Molecular potential energy surface prediction method based on deep learning
By using a deep learning-based molecular potential surface prediction method, combined with machine learning potential models and the GeomeTRIC algorithm, the problem of insufficient computational efficiency and accuracy of traditional methods in high-throughput screening is solved, enabling rapid and accurate optimization and batch processing of molecular structures.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN UNIV OF TECH
- Filing Date
- 2026-02-25
- Publication Date
- 2026-05-19
AI Technical Summary
In drug discovery, materials science, and chemical product design, existing technologies often struggle to balance computational efficiency and prediction accuracy in high-throughput screening. Traditional quantum chemical calculations are time-consuming, and molecular force field methods lack sufficient prediction accuracy.
A deep learning-based molecular potential surface prediction method, combined with machine learning potential models and the GeomeTRIC algorithm, is employed to achieve rapid optimization and accurate prediction of molecular structures.
It achieves efficient and accurate prediction of molecular energy and force matrices, simplifies the high-throughput calculation process, and improves computational efficiency and accuracy.
Smart Images
Figure CN122067657A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computational chemistry, artificial intelligence, and drug design technology, and specifically to a method for predicting molecular potential energy surfaces based on deep learning. Background Technology
[0002] In fields such as drug discovery, materials science, and chemical product design, obtaining accurate molecular geometry is a prerequisite for subsequent property calculations and functional predictions. While traditional quantum chemical calculation methods (such as density functional theory) offer high accuracy, their computational time increases exponentially with the size of the molecular system, making them unsuitable for high-throughput screening. On the other hand, traditional molecular force field methods, although computationally fast, often fail to fully capture the complex interatomic interactions within molecules, resulting in insufficient prediction accuracy when dealing with complex small organic molecules. Therefore, developing a molecular structure optimization tool that balances computational efficiency and prediction accuracy is an urgent need in the current technological field. Summary of the Invention
[0003] The purpose of this invention is to provide a method for predicting molecular potential energy surfaces based on deep learning. This method combines the high precision of quantum chemistry with the high efficiency of deep learning, and uses machine learning potential models to replace expensive quantum chemical calculations, thereby achieving rapid prediction of the potential energy surfaces of small organic molecules and accurate optimization of their geometric structures.
[0004] The technical solution of the present invention: A method for predicting molecular potential energy surfaces based on deep learning includes the following steps: Preparation phase: Prepare the xyz file for the unoptimized molecule.
[0005] Prediction Phase: The machine learning potential model reads the xyz file and predicts the energy and forces of the current molecular structure. This invention, through deep learning technology, can fully capture and utilize the complex interatomic interactions within the molecule.
[0006] Optimization loop phase: The GeomeTRIC algorithm is used to determine whether the molecular structure has converged based on the predicted energy and force. If the molecular structure has not converged, the molecular coordinates are adjusted to generate a new structure, and the energy and force of the new structure are predicted again using machine learning potential. This process is repeated until the structure converges.
[0007] Output stage: Generates a multi-frame xyz file (_optim.xyz) containing the 3D coordinates of each step of the optimization process and a log file (.log) recording energy and gradient changes.
[0008] Preferably, the present invention also includes a batch processing function, which can automatically traverse all molecular files in a specified directory for optimization.
[0009] Preferably, the present invention also provides an interface with quantum chemical calculation software (such as Gaussian), which can automatically convert the optimized equilibrium conformation into a quantum chemical calculation input file, facilitating subsequent property calculations.
[0010] The beneficial effects of this invention are: Accurate and efficient: By combining deep learning models, accurate and efficient prediction of molecular energy and force matrices is achieved, which significantly improves computational efficiency compared with traditional quantum chemistry methods.
[0011] Rapid optimization: It can quickly and accurately optimize the molecular geometry.
[0012] High degree of automation: It supports batch optimization and automatic generation of subsequent calculation files, which greatly simplifies the high-throughput computing process. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of the directory structure of the unoptimized molecular xyz file in the embodiment.
[0014] Figure 2 This is a schematic diagram illustrating the operation of running the main program to optimize a single molecule in the embodiment.
[0015] Figure 3 This is a schematic diagram of the optimized output file directory in the embodiment.
[0016] Figure 4 This is a schematic diagram of the optimized molecular equilibrium structure file extracted in the example.
[0017] Figure 5 This is a schematic diagram illustrating the batch molecular structure optimization operation in the embodiment.
