Information processing device, information processing method, program, and non-transitory computer-readable medium

Abstract

To provide an information processing device, information processing method, program, and non-transitory computer-readable medium, which allows for speeding up computation by using a trained model that outputs at least one of quantities obtained during computation.SOLUTION: An information processing device is provided, comprising one or more memories and one or more processors. When a three-dimensional arrangement of two or more atoms constituting a molecule is entered, the one or more processors search for a reaction path using one or more trained models designed to output a physical quantity pertaining to the molecule.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an information processing device, an information processing method, a program, and a non-transitory computer-readable medium. [Background technology]

[0002] Reaction path and transition state searches are important for obtaining information such as transition states, activation parameters, and rate constants in intermolecular reaction paths. While it is relatively easy to search for stable structures in reaction path searches, it is generally difficult to search for saddle points that could become reaction paths from these stable structures. Reaction path searches can be performed using, for example, stationary point searches and the Nudged Elastic Band (NEB) method. However, the former requires searching from the vicinity of the saddle point, and the latter requires the acquisition of a rough reaction path as a prerequisite. Methods such as the Anharmonic Downward Distortion Following (ADDF) method and the Artificial Force Induced Reaction (AFIR) method, or GRRM (Registered Trademark) (Global Reaction Route Mapping), which combines these methods, can solve these problems.

[0003] While the above methods allow for automated search for reaction paths, the ADDF and AFIR methods are computationally expensive and are limited in the size of molecules that can be calculated and the number of molecules that can be searched. This is because the energy, force, or Hessian must be calculated using a massive number of first-principles calculations while moving the nuclei. Calculations must also be performed for structures far from stable molecules, making application to classical force fields difficult. In the ADDF method, the direction of molecular deformation is the direction of large downward anharmonic distortion. However, calculating downward anharmonic distortion using first-principles calculations is extremely computationally expensive, so there is a risk of search failure as a trade-off for using a high-speed method. Furthermore, in the AFIR method, it is difficult to determine the artificial force parameters, and it is also difficult to determine whether the reaction has been correctly reproduced after the calculation is complete. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] S. Maeda, et. al., “Systematic Exploration of the Mechanism of Chemical Reactions: Global Reaction Rout Mapping (GRRM) Strategy by the ADDF and AFIR Methods”, Phys. Chem. Chem. Phys., 15, 3683-3701 (2013) [Non-patent document 2] S. Maeda, et. al., “Implementation and performance of the artificial force induced reaction method in the GRRM17 program”, J. Comput. Chem., 39, 233-250(2018) [Non-patent document 3] WMC Sameera, S. Maeda, and K. Morokuma, “Computational Catalysis Using the Artificial Force Induced Reaction Method”, Acc. Chem. Res., 49, 763-773 (2016) Summary of the Invention [Problem to be solved by the invention]

[0005] The present disclosure proposes an information processing device that aims to speed up calculations by using a trained model that outputs at least one of the quantities obtained during calculations. [Means for solving the problem]

[0006] According to one embodiment, the information processing device comprises: one or more memories; one or more processors; Equipped with. The one or more processors: When the three-dimensional arrangement of two or more atoms that make up a molecule is input, reaction paths are explored using one or more trained models that output physical quantities related to the molecule. [Brief explanation of the drawings]

[0007] [Figure 1] 1 is a flowchart illustrating a process according to an embodiment. [Figure 2] 1 is a flowchart illustrating a process according to an embodiment. [Figure 3] 1 is a flowchart illustrating a process according to an embodiment. [Figure 4] 1 is a flowchart illustrating a process according to an embodiment. [Figure 5] 1 is a flowchart illustrating a process according to an embodiment. [Figure 6] 1 is a flowchart illustrating an example of a hardware implementation according to one embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0008] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. The drawings and the description of the embodiments are given by way of example only and are not intended to limit the present invention.

[0009] First, the ADDF method and the AFIR method used for the automatic search in this disclosure will be briefly described, and then each embodiment implementing each method will be described in detail with reference to the drawings.

[0010] (ADDF Act) The ADDF method is a method for determining a transition state by deforming a molecule in a direction in which the potential energy is smaller than the harmonic approximation, i.e., in a direction in which the anharmonic downward distortion (ADD) is minimized. The explanation is provided as an example, and the embodiments of the present disclosure are not limited to these.

[0011] In both first-principles calculations and pre-trained models, the energy E is calculated as a function of the nuclear coordinate x.

number

[0012] Therefore, whether using first-principles calculations or pre-trained models, various optimization methods, such as steepest descent, conjugate gradient, and BFGS (Broyden-Fletcher-Goldfarb-Shanno) algorithms, can be used to search for the structure with the lowest energy. In other words, the nuclear coordinates of the stable structure, x0 = (x01, x02, x03, ..., x0 3N ) T can be explored.

[0013] At x0, the first derivative of the energy with respect to the coordinate, i.e., the force, is 0, so in the vicinity of x0, x = (x1, x2, x3, ..., x 3N ) T The energy at is harmonically approximated as follows:

number

[0014] where Δx = x - x0, H is the Hessian matrix at x0, and T denotes the transpose of a vector or matrix.

number

[0015] Consider the coordinate transformation of Δ x to y using the orthogonal matrix R.

number

[0016] In this case, the harmonic approximated energy can be expressed as:

number

[0017] Since the Hessian matrix H is a symmetric matrix, by choosing R appropriately, RHR T becomes the diagonal matrix D.

number

[0018] Furthermore, z i = D ii 1 / 2 y i and by coordinate transformation, the following equation can be obtained:

number

[0019] On the 3N-dimensional hypersphere thus obtained, the direction in which the actual energy without harmonic approximation is minimal is the direction in which the transition state exists, and the transition state can be reached by deforming the molecule in that direction.

