Design apparatus, design method, program, and design system
The use of a deep learning force field trained on first-principles calculations addresses the inefficiencies of existing force fields by enabling efficient and accurate design of classical force fields, facilitating simulations of complex systems.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- RESONAC CORP
- Filing Date
- 2024-05-14
- Publication Date
- 2026-06-23
AI Technical Summary
Existing force fields for simulating chemical reactions and chemisorption face challenges in balancing computational cost, learning cost, and versatility, with first-principles calculations being too costly, classical force fields requiring high learning costs, and deep learning force fields having lower computational efficiency.
A design device and method using a deep learning force field trained on first-principles calculations to infer physical properties, determining parameters for a classical force field function system, thereby reducing computational costs and improving versatility.
The solution enables efficient design of classical force fields with reduced computational costs and improved accuracy, allowing simulations of larger systems like polymers or proteins with millions of atoms.
Smart Images
Figure 0007878577000001 
Figure 0007878577000002 
Figure 0007878577000003
Abstract
Description
Technical Field
[0001] The present disclosure relates to a design device, a design method, a program, and a design system.
Background Art
[0002] With the recent development of computing technology, various physical property predictions such as chemical reactions or chemisorption are performed using various force fields (sets of function systems and parameters for simulation). Although force fields have various characteristics including the chemical species to be targeted or their effects, all are specialized for a specific system, and thus, rapid development of a generally applicable force field is required.
[0003] For example, Patent Document 1 describes NNP (Neural Network Potential) as a method for realizing an inference based on results obtained by first-principles calculations.
Prior Art Documents
Patent Documents
[0004]
Patent Document 1
Summary of the Invention
Problems to be Solved by the Invention
[0005] For the analysis of the behavior of chemical reactions or chemisorption, for example, first-principles calculations or quantum chemical calculations are used. First-principles calculations or quantum chemical calculations have high accuracy and generality, but have a problem that the calculation cost is very high. For example, in the case of first-principles calculations or quantum chemical calculations, calculations of about 100 atoms may be a practical limit.
[0006] Furthermore, simulations using classical force fields, which are an example of force fields, have very low computational costs and can be used to analyze polymers or proteins with around one million atoms. However, classical force fields have the problem of high learning costs and low versatility because they require repeated first-principles calculations to determine the parameters of a function system specific to the target system.
[0007] Furthermore, simulations using deep learning force fields, which are an example of force fields, reproduce first-principles calculation results with high accuracy and versatility through deep learning. Trained deep learning force fields possess high accuracy and versatility. However, deep learning force fields have the problem of lower computational cost or memory utilization compared to classical force fields.
[0008] This disclosure aims to provide design devices, design methods, programs, and design systems that support the design of classical force fields. [Means for solving the problem]
[0009] This disclosure comprises the following configuration.
[0010] [1] A design device for designing classical force fields, An inference means for inferring the physical properties of a chemical structure necessary for creating a classical force field, using a deep learning force field that has learned the results of first-principles calculations of an arbitrary chemical structure, A determination means for determining the parameters of a classical force field function system using the inferred physical properties of the chemical structure, A design device equipped with the following features.
[0011] [2] The determination means fits the inferred physical properties of the chemical structure to the function system of the classical force field and determines the parameters. [1] The design apparatus described above.
[0012] [3] The deep learning force field is a neural network used in NNP (Neural Network Potential) and has been trained to output the first-principles calculation result of the chemical structure when the chemical structure is input. The design apparatus described in [1] or [2].
[0013] [4] The physical properties of the chemical structure are the energy obtained when the bond length, bond angle, and dihedral angle are changed. A design apparatus as described in any one of items [1] to [3].
[0014] [5] A design method performed by a design device for designing classical force fields, An inference procedure for inferring the physical properties of a chemical structure necessary for creating a classical force field, using a deep learning force field that has learned the results of first-principles calculations for an arbitrary chemical structure, A determination procedure for determining the parameters of a classical force field function system using the inferred physical properties of the aforementioned chemical structure, A design method that includes the following features.
[0015] [6] A design device for designing classical force fields, An inference step in which the physical properties of the chemical structure necessary for creating a classical force field are inferred using a deep learning force field that has learned the results of first-principles calculations for an arbitrary chemical structure. A decision step in which the parameters of the classical force field function system are determined using the inferred physical properties of the chemical structure, A program to execute.