[0018] Figure 6 This is a schematic diagram of the output file directory after batch optimization in the embodiment.
[0019] Figure 7 This is a schematic diagram of the directory of optimized molecular structure files extracted in batches in the embodiment.
[0020] Figure 8 This is a schematic diagram of the contents of the quantum chemical calculation parameter configuration file in the embodiment.
[0021] Figure 9 This is a schematic diagram of the quantum chemical calculation input file directory generated in the embodiment.
[0022] Figure 10 This is a schematic diagram of the directory of quantum chemical calculation input files generated in batches in the embodiment.
[0023] Figure 11 This is a schematic diagram of the output file directory for quantum chemical property calculations in the embodiment.
[0024] Figure 12 This is a flowchart of the method of the present invention. Detailed Implementation
[0025] The present invention will be further described below with reference to specific embodiments and accompanying drawings.
[0026] A method for predicting molecular potential energy surfaces based on deep learning, characterized by the following steps: Step 1: Obtain the molecular geometry file to be optimized, which contains the initial three-dimensional coordinate information of the molecule; Step 2: Use a pre-trained machine learning potential model to read the current molecular geometry and predict the molecular energy matrix and force matrix under that structure; Step 3: Using the geometry optimization algorithm (GeomeTRIC), determine whether the current molecular structure satisfies the convergence condition based on the energy and force predicted in Step 2; Step 4: If the molecular structure does not meet the convergence condition, the molecular coordinates are adjusted using the geometric optimization algorithm to generate a new molecular structure, and the new structure is returned to step 2. The prediction and judgment process is repeated until the molecular structure meets the convergence condition. Step 5: When the molecular structure meets the convergence condition, output the optimized molecular equilibrium conformation file and the log file recording the optimization process.
[0027] Step 5 is followed by a step of extracting the equilibrium conformation: Extract the last frame of data from the multi-frame trajectory file (.xyz) generated during the optimization process and save it as an independent molecular equilibrium conformation file.
[0028] The method supports batch processing mode: By reading a directory containing multiple unoptimized molecular geometry files, steps 1 to 5 are executed sequentially for each molecular structure file in the directory to achieve batch optimization of molecular structures.
[0029] The method also includes a step of generating a quantum chemical computing interface: The system reads a preset calculation parameter configuration file, which includes parameters such as the number of cores, memory size, functional basis set, charge number, and spin multiplicity. Based on the calculation parameter configuration file and the extracted molecular equilibrium conformation file, it automatically generates the input file (.gjf) required by the quantum chemistry calculation software.
[0030] The calculation parameter configuration file (Gau_input.ini) contains a batch conversion flag; when the batch conversion flag is detected to be enabled, all optimized molecular equilibrium conformation files in the specified directory are automatically converted into corresponding quantum chemical calculation input files.
[0031] A deep learning-based molecular potential surface prediction device includes: The input module is used to receive unoptimized molecular XYZ coordinate files; The prediction module is used to run machine learning potential models, capture the interactions between atoms within the molecule, and predict the energy and forces of the current conformation. The optimization module is used to run the GeomeTRIC algorithm, determine convergence based on the prediction results, and adjust the molecular coordinates. The output module is used to output optimized multi-frame trajectory files and log files; The interface module is used to convert the optimized structure into the standard input format of quantum chemistry calculation software.
[0032] Example 1: Prediction of Potential Energy Surface and Geometric Optimization of Single Molecules This embodiment demonstrates how to use this method to optimize the structure of a single molecule.
[0033] Environment setup: On a Windows system, install Anaconda and configure the Python 3.7 environment, and install the required Python libraries.
[0034] File preparation: Place the unoptimized molecular structure file (e.g., example-1.xyz) in the DeepEmpest / example / UnOpt_xyz directory, such as... Figure 1 As shown.
[0035] To run the program: Activate the environment in the PowerShell command prompt window and enter the command `python opt.py --fname example-1.xyz`. The program will call a machine learning potential model to predict energy and force, and combine it with the GeomeTRIC algorithm for iterative optimization, such as... Figure 2 As shown.
[0036] Results Analysis: After the program finishes running, example-1.log (recording energy, gradient, and time) and example-1_optim.xyz (recording the optimization trajectory) are generated in the original directory, as follows: Figure 3 As shown.
[0037] Extracting the conformation: Run the command `python saveLastFrame.py --fname example-1_optim.xyz`. The program will extract the last frame (equilibrium conformation) of the trajectory file and save it to the `DeePEST / example / Opt_xyz` directory. Figure 4 As shown.