[0020] In addition, the constraint Σz i 2 = r 2 In order to search for the direction of the minimum energy under the condition, various methods can be used. For example, Lagrange's method of non-stationary numbers can be used. Also, in polar coordinates with fixed r, as shown below, 3N-1 may be optimized.

number

[0021] We have explained how to search for the direction of deformation of a molecule starting from a structure with minimum energy, but once the molecule is deformed, it is no longer in the minimum energy state. In this case, the harmonic approximation energy contains a first-order term with respect to Δx.

number

[0022] However, in this case too, Δx T By applying a similar treatment to HΔx, the direction of molecular deformation can be determined.

[0023] By continuing to deform the molecule in the direction that minimizes the deviation from the harmonic approximation of the energy, it is possible to find a transition state that is a saddle point of energy. Note that if you can reach the vicinity of the saddle point, even if it is not the saddle point itself, you can more easily reach the saddle point by searching for a stationary point using Newton's method.

[0024] If a transition state that is an energy saddle point can be found, a reaction path can be obtained by minimizing the energy from that state. A method for automatically finding a reaction path by calculating the energy or Hessian matrix using a trained model will be described as the first embodiment.

[0025] If the trained model outputs the difference between the harmonic approximation on the hypersurface where the harmonic approximation energy is equal to the direction of deformation Δx of the molecule, the transition state can be searched for by deforming the molecule in the direction where the value of this energy difference is minimized. A method for searching for this transition state using a trained model will be described as the second embodiment.

[0026] In the ADDF method, it takes time to calculate the Hessian matrix (the second derivative of the energy) and determine the direction of molecular deformation based on this result. Therefore, once the direction of molecular deformation is determined, the number of Hessian matrix calculations can be reduced by continuing to deform the molecule in that direction while calculating the force (the first derivative of the energy). Calculating the Hessian matrix is ​​computationally expensive and time-consuming. Therefore, reducing the number of calculations can increase the speed of the method.

[0027] Furthermore, even when the direction of molecular deformation is recalculated, it is not necessarily necessary to perform expensive Hessian matrix calculations for each iteration. One example of this method is the Hessian update method, which uses gradient changes associated with structural changes to update the directional components of the Hessian matrix along the structural changes.

[0028] Since the obtained Hessian matrix is ​​an approximation, its accuracy affects the behavior of the stationary point search calculation. In the stage of approaching the structure to the stationary energy point, the accuracy of the Hessian matrix is ​​not so important. On the other hand, the speed of convergence from the vicinity of the stationary energy point to the stationary energy point depends heavily on the accuracy of the Hessian matrix.

[0029] Therefore, when the Hessian update method is used, the number of steps for structural optimization increases in the vicinity of the poles. However, since it is not necessary to obtain the Hessian matrix by calculation, the time required for each optimization step can be significantly reduced.

[0030] The total computational time is determined by the trade-off between these calculations, but in many cases, the computation time can be reduced by using the Hessian update method. One application of this method is to calculate the exact Hessian matrix once every few steps, and use the Hessian update method for the other steps.

[0031] where the structure x k and structure x l Change in Δx kl = x k -x l and the corresponding gradient change Δg kl = g k -g l By taking the difference using kl DIFF is obtained.

number

[0032] where H0 OLD is the Hessian matrix before updating. When performing a structural optimization calculation, H0 OLD H kl DIFF The matrix to which this is added is used as the Hessian matrix to be determined in the optimization step. The first to third terms on the right-hand side are Δx kl The fourth to sixth terms are the difference in gradient along kl corresponds to a term that eliminates the projected part of the direction.

[0033] Although detailed explanation is omitted, there is another Hessian update method that gives good results in structural optimization of local minimum points and first-order saddle points.

[0034] A method of using a trained model to calculate or update the Hessian matrix at appropriate times will be described as embodiment 3. As described above, in the present disclosure using a trained model, it is possible to ensure a reliable search without implementing these calculation time reductions, but by implementing these calculation time reductions, it is possible to achieve even faster speeds.

[0035] (AFIR Act) While the ADDF method is a search method that starts with a single molecule, the AFIR method is also capable of searching for pathways in which multiple molecules react.

[0036] In the AFIR (Artificial Force Induction of Reaction) method, a reaction is induced by pushing reactant molecules together. For example, when reactants A and B are reacting, the distance between A and B is r AB By adding a term proportional to the energy, structural optimization is performed on a potential surface that cancels out the reaction barrier.

number

[0037] By using this, the optimization problem that requires searching in the direction up the potential surface, which was difficult in reaction path search, can be replaced with a structural optimization problem that searches down the potential surface. According to the AFIR method, after the path calculation is completed, the artificial force term αr AB By removing the , we can obtain an approximate reaction path and identify the location of the barrier.

[0038] In addition, unlike the ADDF method, the AFIR method is a method that can explore reaction paths without calculating the downward anharmonic distortion, which has a high computational cost.

[0039] In applications to polyatomic molecules, we use an artificial force function that can be expressed as follows:

number

[0040] where E(Q) is the potential energy depending on the coordinate Q. When ρ = 1, the artificial force pushes the reactants together, and when ρ = -1, the artificial force pushes them apart. r ij is the distance between atoms i and j. r ij weight function ω ij and calculate the sum for all atoms contained in fragments A and B. Also, Ri and Rj are the covalent bond radii of atoms i and j, respectively.

number

[0041] The calculation is not sensitive to the value of p, and is often set to 6.0 for any real number p. The constant α used to calculate the force strength is given by the following formula:

number

[0042] The constant γ is the model collision energy, which gives an approximate upper limit to the barrier that can be overcome. The model collision energy γ is set by the user. Here, ε = 1.0061 [kJ / mol] and R0 = 3.8164 [Å] may be used. In this case, α corresponds to the average force acting on two Ar atoms from the energy minimum to the turning point of the corresponding Lennard-Jones potential in a direct collision of two Ar atoms at the model collision energy γ.