[0016] [7] A design system for designing classical force fields, A first-principles calculation means for performing first-principles calculations, A learning means for training a deep learning force field using the first-principles calculation results of an arbitrary chemical structure performed by the aforementioned first-principles calculation means, An inference means for inferring the physical properties of the chemical structure necessary for creating a classical force field using a deep learning force field that has learned the results of the first-principles calculations, A determination means for determining the parameters of a classical force field function system using the inferred physical properties of the chemical structure, A design system comprising
Advantages of the Invention
[0017] According to the present disclosure, a design apparatus, a design method, a program, and a design system for assisting in the design of a classical force field can be provided.
Brief Description of the Drawings
[0018] [Figure 1] It is a configuration diagram of an example of the design system according to the present embodiment. [Figure 2] It is a hardware configuration diagram of an example of the computer according to the present embodiment. [Figure 3] It is a functional configuration diagram of an example of the design system according to the present embodiment. [Figure 4A] It is a diagram for explaining an example of a classical force field creation method. [Figure 4B] It is a diagram for explaining an example of a classical force field creation method. [Figure 4C] It is a diagram for explaining an example of a classical force field creation method. [Figure 5] It is a flowchart of an example of the process for creating a deep learning force field according to the present embodiment. [Figure 6] It is a flowchart of an example of the process for creating a classical force field according to the present embodiment.
Modes for Carrying Out the Invention
[0019] Next, embodiments of the present invention will be described in detail. Note that the present invention is not limited to the following embodiments.
[0020] [First Embodiment] <System Configuration> Figure 1 is a configuration diagram of an example of a design system 1 according to this embodiment. The design system 1 in Figure 1 includes a design device 10 and a user terminal 12. The design device 10 and the user terminal 12 of the design system 1 are connected via a communication network 18 such as a local area network (LAN) or the internet, enabling data communication.
[0021] The user terminal 12 is an information processing terminal operated by the worker, such as a PC, tablet, or smartphone. The user terminal 12 displays a screen on its display device that accepts information input from the worker and accepts information input from the worker. The user terminal 12 also transmits the information received from the worker to the design device 10, causing it to execute processing to support the design of classical force fields. The user terminal 12 receives information on the execution results of the processing by the design device 10 and displays it on its display device for the worker to confirm.
[0022] The design device 10 is an information processing device such as a PC or workstation that supports the design of classical force fields by an operator. The design device 10 speeds up the calculation of data sampling necessary for parameter fitting of the function system of the classical force field by using a deep learning force field that has learned from the results of first-principles calculations, rather than directly performing the calculation using first-principles calculations. The design device 10 receives information input by the operator into the user terminal 12 and executes processing to support the design of the classical force field. The design device 10 sends the processing result information to the user terminal 12 and displays the processing result information on the user terminal 12.
[0023] The design system 1 in Figure 1 may be implemented using a design device 10 with web server functionality and a user terminal 12 that runs a web application using a web browser. Alternatively, the design system 1 in Figure 1 may be implemented by having an application installed on the user terminal 12 cooperate with a program installed on the design device 10 to perform processing.
[0024] It should be noted that design system 1 in Figure 1 is merely an example, and there are various system configurations depending on the application and purpose. For example, the design device 10 may be implemented using multiple computers, or it may be implemented as a cloud computing service. Design system 1 may be implemented using a standalone computer.
[0025] <Hardware Configuration> The design apparatus 10 and user terminal 12 in Figure 1 are implemented, for example, by a computer 500 with the hardware configuration shown in Figure 2.
[0026] Figure 2 is a hardware configuration diagram of an example of a computer 500 according to this embodiment. The computer 500 in Figure 2 is equipped with an input device 501, a display device 502, an external interface 503, RAM 504, ROM 505, a CPU 506, a communication interface 507, and an HDD 508, and each is interconnected via bus B. Note that the input device 501 and the display device 502 may be used in a connected configuration.
[0027] The input device 501 includes a touch panel, operation keys and buttons, a keyboard and mouse, etc., used by the operator to input various signals. The display device 502 consists of a display such as a liquid crystal or organic EL that displays the screen, and a speaker that outputs sound data such as voice and sound. The communication interface 507 is an interface for the computer 500 to perform data communication.
[0028] Furthermore, HDD508 is an example of a non-volatile storage device that stores programs and data. The programs and data stored include the OS, which is the basic software that controls the entire computer 500, and applications that provide various functions on the OS. Note that computer 500 may use a drive device that uses flash memory as a storage medium (for example, a solid-state drive: SSD) instead of HDD508.
[0029] External I / F 503 is an interface to external devices. External devices include recording media 503a, etc. This allows computer 500 to read and / or write to recording media 503a via external I / F 503. Recording media 503a include flexible disks, CDs, DVDs, SD memory cards, USB memory, etc.
