynamics simulation by using the deep learning potential function for cellulose pyrolysis to obtain the types and structural information of cellulose pyrolysis products.

[0046] Optionally, the formula for calculating the volume parameters of the simulation box is:

[0047]

[0048]

[0049] where V represents the pyrolysis chamber volume, m represents the mass of the cellulose sample, mW represents the relative molecular mass of the cellulose sample; ρ1 represents the molecular density, Vs represents the volume parameter of the simulation box, and N A represents Avogadro's constant.

[0050] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0051] The present invention discloses a method and system for predicting the products of cellulose pyrolysis reaction. The method includes the following steps: constructing a cellulose molecular structure and calculating the volume parameters of the simulation box; performing energy minimization and structural relaxation on the cellulose molecular structure and the simulation box to obtain an initial cellulose molecular configuration with optimized atomic spacing, bond angles, and dihedral angles; performing ab initio molecular dynamics simulation on the initial cellulose molecular configuration for a preset period of time to obtain multiple simulated cellulose molecular configurations and the atomic forces of each cellulose molecular configuration, and constructing an initial training set; training a multi-layer neural network model by means of parallel learning based on the initial training set to obtain a deep learning potential function for cellulose pyrolysis; and performing molecular dynamics simulation using the deep learning potential function for cellulose pyrolysis to obtain the types and structural information of cellulose pyrolysis products. The present invention uses a deep learning algorithm to train a learning model and performs molecular dynamics simulation using the obtained deep learning potential function for cellulose pyrolysis, realizing the prediction of cellulose pyrolysis products with low cost, high speed, and scalable application. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0053] Figure 1 It is a flowchart of a method for predicting the products of cellulose pyrolysis reaction provided by an embodiment of the present invention;

[0054] Figure 2 Provided by an embodiment of the present invention Figure 1 It is a flowchart of the specific implementation of step 1 in

[0055] Figure 3 Provided by an embodiment of the present invention Figure 1 It is a flowchart of the specific implementation of step 2 in

[0056] Figure 4 Provided by an embodiment of the present invention Figure 1 It is a flowchart of the specific implementation of step 3 in DETAILED DESCRIPTION OF THE EMBODIMENTS

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0058] The purpose of the present invention is to provide a method and system for predicting the products of cellulose pyrolysis reaction, so as to realize the prediction of the distribution and structure of cellulose pyrolysis products with low cost and high speed.

[0059] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] Embodiment 1

[0061] As Figure 1 shown, the embodiment of the present invention provides a method for predicting the products of cellulose pyrolysis reaction, and the method includes the following steps:

[0062] Step 1: Construct the cellulose molecular structure and perform energy minimization and structural relaxation processing on the cellulose molecular structure based on the empirical potential field. As Figure 2 shown, step 1 specifically includes:

[0063] Step 101, obtain the cellulose molecular structure through the Cellulose Builder toolbox according to the experimental parameters, record the relative molecular mass (mW) of the cellulose, and calculate the volume parameter (V s ) of the simulation box to ensure that the subsequent cellulose pyrolysis molecular dynamics simulation process is closer to the industrial production situation.

[0064] The above experimental parameters include the pyrolysis chamber volume (V), the mass (m) of the cellulose sample, and the relative molecular mass (mW) of the cellulose sample. Through the relationship:

[0065]

[0066]

[0067] calculate the molecular density (ρ1) and adjust the volume parameter (V s ) of the simulation box

[0068] Since the system pyrolysis simulation is carried out in the canonical ensemble, the setting of V s will affect the pressure received by the system. To ensure that the cellulose pyrolysis pressure meets 101325 Pa used in industrial production, it is necessary to calculate V s according to the above formula., to ensure that the subsequent molecular dynamics simulation process of cellulose pyrolysis is closer to the industrial production scenario

[0069] Step 102: Use molecular dynamics simulation software to perform energy minimization and structural relaxation on the above system containing cellulose molecular structure and simulation box, so as to obtain cellulose molecular configurations with reasonable parameters such as atomic spacing, bond angle, dihedral angle, etc., for subsequent training of cellulose pyrolysis deep learning potential function and molecular dynamics simulation.