[0043] The GRRM program implements the SC-AFIR (Single-Component AFIR) algorithm, an automated reaction path search method using the AFIR method. The SC-AFIR method automatically defines fragments to which artificial forces are applied, and then applies the AFIR method to perform path search and search for new stable structure EQs. The AFIR method is also applied to new EQs to obtain new reaction paths and EQs.

[0044] The SC-AFIR method performs this iterative process automatically. In this automatic search, not only the reaction of pushing forces (ρ = 1) but also the reaction of pulling forces (ρ = -1) are applied, so it is possible to search for reaction paths involving bond cleavage.

[0045] Therefore, the user's experience may be required to set the calculation conditions, such as an appropriate γ setting. Note that Non-Patent Document 3 gives an example of γ setting of 100 [kJ / mol], which can be used as a guide.

[0046] A method for using a trained model in this AFIR will be described as the fourth embodiment.

[0047] Each of the above will be described below as an embodiment showing a non-limiting example. In each embodiment, the processing is specifically realized, for example, by software information processing using hardware resources. The hardware information processing device is implemented, for example, using one or more processors (processing circuits) and one or more memories (storage circuits) connected to the processors.

[0048] The processing circuitry included in the information processing device writes or reads data to or from the storage circuitry at appropriate times. The data may be data required for processing, data to be input or output, or a program for executing information processing.

[0049] In the following, a trained model is, for example, a model related to NNP (Neural Network Potential). The NNP model is a model that outputs energy for each atom when information such as the three-dimensional arrangement of atoms is input. An information processing device can use this model by having a processing circuit refer to a memory circuit.

[0050] The processes enclosed in bold lines in the flowchart are processes that can be calculated using a trained model in the present disclosure. In the following, energy and the like are used as physical quantities, but the physical quantity used can be selected arbitrarily. Furthermore, as an example, the processing circuit uses a model centered on the training model in NNP for acquiring energy, but this can also be interpreted as a training model for acquiring any physical quantity, as appropriate.

[0051] (First embodiment) 1 is a flowchart showing the processing of the processing circuit according to the first embodiment. In this embodiment, an example of using a trained model to speed up processing in the ADDF method will be described.

[0052] The processing circuit reads data relating to the structure of a target molecule, etc. (S100). The data relating to the structure includes, for example, the types of atoms that make up the molecule and the three-dimensional coordinates of the atoms.

[0053] The processing circuit calculates a physical quantity, such as energy, for the loaded structure, and if this physical quantity is not a minimum, searches for a structure with a minimum energy (S102). The processing circuit may perform this process using a trained model. Note that since this process is outside the iteration, even if a trained model is not used, it is less likely to become a bottleneck than modeling other processes.

[0054] If a trained model is used in S102, the trained model can be a model that outputs the value of equation (1) when a structure is input and outputs energy. The processing circuit can also use a trained model trained using stable structures obtained by the various optimization methods described above as training data. If the energy is not a minimum, the processing circuit obtains a stable structure x0 for the structure x read in S100 via the trained model.

[0055] If the physical quantity obtained from the read structure x is minimal, or after searching for a structure x0 where the physical quantity is minimal in S102, the processing circuit executes a transition state search process (S104 to S112) using this minimal structure as the initial value.

[0056] The processing circuit acquires physical quantities from the structure (S104). As an example, the processing circuit acquires energy and a Hessian matrix, as shown in FIG. 1. This process may be performed using a trained model. The processing circuit may acquire energy from the structure using the trained model and acquire the Hessian matrix by second-order differentiation of this energy. Alternatively, the trained model may output a physical quantity from which the second derivative of energy can be calculated, and the second derivative of energy (Hessian matrix) may be acquired using this physical quantity. Examples of physical quantities from which the second derivative of energy can be calculated include natural frequency and reduced mass.

[0057] As another example, the processing circuit may use a pre-trained model that inputs structural data and outputs a Hessian matrix. As yet another example, if the force is available, the processing circuit may obtain the Hessian matrix by positionally differentiating the force value. For example, in the second iteration and thereafter, the processing circuit can reduce the calculation time for the Hessian matrix by using the processing result of S112 (described later) in the previous iteration.

[0058] In ADDF, the direction of deformation of the molecule, such as the direction of translation, rotation, bending, etc., can be determined. The processing circuit sets the width of deformation of this molecule within the range of the maximum molecular deformation displacement amount specified by the user (S106). The initial value may be input by the user. From the next iteration, the width may be automatically determined from the width obtained in the previous iteration, for example, the width may be narrowed or widened depending on the state obtained in the iteration. The molecular deformation width may be set using a known method.

[0059] After setting the molecular deformation range, the processing circuit calculates the direction in which the difference in energy from the harmonic approximation is minimal, i.e., the molecular deformation direction (S108). The processing circuit may calculate the difference from the harmonic approximation based on the harmonic approximation energy calculated from equations (2, 3, 4, 5, 6, 7, 9), or may search for an energy minimum on a hypersphere with equal harmonic approximation energy, as shown in equation (8). The processing circuit may also obtain this direction using a trained model. The trained model used in this step may be, for example, a model trained using various training data obtained using equations (2) to (9).

[0060] The processing circuitry obtains the coordinates of each atom of the molecule that has been deformed in the direction obtained in S108 within the width set in S106 (S110).

[0061] The processing circuit calculates the force acting on each atom using the structure including the deformed coordinates (S112). This step may be performed using a trained model. For example, the processing circuit may obtain the energy of each atom using a trained model that obtains energy from structural data, and obtain the force by positionally differentiating (first-order differential) this energy. As another example, the processing circuit may obtain the force using a trained model that obtains force from structural data.

[0062] The processing circuit determines whether the structure acquired in S110 is a stationary point (S114). The processing circuit may perform this determination, for example, based on whether the force acquired in S112 is 0. Alternatively, the processing circuit may use, as a determination criterion, whether the force is equal to or less than a predetermined minute value.

[0063] If it is determined that the point is not a stationary point (S114: NO), the processing circuit repeats the process from S104 using the structure modified in S110.