[0030] ROM505 is an example of non-volatile semiconductor memory (storage device) that can retain programs and data even when the power is turned off. ROM505 stores programs and data such as the BIOS, OS settings, and network settings that are executed when the computer 500 starts up. RAM504 is an example of volatile semiconductor memory (storage device) that temporarily holds programs and data.
[0031] The CPU 506 is an arithmetic unit that controls and implements the functions of the entire computer 500 by reading programs and data from storage devices such as the ROM 505 and HDD 508 onto the RAM 504 and executing processing. In this embodiment, the computer 500 can implement various functions of the design device 10 and user terminal 12 described later by executing programs.
[0032] The program to be installed in RAM 504 is installed, for example, by reading a program recorded on the recording medium 503a via the external I / F 503. Alternatively, the program may be installed by downloading it from the communication network 18 via the communication I / F 507.
[0033] <Functional Configuration> The functional configuration of the design system 1 according to this embodiment will now be described. Figure 3 is a functional configuration diagram of an example of the design system 1 according to this embodiment. Note that parts of the configuration diagram in Figure 3 that are not necessary for the explanation of this embodiment have been omitted as appropriate.
[0034] The design device 10 of the design system 1 shown in Figure 3 includes a request receiving unit 20, a response transmission unit 22, a calculation system creation unit 24, a first-principles calculation unit 26, a learning unit 28, an inference unit 30, a decision unit 32, a verification unit 34, a display control unit 36, a deep learning force field memory unit 40, and a classical force field memory unit 42. The user terminal 12 of the design system 1 includes an information display unit 50, an operation reception unit 52, a request transmission unit 54, and a response receiving unit 56.
[0035] The information display unit 50 displays information on the display device 502, including a screen for receiving information input from the operator and information on the execution results of the design device 10's processing. The operation reception unit 52 receives operations from the operator, such as information input. The request transmission unit 54 transmits processing requests to the design device 10 in response to the information input from the operator. The response reception unit 56 receives responses from the design device 10 to the processing requests transmitted by the request transmission unit 54.
[0036] The request receiving unit 20 receives a processing request from the user terminal 12. The response transmission unit 22 responds with the result of the processing performed in accordance with the processing request. The calculation system creation unit 24 creates a calculation system using an arbitrary chemical structure based on information about the arbitrary chemical structure input by the operator. The first-principles calculation unit 26 performs first-principles calculations on structural models in which the bond length, bond angle, and dihedral angle of the arbitrary chemical structure are varied, and calculates the energy of the structural model as a first-principles calculation result. Bond length, bond angle, and dihedral angle are examples of physical properties of a chemical structure. The first-principles calculation is based on, for example, DFT (Density Function Theory) or QM (Quantum Mechanics).
[0037] The learning unit 28 trains the deep learning force field on the relationship between structural models obtained by varying the bond length, bond angle, and dihedral angle of an arbitrary chemical structure, and the energy of the structural model calculated as a result of first-principles calculations, based on an appropriate machine learning method. Thus, the machine learning of the deep learning force field may be performed using an arbitrary chemical structure and the physical properties of that chemical structure calculated by first-principles calculations for that arbitrary chemical structure as training data.
[0038] Furthermore, the deep learning force field is a neural network used in NNP (Neural Network Potential), and is trained to output the first-principles calculation result of a chemical structure when a chemical structure is input. The trained deep learning force field is capable of inferring the first-principles calculation result of the input chemical structure. The learning unit 28 stores the trained deep learning force field in the deep learning force field memory unit 40. The deep learning force field memory unit 40 stores the trained deep learning force field. The deep learning force field memory unit 40 may also store trained deep learning force fields that are publicly available on the internet, etc.
[0039] The inference unit 30 uses the learned deep learning force fields stored in the deep learning force field memory unit 40 to infer the energy of the structural model as a first-principles calculation result for a structural model in which the bond length, bond angle, and dihedral angle of an arbitrary chemical structure are changed.
[0040] The determination unit 32 fits the energy of the structural model inferred by the inference unit 30 to the function system of the classical force field and determines the parameters of the classical force field. The determination unit 32 creates the classical force field by determining the parameters of the function system of the classical force field. Details of the processing of the determination unit 32 will be described later. The classical force field storage unit 42 stores the created classical force field.
[0041] The verification unit 34 verifies whether the parameters of the classical force field function system determined by the determination unit 32 are appropriate. The display control unit 36 controls the display on the user terminal 12. Note that the functional configuration diagram in Figure 3 is an example. The design system 1 according to this embodiment can be realized with various configurations. For example, the deep learning force field memory unit 40 and the classical force field memory unit 42 may be a storage device, computer, or cloud storage that can communicate data with the design device 10.