[0070] Specifically, any molecular dynamics software can be selected for the use of the above molecular dynamics simulation software, such as GROMACS, LAMMPS, NAMD, Amber, OPENMM, etc., and users can select according to their needs.

[0071] Specifically, the above energy minimization process is carried out based on any one or more of algorithms such as stochastic gradient descent, steepest gradient descent, conjugate gradient descent, etc., and users can select according to their needs. To obtain a system with stable energy and energy minimization, and reduce unreasonable atomic arrangements.

[0072] Specifically, the structural relaxation process can be realized through empirical potential fields or force fields including GAFF (Amber force field), CHAMMER (Chemistry at HARvard Macromolecular Mechanics, macromolecular force field), ReaxFF, etc., and users can select according to their needs. Based on the NPT (isothermal isobaric) or NVT (canonical) ensemble (i.e., the above system of cellulose molecular structure and simulation box), perform molecular dynamics simulation at 300K for 200 to 800 ps until the system energy is stable and the cellulose molecules exhibit periodic stable vibrations, and finally obtain cellulose molecular configurations with reasonable parameters such as atomic spacing, bond angle, dihedral angle, etc., for subsequent training of cellulose pyrolysis deep learning potential function and molecular dynamics simulation.

[0073] Step 2: Obtain the cellulose pyrolysis deep learning potential function based on the training of a multi-layer neural network, as Figure 3 shown, Step 2 specifically includes:

[0074] Step 201: Use first-principles calculation software (software for ab initio molecular dynamics simulation) to perform short-time ab initio molecular dynamics simulation (AIMD) on the cellulose molecular configurations with reasonable parameters such as atomic spacing, bond angle, dihedral angle, etc. obtained in Step 1 at a certain temperature, so as to obtain cellulose molecular configurations (atomic coordinate information) at different time points within this short time, as well as the system energy and atomic forces of each cellulose molecular configuration, and construct an initial training set.

[0075] The above AIMD process parameters can be set to 300 to 600 fs and 2000 to 3000 K to ensure the generation and breakage of more chemical bonds in a shorter time. During the pyrolysis process simulated by this AIMD, chemical bonds in cellulose molecules will break and recombine. The process of chemical bond breakage and generation in cellulose molecules involves changes in system energy, atomic forces, atomic coordinates, etc. Therefore, the breakage and generation of more chemical bonds can make the dataset have more effective and non-repetitive information, which is helpful for the subsequent fitting training of the deep learning potential function for cellulose pyrolysis;

[0076] For the above-obtained dataset, the calculated atomic forces, coordinates, and system energy information are extracted every 1 fs to reduce data redundancy. Finally, it is used as the initial training set for deep learning to train the deep learning potential function for cellulose pyrolysis. The obtained deep learning potential function for cellulose pyrolysis will be used in the subsequent reinforcement learning process to expand the training set.

[0077] Based on the parallel deep learning method, the initial training set obtained in the above step 201 is used to construct the deep learning potential function for cellulose pyrolysis, perform molecular dynamics simulation cross-validation, extract other cellulose molecular structures, expand the training set by reinforcement learning, and iteratively train the new deep learning potential function for cellulose pyrolysis to obtain a deep learning potential function for cellulose pyrolysis that meets the requirements.

[0078] Specifically, in the process of establishing the deep learning model for the above parallel deep learning method, three to five (i.e., the value range of the preset number is 3 to 5) training subsets are randomly extracted from the same initial training set, and the generation of the deep learning potential function for cellulose pyrolysis is achieved by minimizing the loss function.

[0079]

[0080] where N is the number of atoms, ΔE and ΔF i respectively represent the root mean square error of the system energy and the root mean square error of the force on the i-th atom in the system, and p ε and p f represent the pre-exponential factors related to time for the energy term and the force term;

[0081] Specifically, in the above cross-validation process, all the obtained deep learning potential functions for cellulose pyrolysis are used to perform molecular dynamics simulations. Based on the selected molecular dynamics simulation software, the force calculation command is written to calculate the atomic forces under each molecular configuration. The obtained atomic forces will be used to calculate the maximum relative model deviation R i .