[0064] If it is determined that the point is a stationary point (S114: YES), the processing circuit determines a reaction path based on the structure deformed in S110 (S116). The processing circuit determines the reaction path, for example, by performing bidirectional structural optimization from the transition state. For the structural optimization process in determining this reaction path, the processing circuit may use a trained model. Note that, as with the process of S102, the processing circuit may use a general optimization method rather than a trained model for reasons such as improving accuracy, since the process of S116 is also outside the iteration.

[0065] As described above, according to this embodiment, the information processing device can execute calculations at any granularity using a pre-trained model. If a pre-trained model is not used, it is necessary to calculate energy by first-principles calculations and then calculate the first derivative (force) and second derivative (Hessian matrix) of this energy. However, since the calculation time for this energy is long, the process of searching for a stationary point, which must be executed sequentially multiple times within an iteration, takes an enormous amount of time.

[0066] According to this embodiment, energy calculation is significantly accelerated, thereby shortening the overall processing time. Furthermore, if the trained model can output the first derivative of the energy, the Hessian matrix can be obtained by first differentiating this value. This also contributes to the speedup. Furthermore, if the trained model can output the second derivative of the energy, the effect of speedup becomes even more significant.

[0067] In the above description, it has been explained that it is possible to select whether or not to use a trained model for each process, but this is not limited to this. For example, it may be possible to select whether to execute the same process using a trained model or by first-principles calculation, etc., depending on the iteration. As a non-limiting example, for at least one processing step, first-principles calculation may be executed every predetermined number of iterations to ensure a certain degree of accuracy. This is not limited to this embodiment, but also applies to any of the embodiments described below.

[0068] (Second embodiment) 2 is a flowchart showing the processing of the processing circuit according to the first embodiment. In this embodiment, an example of using a trained model to speed up processing in the ADDF method will be described.

[0069] The processing circuit reads data relating to the structure of the target molecule, etc. (S200). This process is the same as S100 in the first embodiment.

[0070] The processing circuit calculates a physical quantity, for example, energy, for the read structure, and if this physical quantity is not a minimum, searches for a minimum energy structure (S202). This process is equivalent to S102 in the first embodiment.

[0071] If the physical quantity obtained from the read structure x is minimal, or after searching for a structure x0 where the physical quantity is minimal in S202, the processing circuit executes the transition state search process (S204 to S210) using this minimal structure as the initial value.

[0072] The processing circuit sets the width of the molecule deformation (S204). This process is, for example, the same as the process in S106 in the first embodiment.

[0073] Alternatively, the processing circuitry may perform the process without using a trained model to obtain at least one of the energy, force, or Hessian matrix, and may obtain the deformation direction using a trained model that receives a structure as an input and outputs an inharmonicity.

[0074] For example, the processing circuit may use a model that outputs the difference between the direction of molecular deformation and a harmonic approximation on a hypersurface where the harmonic approximation energy is equal. Using such a model, the processing circuit can estimate the direction of deformation by searching for the direction of molecular deformation where the energy difference is minimal. Appropriate training data can also be obtained for training this training model.

[0075] After setting the molecular deformation range, the processing circuitry calculates the direction in which the deviation from the harmonic approximation of the energy is minimized, i.e., the molecular deformation direction (S206). This process is equivalent to the process in S108 in the first embodiment.

[0076] The processing circuitry obtains the coordinates of each atom by deforming the molecule within the width set in S204 in the direction obtained in S206 (S208). This process is equivalent to the process in S110 in the first embodiment.

[0077] The processing circuit uses the structure including the deformed coordinates to calculate the forces acting on each atom (S210). This process is equivalent to the process of S112 in the first embodiment.

[0078] The processing circuit determines whether the structure acquired in S208 is a stationary point (S212). This process is equivalent to the process in S114 in the first embodiment.

[0079] If it is determined that the point is not a stationary point (S212: NO), the processing circuit repeats the process from S204 using the structure modified in S208.

[0080] If it is determined that the point is a stationary point (S212: YES), the processing circuit determines a reaction path based on the structure transformed in S208 (S214). The subsequent processing is the same as S114 to S116 in the first embodiment.

[0081] According to this embodiment, the information processing device can execute calculations at any granularity using a trained model, as in the first embodiment. Furthermore, according to this embodiment, it becomes possible to obtain data for determining the deformation direction without directly obtaining energy, etc.

[0082] In the search for the direction to deform a molecule, there have been cases where the search was not successful in real time due to the large number of calculations required during the process. According to the information processing device of this embodiment, anharmonicity data can be obtained without calculating physical quantities such as energy, thereby realizing a more reliable search. Furthermore, since it is not necessary to obtain various physical quantities to obtain anharmonicity, it is possible to further speed up the method of the first embodiment.

[0083] (Third embodiment) 3 is a flowchart showing the processing of a processing circuit according to a third embodiment. In this embodiment, an example of using a trained model to speed up processing in the ADDF method is described. The processing according to this embodiment traces the processing of ADDF, which does not use a trained model, and uses a trained model in processing that requires high computational costs.

[0084] The processes of S300 and S302 are equivalent to the processes of S100 and S102 in the first embodiment, respectively.

[0085] The processing circuit acquires physical quantities, such as energy and Hessian matrix, from the structure outside the iteration (S304). The processing content is the same as that of S104 in the first embodiment.

[0086] If the physical quantity obtained from the structure x read in S302 is minimal, or after searching for a structure x0 in which the physical quantity is minimal in S302, the processing circuit executes a transition state search process (S306 to S318) using this minimal structure as an initial value.

[0087] The processes from S306 to S316 are equivalent to the processes from S106 to S116 in the first embodiment, respectively, except for the process branched to at S314.

[0088] If the deformed structure is not a stationary point (S314: NO), the processing circuit proceeds to a process of obtaining a Hessian matrix (S318).

[0089] FIG. 4 is a flowchart showing the process of obtaining the Hessian matrix according to this embodiment.