[0042] <Processing> Figures 4A to 4C illustrate an example of a method for constructing a classical force field. Figure 4A is a schematic diagram of an example of a computational system constructed using an arbitrary chemical structure. In Figure 4A, "l" represents the bond length between two particles. In Figure 4A, "θ" represents the bond angle between three particles. In Figure 4A, "ω" represents the dihedral angle between four particles. Figure 4B is an example of the function system of the classical force field for the chemical structure shown in Figure 4A. The function system of the classical force field has a parameter k b , k a , and V n Includes.
[0043] Figure 4C plots the energy when the bond length l is changed in increments of 0.5 angstroms, for example. The plot is created by calculating data sampling. In creating the classical force field, the parameters of the classical force field function system are determined by individually optimizing the parameters in Figure 4B so that the function system approximates the plot in Figure 4C.
[0044] The individual optimization of the parameters in Figure 4B may be performed by a program that determines the parameters using, for example, the least squares method, so as to minimize the difference with the plot in Figure 4C, or by adjusting the parameters by the operator.
[0045] By the way, if first-principles calculations are performed for the plot in Figure 4C, it is necessary to perform first-principles calculations for each plot. First-principles calculations have high accuracy and versatility, but the computational cost is very high. Therefore, if first-principles calculations are performed for the plot in Figure 4C, the computational cost will be high. In this embodiment, however, a simulation is performed using a deep learning force field that has been trained to produce first-principles calculation results for the plot in Figure 4C. The simulation using the deep learning force field reproduces the first-principles calculation results with high accuracy. The trained deep learning force field has a lower computational cost than first-principles calculations. Therefore, if a simulation using a trained deep learning force field is performed for the plot in Figure 4C, the computational cost will be lower than if first-principles calculations were performed. Furthermore, if a simulation using a trained deep learning force field is performed for the plot in Figure 4C, the memory utilization is better than if first-principles calculations were performed. Better memory utilization can be expected to increase the number of atoms that can be calculated.
[0046] Furthermore, publicly available deep learning force fields can be used; for example, those published by the Open Catalyst Project can be utilized. By using publicly available deep learning force fields, the training cost of the deep learning force fields can be reduced in this embodiment.
[0047] If no publicly available deep learning force field exists, in this embodiment, a deep learning force field is created using the procedure shown in Figure 5, for example. Figure 5 is a flowchart of an example of the process for creating a deep learning force field according to this embodiment.
[0048] In step S10, the calculation system creation unit 24 creates a calculation system using an arbitrary chemical structure, such as the one shown in Figure 4A, based on the information about the arbitrary chemical structure entered by the operator.
[0049] In step S12, the first-principles calculation unit 26 performs first-principles calculations on structural models in which the bond length, bond angle, and dihedral angle of an arbitrary chemical structure are varied, for example, as shown in Figure 4A, and calculates the energy of the structural model as a first-principles calculation result, for example, as plotted in Figure 4C.
[0050] In step S14, the learning unit 28, based on an appropriate machine learning method, trains the deep learning force field on the relationship between structural models in which the bond length, bond angle, and dihedral angle of an arbitrary chemical structure are varied, and the energy of the structural model calculated as a result of first-principles calculations.
[0051] Figure 6 is a flowchart of an example of the process for creating a classical force field according to this embodiment.
[0052] In step S20, the calculation system creation unit 24 creates a calculation system using an arbitrary chemical structure, such as the one shown in Figure 4A, based on the information about the arbitrary chemical structure entered by the operator.
[0053] In step S22, the inference unit 30 uses the learned deep learning force fields stored in the deep learning force field memory unit 40 to infer the energy of a structural model in which the bond length, bond angle, and dihedral angle of an arbitrary chemical structure are changed, and calculates the energy of the structural model, for example, as plotted in Figure 4C.
[0054] In step S24, the determination unit 32 fits the energy of the structural model inferred by the inference unit 30 in step S22 to the function system of the classical force field and determines the parameters of the classical force field. For example, the determination unit 32 determines the parameters of the function system of the classical force field by individually optimizing the parameters in Figure 4B so that the function system approximates the plot in Figure 4C.
[0055] In step S26, the verification unit 34 verifies whether the parameters determined by the determination unit 32 in step S24 are appropriate. For example, the verification unit 34 uses the parameters determined in step S24 in the classical force field function system shown in Figure 4B to perform molecular dynamics calculations and verifies whether the parameters are appropriate based on the results.
