Machine learning apparatus and machine learning method

Abstract

The purpose of the present invention is to provide a machine learning device, a machine learning method, and a program capable of performing machine learning by combining experimental data and calculation data.　A machine learning device according to the present disclosure comprises: an input unit that acquires a first data set and a second data set that is different from the first data set in at least one of structure, object, and generation means; a conversion unit that converts the second data set into a data space to which the first data set belongs; a training unit that uses the converted second data set and the first data set to train a prediction model for predicting material performance; and an output unit that outputs a learning result using the prediction model.

Description

Machine learning device, machine learning method, and program

[0001] The present disclosure relates to a machine learning device, a machine learning method, and a program, and in particular to a machine learning device, a machine learning method, and a program for predicting material performance.

[0002] Patent Document 1 discloses a machine learning method for creating a prediction module that predicts values ​​related to feature quantities from an original data set that includes experimental data and calculation data obtained by simulation.

[0003] Japanese Patent Application Laid-Open No. 2021-22276

[0004] There are two approaches to materials development: experimental approaches and theoretical calculations, and data collection and the use of AI (Artificial Intelligence) are being promoted in each of these approaches. Furthermore, theoretical calculations are broadly divided into first-principles calculations, which are calculation methods based on quantum mechanics that do not rely on experimental values ​​other than fundamental physical constants, and empirical methods that use experimental values ​​as parameters, and predictions for unknown materials are only made by the former.

[0005] To utilize AI in materials development, a sufficient number of samples of the target system are required. However, experimental data for the newly explored areas is scarce to begin with, and the cost of data collection is high. On the other hand, while calculation data from first-principles calculations can be generated at a lower cost than experimental data, it can only provide microscopic information about the material. Therefore, it is difficult to compare calculation data with experimental data or to directly incorporate calculation data into experimental data for transfer learning.

[0006] Furthermore, Patent Document 1 discloses a machine learning method for creating a prediction module from an original data set including experimental data and calculation data obtained by simulation. However, the machine learning method disclosed in Patent Document 1 is based on the premise that experimental data corresponding to the calculation data exists, and therefore cannot be said to be an effective means for predicting unknown materials.

[0007] The present disclosure has been made in consideration of these issues, and aims to provide a machine learning device, a machine learning method, and a program that can perform machine learning by combining experimental data and calculation data.

[0008] The machine learning device according to the present disclosure includes an input unit that acquires a first data set and a second data set that differs from the first data set in at least one of structure, object, and generation means; a conversion unit that converts the second data set into a data space to which the first data set belongs; a learning unit that uses the converted second data set and the first data set to learn a predictive model that predicts material performance; and an output unit that outputs learning results using the predictive model.

[0009] The machine learning method according to the present disclosure acquires a first data set and a second data set that differs from the first data set in at least one of structure, object, and generation means, transforms the second data set into a data space to which the first data set belongs, uses the transformed second data set and the first data set to train a predictive model that predicts material performance, and outputs a learning result using the predictive model.

[0010] The program according to the present disclosure causes a computer equipped with a machine learning device to execute the following steps: acquiring a first data set and a second data set that differs from the first data set in at least one of structure, object, and generation means; converting the second data set into a data space to which the first data set belongs; learning a predictive model that predicts material performance using the converted second data set and the first data set; and outputting the learning results using the predictive model.

[0011] The present disclosure makes it possible to provide a machine learning device, a machine learning method, and a program that can perform machine learning by combining experimental data and calculation data.

[0012] Fig. 1 is a configuration diagram of a machine learning device according to the present disclosure; Fig. 2 is a flowchart showing processing of a machine learning device according to the present disclosure; Fig. 3 is a configuration diagram of a machine learning device according to the present disclosure; Fig. 4 is a flowchart showing processing of a machine learning device according to the present disclosure;

[0013] First Embodiment An example configuration of a machine learning device 10 according to the present disclosure will be described below with reference to Fig. 1. Also, an example process performed by the machine learning device 10 according to the present disclosure will be described with reference to the flowchart shown in Fig. 2.

[0014] The machine learning device 10 shown in FIG. 1 predicts material performance, and includes an input unit 101, a conversion unit 102, a learning unit 103, and an output unit 104.

[0015] The input unit 101 receives a first data set including target input information and a second data set including input information different from the target (S101). The second data set differs from the first data set in at least one of structure, object, and generation means, and details will be described in the second embodiment and subsequent embodiments.

[0016] The conversion unit 102 converts the second data set input by the input unit 101 into a data space to which the first data set belongs (S102). Details of the conversion of the data space will be described in the second embodiment and onwards.

[0017] The learning unit 103 creates and learns a prediction model for predicting material performance using the first data set and the second data set converted by the conversion unit 102, and outputs the learning results using the prediction model at the output unit 104 (S103). At this time, transfer learning is performed using the first data set as target data and the converted second data set as source data.