[0090] When the processing circuit determines to obtain the Hessian matrix, it determines whether or not to perform second-order differentiation (S320). As described above, when the direction of deformation is further obtained after the molecule is deformed, it is not necessarily necessary to perform expensive Hessian matrix calculations for each iteration.

[0091] The processing circuit determines whether to perform second-order differentiation to calculate a highly accurate Hessian matrix or to update the Hessian matrix without performing second-order differentiation, depending on various conditions. This condition may be determined, for example, by the number of iterations or based on an energy difference or the like.

[0092] If it is determined that second-order differentiation is to be performed (S320: YES), the processing circuit performs second-order differentiation to obtain a Hessian matrix (S322). The processing circuit may obtain the Hessian matrix using a trained model. Alternatively, the trained model may output a physical quantity from which the second-order derivative of energy can be calculated, and the second-order derivative of energy (Hessian matrix) may be obtained using the physical quantity. Examples of physical quantities from which the second-order derivative of energy can be calculated include the natural frequency and the reduced mass.

[0093] The processing circuitry may, for example, obtain the energy using a trained model and take the second derivative, obtain the force using a trained model and take the first derivative, or obtain the Hessian matrix using a trained model.

[0094] If it is determined that second differentiation is not to be performed (S320: NO), the processing circuit obtains a Hessian matrix without performing second differentiation (S324). The processing circuit updates the Hessian matrix using, for example, a Hessian update method. The Hessian update method updates the Hessian matrix based on Equation (10). The processing circuit may perform this process using a trained model, or may perform this process without using a trained model.

[0095] As described above, according to this embodiment, the information processing device can execute calculations at any granularity using a pre-trained model. If a pre-trained model is not used, it is necessary to calculate energy by first-principles calculations and then calculate the first derivative (force) and second derivative (Hessian matrix) of this energy. However, since the calculation time for this energy is long, the process of searching for a stationary point, which must be executed sequentially multiple times within an iteration, takes an enormous amount of time.

[0096] According to this embodiment, by appropriately omitting the calculation of the Hessian matrix, it is possible to achieve even faster speeds than the previous embodiments.

[0097] In the first to third embodiments, the ADDF method has been described. In this ADDF method, the calculation of the Hessian matrix determines the speed. As an example, when performed using first-principles calculations, it currently takes approximately 1028 minutes on a 4-core processor and approximately 340 minutes on a 16-core processor. On the other hand, according to the embodiment of the present disclosure, it takes approximately 1.33 seconds without I / O and approximately 33 seconds with I / O. Therefore, even at a conservative estimate, a speed increase of approximately 600 times can be achieved.

[0098] (Fourth embodiment) The information processing device according to this embodiment performs processing at high speed using a trained model for the AFIR method.

[0099] FIG. 5 is a flowchart showing an example of processing according to this embodiment.

[0100] The processing circuit reads the initial molecular configuration (S400). For example, the processing circuit reads the molecular configuration of interest in the AFIR method.

[0101] The processing circuit calculates the artificial force parameter (artificial force coefficient) ω in equation (12) between these loaded molecules. ij(or α in Equation (11)) is calculated (S402). The processing circuitry obtains a weighting coefficient for the artificial force based on, for example, Equation (14) from the model collision energy γ. The processing circuitry may obtain this parameter by, for example, estimating the energy maximum point or calculating the energy difference ΔE from the starting structure.

[0102] The processing circuit may calculate the artificial force parameters from input values ​​or define them for automatic estimation. In the automatic estimation definition, the processing circuit estimates the two reacting atoms from the specified fragment and the force direction (attraction or repulsion), estimates the activation energy near the transition state, and automatically defines the artificial force required for path search. The following may be used as some non-limiting examples of methods for determining the artificial force parameters.

[0103] The processing circuitry may, for example, determine this artificial force parameter from the input parameters.

[0104] The artificial force parameters may be estimated using a list of at least one of typical activation parameters, bond energies, or reaction rate constants for bond formation / dissociation between two atoms of a presumed reaction as input parameters.

[0105] As input parameters, the intermediate may be estimated using the typical interatomic distances of the transition state for the bond formation / dissociation of the estimated two atoms in the reaction, and the artificial force parameters may be estimated by performing a one-point energy calculation.

[0106] As input parameters, a typical reaction coordinate distance of the transition state for the bond formation / dissociation of two atoms in a pre-defined estimated reaction can be used. A structure can be created by moving the typical reaction coordinate distance on a PES (Potential Energy Surface) that adds a very large artificial force (approximately 10 times or more than the value discussed in a normal chemical reaction) between the specified fragments, and that point can be used as an estimated intermediate. The artificial force parameters can then be estimated by performing a single-point energy calculation.

[0107] The processing circuit uses the calculated artificial force parameters to search for the positions of the atoms (S404 to S410).

[0108] The processing circuit uses the parameters determined in S402 to obtain a force according to Equation (11) or Equation (12) (S404). The processing circuit may perform this force obtaining calculation using a trained model that obtains forces from atomic configurations and artificial force parameters.

[0109] The processing circuit moves the position of each atom according to the force obtained in S404 (S406).

[0110] The processing circuitry obtains energy (S408) based on the structure according to the positions of the atoms moved in S406. The processing circuitry may perform this energy obtainment calculation using the trained model.

[0111] The processing circuit determines whether the energy obtained in S408 is lower than the energy before the atom was moved in S406 (S410). In this comparison, the processing circuit may use an energy value obtained in a previous iteration. Alternatively, in the first iteration, the processing circuit may obtain the energy of a separate, unmoved state of the atom, with or without using a trained model.

[0112] If the energy is decreasing (S410: YES), the processing circuit repeats the process from S404.

[0113] If the energy has not decreased (S410: NO), the processing circuit determines whether to terminate the atomic movement (S412). The processing circuit makes this determination by determining whether the reaction was properly tracked based on, for example, changes in the bonding state. The processing circuit determines that the atomic movement has not terminated if, for example, the bonding state has stopped midway.