[0082]

[0083] where $\overline{F}_i$ is the average force on the $i$-th atom calculated by the deep learning potential function of cellulose pyrolysis, $\varepsilon$ is the maximum standard deviation of the atomic force, and $d$ is a constant. Ensure that atoms with smaller $\varepsilon$ also have smaller relative model deviations.

[0084] Finally, the maximum relative model deviation $R$ of all molecular configurations during the molecular dynamics simulation using 3 - 5 deep learning potential functions of cellulose pyrolysis i will be marked according to user requirements. Usually, $R$ i in the range of 0.3 to 0.5 is selected as the molecular configuration for the input data of reinforcement learning to perform the force calculation based on the first principles to expand the training set. $R$ i less than 0.3 is determined as an accurate configuration, and $R$ i greater than 0.5 is determined as a configuration to be discarded.

[0085] Specifically, the above-mentioned reinforcement learning to expand the training set is achieved through the force calculation based on the first principles. This step can be carried out using software including VASP, Gaussian, CP2K, ORCA, etc., and users can select according to their needs;

[0086] The atomic forces and coordinate information (molecular configuration) of the molecules obtained after the calculation will be used as a new dataset to expand into the deep learning training set and new deep learning potential functions of cellulose pyrolysis will be trained based on the expanded training set. The training of the new deep learning potential functions of cellulose pyrolysis uses the same method steps as above (randomly select three to five deep learning potential functions of cellulose pyrolysis using the same expanded training set). The subsequent methods of cross - validation and force calculation to expand the training set will not be elaborated.

[0087] Iterate through the processes of "molecular dynamics simulation using the deep learning potential function of cellulose pyrolysis", "selecting molecular structures through cross - validation", "force calculation to expand the training set", and "training of new deep learning potential functions of cellulose pyrolysis" until the training set is sufficient to obtain a deep learning potential function of cellulose pyrolysis that meets the requirements.

[0088] Step 3: Perform molecular dynamics simulation of cellulose pyrolysis based on the deep learning potential function model of cellulose pyrolysis, and use Openbabel software combined with the hidden Markov model to capture the products and predict the product types and structures, such as Figure 4 shown. Step 103 specifically includes:

[0089] Step 301, perform molecular dynamics simulation in the NVT ensemble using the finally obtained deep learning potential function model of cellulose pyrolysis, and record the atomic motion trajectory as the simulation calculation result.

[0090] Specifically, the deep learning potential function for cellulose pyrolysis can be used in molecular dynamics simulations only when the accuracy reaches over 90% (the preset ratio is set to 0%) in Step 2;

[0091] For this molecular dynamics simulation, the simulation box calculation formula in Step 1 needs to be adopted and set according to experimental parameters to conform to the actual cellulose pyrolysis experiment as much as possible;

[0092] This molecular dynamics simulation can be carried out through molecular dynamics simulation software such as GROMACS and LAMMPS that support the deep learning potential function for cellulose pyrolysis, and users can select according to their needs.

[0093] During the simulation process, users need to adopt the scripts provided by us or compile scripts according to their own needs to record the atomic motion trajectories as the simulation calculation results.

[0094] These atomic motion trajectories will be used for subsequent hidden Markov model analysis to obtain the types and structural information of cellulose pyrolysis products.

[0095] Step 302: Use Openbabel software and the hidden Markov model to analyze the atomic motion trajectories obtained in the above Step 301 to obtain the prediction results of the types and structural information of cellulose pyrolysis products.

[0096] Specifically, the above Openbabel software is existing open-source software, which can, based on the molecular information recorded by the molecular dynamics simulation software, obtain the molecular types, structural information of all molecules at each sampling time node according to information such as atomic spacing, bond angle, and dihedral angle, and give the corresponding chemical formula as the product trajectory obtained from the molecular dynamics simulation.