[0018] This allows source data, which is input information with high commonality, to be used in transfer learning for target data, which is desired input information, thereby enabling effective machine learning and deep learning even when there is no explicit correspondence between the source data and the target data. Therefore, the machine learning device according to the present disclosure can perform machine learning by combining a first data set including desired input information and a second data set including input information different from the desired information.

[0019] Second Embodiment A configuration example of a machine learning device 10 according to the present disclosure will be described in detail below with reference to Fig. 3. Furthermore, a processing example performed by the machine learning device 10 according to the present disclosure will be described in detail with reference to the flowchart shown in Fig. 4. Note that parts that are similar to the configuration example or processing example described in the first embodiment may be omitted to avoid repetitive explanation.

[0020] 3 predicts material performance, and includes an input unit 101, a conversion unit 102, a learning unit 103, and an output unit 104. The machine learning device 10 may further include an estimation unit 105.

[0021] The input unit 101 acquires and inputs data from the server 100 that stores the first data set and the second data set (S201). Note that the acquisition of the first data set and the second data set is not limited to this. For example, a data storage unit (not shown) that stores the data sets may be provided within the machine learning device 10, and the input unit 101 may acquire data from the data storage unit. Alternatively, the first data set may be acquired from the data storage unit within the machine learning device 10, and the second data set may be acquired from the server 100.

[0022] The second data set differs from the first data set in at least one of the structure, the object, and the generation means, i.e., the features in machine learning are different. In this embodiment, the first data set includes experimental data T, and the second data set includes calculation data S. Furthermore, the calculation data S included in the second data set is obtained by simulation using various parameters, and three or more heterogeneous data may be used.

[0023] The structure includes at least one of a chemical formula, a reaction formula, a composition formula, a structural formula, an elemental species, an atomic arrangement, a molecular graph, a SMILES (Simplified Molecular Input Line Entry System) character string, and a descriptor created using a part of these as input. The object includes at least one of energy, a wave function, a band structure, a density of states, a spectral density, a density matrix, an energy difference between two states, a free energy, or a quantity obtained as a function thereof.

[0024] The experimental data T inputted in the input unit 101 is measurement data (data group Tεt) relating to a chemical reaction R expressed by a chemical formula that includes the relationship between a target reactant, a product, and a catalyst.

[0025] As an example, consider a system for obtaining activation energy for an unknown composition of an alloy used in a catalytic reaction. The alloy composition, which is the input of a model trained by transfer learning (described later), is an m-dimensional real vector c → ｉ = (p １ , p ２ , ..., p ｍ ) and is expressed by the following formula (1). ...(1) where m is the type of metal element under consideration.

[0026] The activation energy ε, which is the output of the model trained by transfer learning ｉ is related to the magnitude of the energy barrier in the elementary reaction, that is, the ease with which the reaction proceeds, and is expressed by the following formula (2). ...(2) where R is a real number.

[0027] In this way, the experimental data T input by the input unit 101 can be expressed as a pair of data that will be input to the model learned by transfer learning and data that will be output. In the case of the above example, the alloy composition c → ｉ and activation energy ε ｉ Therefore, (c → ｉ , ε ｉ) is expressed as:

[0028] The calculation data S input in the input unit 101 is calculation data (data group Sεs) corresponding to the elementary process of the catalytic reaction m. → ｉ Activation energy ε ｉ In the system to obtain the input structure s ｉ is the total number of atoms (atom number N) contained in the unit structure of the catalyst surface. ａ The data consists of the arrangement and element type of the element. Here, the arrangement is the spatial coordinate s expressed as a three-dimensional real vector, and the element type is the label r expressed as the atomic number. → ｊ and is expressed by the following formulas (3) and (4). ...(3) ... (4) where e ｊ is the atomic number of the jth atom.

[0029] The calculation data S is, for example, the adsorption energy E obtained by performing a simulation based on first-principles calculations. ｉ The adsorption energy E ｉ is related to the energy gain when the target molecule is adsorbed onto the surface, that is, the ease of adsorption of the molecule, and is expressed by the following formula (5). ...(5)

[0030] In this way, the calculation data S input by the input unit 101 has the structure s ｉ and adsorption energy E ｉ Therefore, (s ｉ , E ｉ ) is expressed as:

[0031] Furthermore, the input unit 101 may input the calculation data S obtained in advance by a simulation, or may acquire the calculation data S by executing a simulation in the machine learning device 10. In this case, the input unit 101 may input parameters for the simulation, and the simulation may be executed in a calculation unit (not shown) provided in the machine learning device 10. Examples of parameters for the simulation include parameters necessary for executing a simulation based on first-principles calculation.