[0114] The processing circuit may determine whether or not the target chemical reaction has been tracked at the end of the reaction path search on the PES to which artificial force has been applied. The following may be used as a non-limiting example of a method for determining whether or not the target chemical reaction has been tracked.

[0115] The processing circuit may determine whether or not there is a maximum point in the route search with the artificial force removed, and make the determination in S412 based on the result of this determination.

[0116] The processing circuit may use the typical bond distance to make the determination by comparing a symmetric matrix that stores the bond information before and after the path search. The symmetric matrix may be, for example, a matrix that is 1 if the bond distance is less than the typical distance and 0 otherwise.

[0117] The processing circuitry may determine S412 by using the Root Mean Square Deviation (RMSD) of the atomic positions of the pre- and post-reaction structures and comparing it with a reference value.

[0118] If the movement of the atoms is not completed (S412: NO), the processing circuit repeats the process from S404. Here, instead of repeating the process with the same parameters, the processing circuit may, for example, increase the value of γ to update the parameters so that the local maximum point can be exceeded, and repeat the process from S404.

[0119] The processing circuit selects the coordinates of the atoms (approximate transition state) where the energy excluding the artificial force is maximized (S414).

[0120] The processing circuitry may perform a transition state search (S416) starting from the approximate transition state selected in S414. The processing circuitry may perform at least one of the steps in the transition state search using the trained model.

[0121] The processing circuit obtains the structural data resulting from the search in S416 and determines the reaction path by bidirectionally optimizing the structure from the transition state (S418).

[0122] As described above, the information processing device according to this embodiment can also execute the AFIR method at high speed. For example, by executing operations such as S404 and S408 at high speed using a trained model, the information processing device can significantly speed up the AFIR method, which does not need to calculate energy and forces for many structures.

[0123] In addition, for example, by performing the process of S402, the parameters can be easily determined. Furthermore, for example, by performing the determination of S412, the reaction path can be reliably and correctly searched.

[0124] In the learning process to obtain the trained model used in these embodiments, the three-dimensional structure provided as training data does not necessarily include a reaction transition state. For example, by providing a three-dimensional structure that deviates from the energy minimum structure obtained in molecular dynamics calculations as training data, it is possible to obtain a trained model capable of estimating the energy, first derivative of the energy, second derivative of the energy, and energy anharmonicity of the reaction transition state. In this case, since there is no need to calculate the reaction transition state when creating the training data, a large amount of training data can be prepared quickly, thereby improving the accuracy of the trained model. Note that the trained model in the above-described embodiments may also be a concept that includes, for example, a model that has been trained as described above and then further distilled using a general method.

[0125] Some or all of the devices (information processing devices) in the above-described embodiments may be configured as hardware, or may be configured as software (program) information processing executed by a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), etc. When configured as software information processing, software that realizes at least some of the functions of each device in the above-described embodiments may be stored on a non-transitory storage medium (non-transitory computer-readable medium) such as a CD-ROM (Compact Disc-Read Only Memory) or a USB (Universal Serial Bus) memory, and the software information processing may be executed by loading the software into a computer. The software may also be downloaded via a communications network. Furthermore, all or part of the software processing may be implemented in a circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array), thereby allowing the software information processing to be executed by hardware.

[0126] The storage medium that stores the software may be a removable medium such as an optical disk, or a fixed medium such as a hard disk or memory. The storage medium may be located inside the computer (such as a main memory or auxiliary memory), or may be located outside the computer.

[0127] 6 is a block diagram showing an example of the hardware configuration of each device (information processing device) in the above-described embodiment. Each device may be realized as a computer 7 including, for example, a processor 71, a main storage device 72 (memory), an auxiliary storage device 73 (memory), a network interface 74, and a device interface 75, all of which are connected via a bus 76.

[0128] Although the computer 7 in FIG. 6 includes one of each component, it may include multiple of the same component. Also, while FIG. 6 shows one computer 7, the software may be installed on multiple computers, and each of the multiple computers may execute the same or different parts of the software. In this case, a distributed computing configuration may be used in which each computer communicates with the other computers via a network interface 74 or the like to execute the processing. In other words, each device (information processing device) in the above-described embodiment may be configured as a system in which one or more computers execute instructions stored in one or more storage devices to realize its functions. Furthermore, the system may be configured such that information transmitted from a terminal is processed by one or more computers on a cloud, and the processing results are transmitted to the terminal.

[0129] The various calculations of each device (information processing device) in the above-described embodiments may be executed in parallel using one or more processors, or using multiple computers via a network. Furthermore, the various calculations may be distributed to multiple processor cores within a processor and executed in parallel. Furthermore, some or all of the processes, means, etc. disclosed herein may be implemented by at least one processor and storage device provided on a cloud that can communicate with computer 7 via a network. Thus, each device in the above-described embodiments may be implemented in the form of parallel computing using one or more computers.

[0130] The processor 71 may be an electronic circuit (processing circuit, processing circuitry, CPU, GPU, FPGA, ASIC, etc.) that at least controls or performs calculations on a computer. The processor 71 may also be a general-purpose processor, a dedicated processing circuit designed to perform a specific calculation, or a semiconductor device that includes both a general-purpose processor and a dedicated processing circuit. The processor 71 may also include an optical circuit, or may include a calculation function based on quantum computing.

[0131] The processor 71 may perform arithmetic processing based on data or software input from each device or the like configured inside the computer 7, and may output the calculation results or control signals to each device or the like. The processor 71 may control each component constituting the computer 7 by executing the OS (Operating System) of the computer 7, applications, etc.

[0132] Each device (information processing device) in the above-described embodiments may be realized by one or more processors 71. Here, the processor 71 may refer to one or more electronic circuits arranged on one chip, or may refer to one or more electronic circuits arranged on two or more chips or two or more devices. When multiple electronic circuits are used, the electronic circuits may communicate with each other via wire or wirelessly.