[0038] Example 2: Batch Molecular Structure Optimization This example demonstrates how to perform high-throughput optimization on multiple molecules.
[0039] Batch execution: In the configured environment, enter the command `python opt.py --all yes`. The program will automatically recognize all xyz files in the UnOpt_xyz directory (e.g., ...). Figure 1 (examples 1 to 5), and perform structural optimization in sequence, such as Figure 5 As shown.
[0040] Results output: Log files and optimized trajectory files for all molecules are generated in the same directory, such as... Figure 6 As shown.
[0041] Batch extraction: Enter the command `python saveLastFrame.py --all yes`, and the program will automatically extract the equilibrium conformations of all molecules in batches and save them to the `Opt_xyz` directory, such as... Figure 7 As shown.
[0042] Example 3: Application of Quantum Chemistry Computation Interface This example demonstrates how to perform high-throughput optimization on multiple molecules.
[0043] Parameter configuration: Edit the Gau_input.ini file in the DeePEST / example / QC-based directory. Set the [Parameters] (e.g., number of cores nproc=8, memory mem=8GB, functional basis set b3lyp 6-31g(d,p), em=GD3BJ, etc.) and [Task] (e.g., force). Figure 8 As shown.
[0044] Generate input files: Set all=yes in the configuration file and run the command python xyz2GauInput.py. The program will automatically read all optimized structures in the Opt_xyz directory and generate Gaussian input files (.gjf) in batches according to the configuration parameters, saving them to the gjf directory, such as... Figure 10 As shown.
[0045] Property Calculation: Running the generated batch file (all.bat) will invoke the Gaussian program to perform batch calculations, ultimately generating an .out file containing quantum chemical properties, such as... Figure 11 As shown. The final result can be viewed through graphical interface software (such as GaussView).
Claims
1. A method for predicting molecular potential energy surfaces based on deep learning, characterized in that, Includes the following steps: Step 1: Obtain the molecular geometry file to be optimized, which contains the initial three-dimensional coordinate information of the molecule; Step 2: Use a pre-trained machine learning potential model to read the current molecular geometry and predict the molecular energy matrix and force matrix under that structure; Step 3: Using a geometric optimization algorithm, determine whether the current molecular structure satisfies the convergence condition based on the energy and force predicted in Step 2; Step 4: If the molecular structure does not meet the convergence condition, the molecular coordinates are adjusted using the geometric optimization algorithm to generate a new molecular structure, and the new structure is returned to step 2. The prediction and judgment process is repeated until the molecular structure meets the convergence condition. Step 5: When the molecular structure meets the convergence condition, output the optimized molecular equilibrium conformation file and the log file recording the optimization process.
2. The method according to claim 1, characterized in that, Step 5 is followed by a step of extracting the equilibrium conformation: Extract the last frame of data from the multi-frame trajectory file generated during the optimization process and save it as an independent molecular equilibrium conformation file.
3. The method according to claim 1, characterized in that, The method supports batch processing mode: By reading a directory containing multiple unoptimized molecular geometry files, steps 1 to 5 are executed sequentially for each molecular structure file in the directory to achieve batch optimization of molecular structures.
4. The method according to claim 2, characterized in that, The method also includes a step of generating a quantum chemical computing interface: The system reads a preset calculation parameter configuration file, which includes parameters such as the number of cores, memory size, functional basis set, charge number, and spin multiplicity. Based on the calculation parameter configuration file and the extracted molecular equilibrium conformation file, the system automatically generates the input files required by the quantum chemistry calculation software.
5. The method according to claim 4, characterized in that, The calculation parameter configuration file contains a batch conversion flag; when the batch conversion flag is detected to be enabled, all optimized molecular equilibrium conformation files in the specified directory are automatically converted into corresponding quantum chemical calculation input files.
6. A molecular potential energy surface prediction device based on deep learning, characterized in that, include: The input module is used to receive unoptimized molecular XYZ coordinate files; The prediction module is used to run machine learning potential models, capture the interactions between atoms within the molecule, and predict the energy and forces of the current conformation. The optimization module is used to run the GeomeTRIC algorithm, determine convergence based on the prediction results, and adjust the molecular coordinates. The output module is used to output optimized multi-frame trajectory files and log files; The interface module is used to convert the optimized structure into the standard input format of quantum chemistry calculation software.