[0097] Specifically, the hidden Markov model can achieve the purpose of noise reduction as a filter in the product trajectories obtained from complex molecular dynamics simulations by setting the second-order transition probability matrix T and the second-order output probability matrix O, and finally obtain information such as the types of main pyrolysis product molecules and the reaction process;

[0098] Second-order transition probability matrix:

[0099] T = P(X t |X t-1 )

[0100] Describes the frequency of transitions between different products in the product trajectory sequence obtained from the molecular dynamics simulation, where X t is the state of the Markov chain X at time t; this second-order output probability matrix is:

[0101] O = P(E t |X t )

[0102] Describes the frequency of specific observations generated for a given X t value to produce E t where E t is a time series. Since the pyrolysis products of cellulose are complex and there are many intermediate products that only exist at the f s time level, and these intermediate products are the background noise of the obtained atomic motion trajectories and need to be eliminated. The user can set the above two matrix parameter information according to the needs, change the T matrix that records the transition frequency between products and the O matrix for result output, so as to achieve the purpose of filtering out unnecessary reaction information. The noise reduction principle and specific implementation method of the hidden Markov model are already described in patent documents CN113761814A and CN113950113A, and will not be elaborated here.

[0103] Specifically, the molecular dynamics simulation trajectory can support being imported into molecular trajectory visualization software such as OVITO and VMD to observe the structural characteristics of the obtained products, and the user can select according to the needs.

[0104] Example 2

[0105] The embodiment of the present invention also provides a system for predicting the pyrolysis reaction products of cellulose, characterized in that the system includes:

[0106] A molecular structure construction and volume parameter calculation module, used to construct the cellulose molecular structure and calculate the volume parameters of the simulation box;

[0107] An energy minimization and structure relaxation processing module, used to perform energy minimization processing and structure relaxation processing on the cellulose molecular structure and the simulation box, and optimize the initial cellulose molecular configuration after atomic spacing, bond angle and dihedral angle;

[0108] An ab initio molecular dynamics simulation module, used to perform ab initio molecular dynamics simulation on the initial cellulose molecular configuration for a preset time period, obtain multiple simulated cellulose molecular configurations, and the atomic forces of each cellulose molecular configuration, and construct an initial training set;

[0109] A model training module, used to obtain the deep learning potential function of cellulose pyrolysis by using an iterative method of parallel training, cross-validation, and expansion of the training set based on the initial training set using a multi-layer neural network model;

[0110] A deep learning potential function simulation module for cellulose pyrolysis, used to perform molecular dynamics simulation using the deep learning potential function of cellulose pyrolysis to obtain the types and structural information of cellulose pyrolysis products.

[0111] Among them, the formula for calculating the volume parameters of the simulation box is:

[0112]

[0113]

[0114] Among them, V represents the volume of the pyrolysis chamber, m represents the mass of the cellulose sample, mW represents the relative molecular mass of the cellulose sample; ρ1 represents the molecular density, Vs represents the volume parameter of the simulation box, and N A represents Avogadro's constant.

[0115] In summary, compared with the existing cellulose pyrolysis prediction methods, the present invention does not need to obtain data through repetitive pyrolysis experiments, can achieve an accuracy close to that of first-principles calculations while ensuring the calculation speed, has a simple implementation process, and the deep learning potential function model of cellulose pyrolysis can be extended and applied to different pyrolysis conditions and cellulose configurations. Therefore, it can greatly reduce the experimental cost and improve the efficiency of obtaining the desired products in cellulose pyrolysis.

[0116] Convenient construction of cellulose molecular model: The model can be constructed through the Cellulose Builder described in step 101, without the user manually setting atomic parameters to construct the model through software such as VESTA or Gaussian-view.

[0117] High accuracy and fast speed in molecular dynamics prediction: By introducing the deep learning method described in step 2, a deep learning potential function of cellulose pyrolysis close to the accuracy of first-principles calculations is constructed for molecular dynamics simulation, and the calculation speed is ensured to be close to that of empirical potential fields or force fields.

[0118] Simple implementation process and high degree of freedom: All the software described in the present invention can be freely selected and downloaded and installed on the computer by the user to realize functions such as obtaining cellulose molecular structure data, calculating and analyzing pyrolysis mechanisms, predicting product distributions, and visualizing structures, without the need to conduct repetitive and costly pyrolysis experiments for trial-and-error exploration.