[0133] The main memory device 72 may store instructions executed by the processor 71, various data, etc., and information stored in the main memory device 72 may be read by the processor 71. The auxiliary memory device 73 is a memory device other than the main memory device 72. Note that these memory devices refer to any electronic component capable of storing electronic information, and may be semiconductor memory. The semiconductor memory may be either volatile memory or non-volatile memory. The memory device for saving various data, etc. in each device (information processing device) in the above-described embodiments may be realized by the main memory device 72 or the auxiliary memory device 73, or may be realized by an internal memory built into the processor 71. For example, the memory circuit in the above-described embodiments may be realized by the main memory device 72 or the auxiliary memory device 73.

[0134] When each device (information processing device) in the above-described embodiments is configured with at least one storage device (memory) and at least one processor connected (coupled) to this at least one storage device, at least one processor may be connected to one storage device. At least one storage device may be connected to one processor. A configuration in which at least one processor among multiple processors is connected to at least one storage device among multiple storage devices may also be included. This configuration may also be realized by storage devices and processors included in multiple computers. Furthermore, a configuration in which a storage device is integrated with a processor (for example, a cache memory including an L1 cache and an L2 cache) may also be included.

[0135] The network interface 74 is an interface for connecting to the communication network 8 wirelessly or via a wire. The network interface 74 may be an appropriate interface, such as one that conforms to an existing communication standard. Information may be exchanged with an external device 9A connected via the communication network 8 via the network interface 74. The communication network 8 may be any one of a WAN (Wide Area Network), a LAN (Local Area Network), a PAN (Personal Area Network), etc., or a combination thereof, as long as information is exchanged between the computer 7 and the external device 9A. An example of a WAN is the Internet, an example of a LAN is IEEE 802.11 or Ethernet (registered trademark), and an example of a PAN is Bluetooth (registered trademark) or NFC (Near Field Communication), etc.

[0136] The device interface 75 is an interface such as a USB that directly connects to the external device 9B.

[0137] The external device 9A is a device connected to the computer 7 via a network. The external device 9B is a device connected directly to the computer 7.

[0138] For example, the external device 9A or the external device 9B may be an input device. The input device may be, for example, a camera, a microphone, a motion capture device, various sensors, a keyboard, a mouse, a touch panel, or other device, and provides acquired information to the computer 7. Alternatively, the external device 9A or the external device 9B may be a device equipped with an input unit, a memory, and a processor, such as a personal computer, a tablet terminal, or a smartphone.

[0139] Furthermore, the external device 9A or the external device 9B may be, for example, an output device. The output device may be, for example, a display device such as an LCD (Liquid Crystal Display) or an organic EL (Electro Luminescence) panel, or a speaker that outputs sound or the like. Alternatively, the output device may be a device including an output unit, a memory, and a processor, such as a personal computer, a tablet terminal, or a smartphone.

[0140] Furthermore, the external device 9A or the external device 9B may be a storage device (memory). For example, the external device 9A may be a network storage or the like, and the external device 9B may be a storage such as an HDD.

[0141] Furthermore, the external device 9A or the external device 9B may be a device having some of the functions of the components of each device (information processing device) in the above-described embodiments. That is, the computer 7 may transmit some or all of the processing results to the external device 9A or the external device 9B, or may receive some or all of the processing results from the external device 9A or the external device 9B.

[0142] In this specification (including the claims), when the expression "at least one of a, b, and c" or "at least one of a, b, or c" (including similar expressions) is used, it includes any of a, b, c, a-b, a-c, b-c, or a-b-c. It may also include multiple instances of any element, such as a-a, a-b-b, a-a-b-b-c-c, etc. It also includes the addition of elements other than the listed elements (a, b, and c), such as having d, as in a-b-c-d.

[0143] In this specification (including claims), when expressions such as "using data as input / based on / according to / in response to" (including similar expressions) are used, unless otherwise specified, this includes cases where the data itself is used, or where data that has been processed in some way (e.g., data with noise added, normalized data, features extracted from data, intermediate representations of data, etc.) is used. Furthermore, when a statement is made that a result is obtained "using data as input / based on / according to / in response to" (including similar expressions), this includes cases where the result is obtained based solely on the data, or where the result is influenced by other data, factors, conditions, and / or states other than the data, unless otherwise specified. Furthermore, when a statement is made that "data is output" (including similar expressions), this includes cases where the data itself is used as output, or where data that has been processed in some way (e.g., data with noise added, normalized data, features extracted from data, intermediate representations of data, etc.) is used as output, unless otherwise specified.

[0144] When the terms "connected" and "coupled" are used in this specification (including the claims), they are intended as open-ended terms that encompass any of direct connection / coupling, indirect connection / coupling, electrically connection / coupling, communicatively connection / coupling, functionally connection / coupling, and physically connection / coupling. These terms should be interpreted appropriately according to the context in which they are used, but should be interpreted without limitation to include any form of connection / coupling that is not intentionally or naturally excluded.

[0145] In this specification (including the claims), the expression "A configured to B" may include the physical structure of element A having a configuration capable of performing operation B, and the permanent or temporary setting / configuration of element A being configured / set to actually perform operation B. For example, if element A is a general-purpose processor, it is sufficient that the processor has a hardware configuration capable of performing operation B, and is configured to actually perform operation B by setting a permanent or temporary program (instruction). Also, if element A is a dedicated processor or dedicated arithmetic circuit, it is sufficient that the circuit structure of the processor is implemented to actually perform operation B, regardless of whether control instructions and data are actually attached to it.

[0146] When used in this specification (including the claims), terms implying containing or possessing (e.g., "comprising" or "including" and "having"), they are intended to be open-ended terms that include containing or possessing things other than the object designated by the object of the term. When the object of such terms implies no quantity or a singular number (e.g., expressions using the articles "a" or "an"), the expression should be construed as not being limited to a specific number.