[0056] If the parameters determined in step S24 are not appropriate, the design apparatus 10 returns to the process in step S22 and continues processing. If the parameters determined in step S24 are appropriate, the design apparatus 10 terminates the process shown in the flowchart in Figure 6. Note that the processes shown in the flowcharts in Figures 5 and 6 may be performed after separating the arbitrary chemical structure, for example, if the arbitrary chemical structure is a polymer or protein.
[0057] [Other embodiments] The design apparatus 10 according to this embodiment may be configured in which multiple computers work together to perform the above processing. For example, the design apparatus 10 of the design system 1 according to this embodiment may be implemented by dividing it into an apparatus for performing first-principles calculations, an apparatus for creating a deep learning force field, and an apparatus for designing a classical force field.
[0058] As described above, the design system 1 according to this embodiment can speed up the calculation of data sampling necessary for parameter fitting of the function system of a classical force field, and can efficiently determine the parameters of the function system of a classical force field. Therefore, according to this embodiment, a design device, design method, program, and design system can be provided to support the efficient design of classical force fields.
[0059] Although this embodiment has been described above, it will be understood that various modifications to the form and details are possible without departing from the spirit and scope of the claims. Although the present invention has been described above based on examples, the present invention is not limited to the above examples, and various modifications are possible within the scope described in the claims. This application claims priority to Basic Application No. 2023-080836 filed with the Japan Patent Office on 16 May 2023, the entire contents of which are incorporated herein by reference. [Explanation of Symbols]
[0060] 1. Design System 10 Design equipment 12 User terminals 18. Communication Networks 24 Computation System Creation Department 26 First principles calculation section 28 Learning Department 30 Reasoning section 32 Decision Section 34 Verification Department
Claims
1. A design device for designing classical force fields, An inference means for inferring the physical properties of a chemical structure necessary for creating a classical force field, using a deep learning force field that has learned the results of first-principles calculations of an arbitrary chemical structure, A determination means for determining the parameters of a classical force field function system using the inferred physical properties of the chemical structure, Equipped with, The aforementioned deep learning force field is a neural network used in NNP (Neural Network Potential), which has been trained to output the first-principles calculation result of the chemical structure when the chemical structure is input. The physical properties of the aforementioned chemical structure are the energy obtained when the bond length, bond angle, and dihedral angle are changed, respectively, in the design device.
2. The determination means fits the inferred physical properties of the chemical structure to the function system of the classical force field and determines the parameters. The design apparatus according to claim 1.
3. A design method used by a design device for designing classical force fields, An inference procedure for inferring the physical properties of a chemical structure necessary for creating a classical force field, using a deep learning force field that has learned the results of first-principles calculations for an arbitrary chemical structure, A determination procedure for determining the parameters of a classical force field function system using the inferred physical properties of the aforementioned chemical structure, Equipped with, The aforementioned deep learning force field is a neural network used in NNP (Neural Network Potential), which has been trained to output the first-principles calculation result of the chemical structure when the chemical structure is input. The design method for determining the physical properties of the aforementioned chemical structure is the energy obtained when the bond length, bond angle, and dihedral angle are changed.
4. For designing classical force fields, An inference step in which the physical properties of the chemical structure necessary for creating a classical force field are inferred using a deep learning force field that has learned the results of first-principles calculations for an arbitrary chemical structure. A decision step in which the parameters of the classical force field function system are determined using the inferred physical properties of the chemical structure, Make it run, The aforementioned deep learning force field is a neural network used in NNP (Neural Network Potential), which has been trained to output the first-principles calculation result of the chemical structure when the chemical structure is input. The physical properties of the aforementioned chemical structure are programs that represent the energy obtained when the bond length, bond angle, and dihedral angle are changed, respectively.
5. A design system for designing classical force fields, A first-principles calculation means for performing first-principles calculations, A learning means for training a deep learning force field using the first-principles calculation results of an arbitrary chemical structure performed by the aforementioned first-principles calculation means, An inference means for inferring the physical properties of the chemical structure necessary for creating a classical force field using a deep learning force field that has learned the results of the first-principles calculations, A determination means for determining the parameters of a classical force field function system using the inferred physical properties of the chemical structure, Equipped with, The aforementioned deep learning force field is a neural network used in NNP (Neural Network Potential), which has been trained to output the first-principles calculation result of the chemical structure when the chemical structure is input. The aforementioned chemical structure's physical properties are defined as the energy obtained when the bond length, bond angle, and dihedral angle are changed, respectively, in this design system.