[0119] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other.

[0120] Specific examples are used in this article to elaborate on the principle and implementation manner of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, based on the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for predicting the pyrolysis reaction products of cellulose, characterized in that, The method includes the following steps: Construct a cellulose molecular structure and calculate the volume parameters of the simulation box; Perform energy minimization and structural relaxation on the cellulose molecular structure and the simulation box to obtain an initial cellulose molecular configuration with optimized atomic distances, bond angles, and dihedral angles; Perform ab initio molecular dynamics simulation on the initial cellulose molecular configuration for a preset period of time to obtain multiple simulated cellulose molecular configurations and the atomic forces of each cellulose molecular configuration, and construct an initial training set; Based on the initial training set, use an iterative method of parallel training, cross-validation, and expansion of the training set with a multi-layer neural network model to obtain a deep learning potential function for cellulose pyrolysis, specifically including: S1. Set the value of the initial iteration number k to 1; S2. Set the initial training set as the training set for the k-th iteration; S3. Select a preset number of training subsets from the training set for the k-th iteration; S4. Input the cellulose molecular configurations in each training subset into the multi-layer neural network model corresponding to each training subset one by one to obtain the system energy and atomic forces of each cellulose molecular configuration; S5. Calculate the loss function value of the multi-layer neural network model corresponding to each training subset according to the system energy and atomic forces of each cellulose molecular configuration respectively; S6. When the loss function value is not less than the loss function threshold, update the parameters of the multi-layer neural network model corresponding to each training subset, and return to S4; S7. When the loss function value is less than the loss function threshold, output the multi-layer neural network model corresponding to each training subset as the deep learning potential function for cellulose pyrolysis for the k-th iteration corresponding to each training subset; S8. Calculate the model deviation of each atom of each cellulose molecular configuration according to the atomic forces of each cellulose molecular configuration in each training subset, and obtain the maximum model deviation of each cellulose molecular configuration in each training subset; the maximum model deviation is the maximum value of the model deviations of all atoms of the cellulose molecular configuration; S9. Select the cellulose molecular configurations and the atomic forces of the cellulose molecular configurations at the moments when the maximum model deviation is within the preset range as training samples to expand the training set for the k-th iteration to obtain the training set for the (k + 1)-th iteration, increase the value of k by 1, and return to S3 until the maximum model deviations of the cellulose molecular configurations at different moments of the pre-proportion are less than the deviation threshold; Use the deep learning potential function for cellulose pyrolysis to perform molecular dynamics simulation to obtain the types and structural information of cellulose pyrolysis products.

2. The method for predicting the pyrolysis reaction products of cellulose according to claim 1, wherein The formula for calculating the volume parameters of the simulation box is: Among them, V represents the volume of the pyrolysis chamber, m represents the mass of the cellulose sample, mW represents the relative molecular mass of the cellulose sample; ρ1 represents the molecular density, Vs represents the volume parameter of the simulation box, and N A represents Avogadro's constant.

3. The method for predicting the pyrolysis reaction products of cellulose according to claim 1, wherein The performing energy minimization and structural relaxation on the cellulose molecular structure and the simulation box to obtain an initial cellulose molecular configuration with optimized atomic distances, bond angles, and dihedral angles specifically includes: Perform energy minimization on the cellulose molecular structure in the simulation box using one or more of the stochastic gradient descent algorithm, the steepest gradient descent algorithm, and the conjugate gradient descent algorithm to obtain the cellulose molecular configuration after energy minimization; Perform structural relaxation on the cellulose molecular configuration after energy minimization to obtain the initial cellulose molecular configuration with optimized atomic distances, bond angles, and dihedral angles after structural relaxation.

4. The method for predicting cellulose pyrolysis reaction products according to claim 1, characterized in that: The calculation formula for the value of the loss function is: Among them, L(p ε , p f ) represents the loss function value, N is the number of atoms in the cellulose molecule, and ΔE is the root mean square error of the system energy corresponding to each training subset; ΔF i is the root mean square error of the atomic force on the i-th atom of the cellulose molecule configuration in each training subset, and p ε and p f respectively represent the pre-exponential factors related to time for the energy term and the force term.