[0147] In this specification (including the claims), although expressions such as "one or more" or "at least one" are used in some places and expressions that do not specify a quantity or that imply a singular number (expressions using the articles a or an) are used in other places, the latter expressions are not intended to mean "one." In general, expressions that do not specify a quantity or that imply a singular number (expressions using the articles a or an) should be interpreted as not necessarily being limited to a specific number.

[0148] In this specification, when a particular advantage / result is described as being obtained from a particular configuration of an embodiment, it should be understood that the same advantage / result can also be obtained from one or more other embodiments having the same configuration, unless otherwise stated. However, it should be understood that the presence or absence of the effect generally depends on various factors, conditions, and / or circumstances, and that the effect is not necessarily obtained by the configuration. The effect is merely obtained by the configuration described in the embodiment when various factors, conditions, and / or circumstances are satisfied, and the effect does not necessarily occur in a claimed invention that defines the same or a similar configuration.

[0149] When terms such as "maximize" are used in this specification (including the claims), they include finding a global maximum, finding an approximation of a global maximum, finding a local maximum, and finding an approximation of a local maximum, and should be interpreted accordingly according to the context in which the term is used. They also include finding approximations of these maxima probabilistically or heuristically. Similarly, when terms such as "minimize" are used, they include finding a global minimum, finding an approximation of a global minimum, finding a local minimum, and finding an approximation of a local minimum, and should be interpreted accordingly according to the context in which the term is used. They also include finding approximations of these minima probabilistically or heuristically. Similarly, when terms such as "optimize" are used, they include finding a global optimum, finding an approximation of a global optimum, finding a local optimum, and finding an approximation of a local optimum, and should be interpreted accordingly according to the context in which the term is used. It also includes finding approximations of these optimum values ​​probabilistically or heuristically.

[0150] In this specification (including claims), when multiple pieces of hardware perform a predetermined process, the pieces of hardware may cooperate to perform the predetermined process, or some of the hardware may perform all of the predetermined process. Furthermore, some of the hardware may perform part of the predetermined process, and other hardware may perform the rest of the predetermined process. In this specification (including claims), when expressions such as "one or more pieces of hardware perform a first process, and the one or more pieces of hardware perform a second process" (including similar expressions) are used, the hardware performing the first process and the hardware performing the second process may be the same or different. In other words, it is sufficient that the hardware performing the first process and the hardware performing the second process are included in the one or more pieces of hardware. Note that hardware may also include an electronic circuit or a device including an electronic circuit.

[0151] In this specification (including the claims), when multiple storage devices (memories) store data, each of the multiple storage devices may store only a portion of the data, or may store the entire data. Also, a configuration in which only some of the multiple storage devices store data may be included.

[0152] Although the embodiments of the present disclosure have been described in detail above, the present disclosure is not limited to the individual embodiments described above. Various additions, modifications, substitutions, and partial deletions are possible within the scope of the conceptual idea and spirit of the present disclosure, which is derived from the content defined in the claims and their equivalents. For example, when numerical values ​​or formulas are used in the above-described embodiments, they are shown for illustrative purposes and do not limit the scope of the present disclosure. Furthermore, the order of each operation shown in the embodiments is also illustrative and does not limit the scope of the present disclosure.

Claims

1. one or more memories; one or more processors; Equipped with The one or more processors: Searching for reaction paths using one or more trained models that output physical quantities related to a molecule when a three-dimensional arrangement of two or more atoms that constitute the molecule is input; Information processing device.

2. The one or more trained models are models that output at least one of energy, a first derivative of energy, a second derivative of energy, a physical quantity from which the second derivative of energy can be calculated, or anharmonicity. The information processing device according to claim 1.

3. The trained model is a model trained using teacher data that does not include a reaction transition state.

3. The information processing device according to claim 1 or 2.

4. The one or more processors: Search for transition states using the ADDF (Anharmonic Downward Distortion Following) method.

4. An information processing device according to claim 1.

5. The one or more processors: In the ADDF method, a Hessian matrix representing a second derivative of energy is obtained using at least one of the one or more trained models each time a molecular deformation is performed; The information processing device according to claim 4.

6. The one or more processors: Search for transition states using the AFIR (Artificial Force Induced Reaction) method.

4. An information processing device according to claim 1.

7. The one or more processors: determining artificial force parameters in the AFIR method based on the input parameters; The information processing device according to claim 6.

8. The one or more processors: After performing the AFIR calculation, it is determined whether the target chemical reaction has been found. If you want to continue exploring, Update the artificial force parameters, Perform route planning, 8. The information processing device according to claim 6 or 7.

9. The one or more processors: inputting the three-dimensional configuration of two or more atoms constituting a molecule into the one or more trained models; Searching for the reaction pathway based on physical quantities related to the molecule output from the one or more trained models; 9. An information processing device according to claim 1.

10. The one or more processors: inputting the three-dimensional configuration into the one or more trained models, and repeating a process of deforming the three-dimensional configuration based on the physical quantity outputted from the one or more trained models, and searching for the reaction path based on the three-dimensional configuration after the deformation obtained by the repetition of the process; The information processing device according to claim 9.

11. The one or more processors repeat the process until the three-dimensional layout reaches a stationary point. The information processing device according to claim 10.

12. The one or more processors repeat the process while applying artificial force until a target reaction is found. The information processing device according to claim 10.

13. one or more processors Searching for reaction paths using one or more trained models that output physical quantities related to a molecule when a three-dimensional arrangement of two or more atoms that constitute the molecule is input; Information processing methods.

14. One or more processors, Searching for reaction paths using one or more trained models that output physical quantities related to a molecule when a three-dimensional arrangement of two or more atoms that constitute the molecule is input; A program for executing an information processing method.

15. One or more processors, Searching for reaction paths using one or more trained models that output physical quantities related to a molecule when a three-dimensional arrangement of two or more atoms that constitute the molecule is input; A non-transitory computer-readable medium storing a program for executing an information processing method.

Images