5. The method for predicting the pyrolysis reaction products of cellulose according to claim 1, wherein The calculation formula for the model deviation is: in, is the average force on the i-th atom calculated based on the output of the deep learning potential function for cellulose pyrolysis, ε is the maximum standard deviation of the atomic force, and d is a constant.

6. The method for predicting the pyrolysis reaction products of cellulose according to claim 1, characterized in that, Use the deep learning potential function for cellulose pyrolysis to perform molecular dynamics simulations to obtain the types and structural information of cellulose pyrolysis products, specifically including: Use the deep learning potential function for cellulose pyrolysis to perform molecular dynamics simulations to obtain the atomic motion trajectory of the cellulose molecular configuration varying with the molecular dynamics simulation step t; Use the open-source Openbabel software to capture the molecular configurations and chemical formulas of the obtained products from the atomic motion trajectory; Based on the molecular configurations and chemical formulas of the obtained products, obtain the types, structures, and quantities of cellulose pyrolysis products.

7. A system for predicting the products of cellulose pyrolysis reaction, characterized in that, The system includes: A molecular structure construction and volume parameter calculation module for constructing the cellulose molecular structure and calculating the volume parameters of the simulation box; An energy minimization and structural relaxation processing module for performing energy minimization and structural relaxation on the cellulose molecular structure and the simulation box to obtain the initial cellulose molecular configuration with optimized atomic distances, bond angles, and dihedral angles; An ab initio molecular dynamics simulation module for performing ab initio molecular dynamics simulations on the initial cellulose molecular configuration for a preset time period to obtain multiple simulated cellulose molecular configurations and the atomic forces of each cellulose molecular configuration, and constructing an initial training set; A model training module for obtaining the deep learning potential function for cellulose pyrolysis based on the initial training set using an iterative method of parallel training, cross-validation, and expansion of the training set with a multi-layer neural network model, specifically including: S1. Set the value of the initial iteration number k to 1; S2. Set the initial training set as the training set for the k-th iteration; S3. Select a preset number of training subsets from the training set for the k-th iteration; S4. Input the cellulose molecular configurations in each training subset into the multi-layer neural network model corresponding to each training subset one by one to obtain the system energy and atomic forces of each cellulose molecular configuration; S5. Calculate the loss function values of the multi-layer neural network models corresponding to each training subset respectively according to the system energy and atomic forces of each cellulose molecular configuration; S6. When the value of the loss function is not less than the loss function threshold, update the parameters of the multi-layer neural network model corresponding to each training subset, and return to S4; S7. When the value of the loss function is less than the loss function threshold, output the multi-layer neural network model corresponding to each training subset as the deep learning potential function for cellulose pyrolysis for the k-th iteration corresponding to each training subset; S8. Calculate the model deviation of each atom of each cellulose molecular configuration according to the atomic forces of each cellulose molecular configuration in each training subset, and obtain the maximum model deviation of each cellulose molecular configuration in each training subset; the maximum model deviation is the maximum value of the model deviations of all atoms of the cellulose molecular configuration. S9. Select the cellulose molecular configuration and the atomic forces of the cellulose molecular configuration at the moment when the maximum model deviation is within the preset range as training samples, expand the training set of the k-th iteration to obtain the training set of the (k + 1)-th iteration, increase the value of k by 1, return to S3, and continue until the maximum model deviation of the cellulose molecular configurations at a pre-specified proportion of different moments is less than the deviation threshold. Cellulose pyrolysis deep learning potential function simulation module, which is used to perform molecular dynamics simulation using the cellulose pyrolysis deep learning potential function to obtain the types and structural information of cellulose pyrolysis products.

8. The system for predicting the pyrolysis reaction products of cellulose according to claim 7, characterized in that, The formula for calculating the volume parameter of the simulation box is: Among them, V represents the volume of the pyrolysis chamber, m represents the mass of the cellulose sample, mW represents the relative molecular mass of the cellulose sample; ρ1 represents the molecular density, Vs represents the volume parameter of the simulation box, and N A represents Avogadro's constant.

Images