[0032] The estimation unit 105 creates a function (F(S)∈t) that converts the calculation data S input by the input unit 101 into a format that can be interpreted by the machine learning model (S202). The function is preferably configured by the machine learning model. This enables data conversion without explicitly specifying the specific system of the function. In the above example, the problem can be formulated as a black-box optimization problem between the calculation data S and the experimental data T.

[0033] The conversion unit 102 applies the function obtained by the estimation unit 105 to the calculation data S, and performs conversion of the calculation data S (S203). This makes it possible to effectively utilize the calculation data obtained by the simulation for machine learning.

[0034] The conversion unit 102 has a mechanism for projecting a plurality of data in the first data set and the second data set into a common data space, and in the above example, converts the calculation data S into the format of the experimental data T.

[0035] Structure s in calculation data S ｉ From the experimental data T, the composition c → ｉ As a conversion (h:s→c) to (h:s), a conversion using the following formula (6) can be given. ...(6) In formula (6), the structure s ｉ The conversion is performed by ignoring the coordinate information and counting the number for each element label.

[0036] Adsorption energy E in calculation data S ｉ From the above, the activation energy ε in the experimental data T ｉ Conversion to (f:s ｎ ×R ｎ →R), a conversion using the following formula (7) can be given. ...(7) In equation (7), a linear relational equation (Bronsted-Evans-Polanyi law) is used to obtain (ε ｉ ~aE ｉ + b) transformation is performed. Here, a and b are constants, and p(E ｉ ) is the weight.

[0037] The conversion performed by the conversion unit 102 is not limited to the above-mentioned method. For example, data conversion may be enabled by configuring a function that converts calculation data S into experimental data T using a machine learning model. Furthermore, the conversion performed by the conversion unit 102 does not necessarily have to be applied to calculation data S, and the experimental data T may be converted in a manner that adapts it to calculation data S. Alternatively, a method may be adopted in which both the experimental data T and the calculation data S are converted and projected onto a common space.

[0038] The learning unit 103 creates a prediction model for predicting material performance using the first data set input from the input unit 101 as target data and the second data set converted by the conversion unit 102 as source data, and performs learning (S204). After that, the output unit 104 outputs the learning results using the prediction model. This makes it possible to share knowledge between different data sets or tasks, such as experimental data and calculation data, and improve learning efficiency and performance.

[0039] In the example of the catalytic reaction described above, the content output by the output unit 104 is the activation energy value predicted for the unknown composition of a given alloy based on the knowledge, prediction model, and learning results stored in the learning unit 103.

[0040] The machine learning device 10 according to the present disclosure may include a feedback mechanism that proposes promising compositions for new measurements based on the model obtained by transfer learning in the learning unit 103 or the results of the prediction model (S205). The feedback mechanism preferably performs active learning, and preferably proposes an optimal execution policy using, for example, a Bayesian optimization method such as a differential Gaussian process model, thereby enabling the accuracy of learning of the prediction model to be improved.

[0041] In this way, it is possible to provide a machine learning device, a machine learning method, and a program that can perform machine learning by combining experimental data and calculation data.

[0042] The machine learning device according to the present disclosure can efficiently improve the predictive performance for small data such as experimental data by using large-scale computational data generated by a relatively low-cost method such as density functional theory as source data in transfer learning, thereby accelerating material development and reducing costs.

[0043] Furthermore, because the prediction model includes experimental data, it is possible to reflect information not included in the calculation data, such as electron correlation, disturbances, and device characteristics, in the prediction, resulting in prediction results with higher fidelity than models created from calculation data alone.

[0044] Furthermore, by incorporating active learning through a feedback mechanism, efficient sampling is possible even when experimental data is scarce, minimizing the number of experiments required to develop catalysts with desired properties, leading to accelerated catalyst development and cost reduction.

[0045] Although the present disclosure has been described above with reference to the embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate.

[0046] Each drawing is merely an example for describing one or more embodiments. Each drawing may not relate to only one particular embodiment, but may also relate to one or more other embodiments. As will be understood by those skilled in the art, various features or steps described with reference to any one drawing can be combined with features or steps shown in one or more other drawings to create, for example, an embodiment not explicitly shown or described. Not all features or steps shown in any one drawing are necessary to describe an exemplary embodiment, and some features or steps may be omitted. The order of steps described in any drawing may be changed as appropriate.

[0047] Some or all of the above embodiments can be described as, but are not limited to, the following supplementary notes. (Supplementary Note 1) A machine learning device comprising: an input unit that acquires a first data set and a second data set that differs from the first data set in at least one of structure, object, and generation means; a conversion unit that converts the second data set into a data space to which the first data set belongs; a learning unit that uses the converted second data set and the first data set to learn a predictive model that predicts material performance; and an output unit that outputs a learning result using the predictive model. (Supplementary Note 2) The machine learning device according to Supplementary Note 1, wherein the first data set is a data set based on experimental data, and the second data set is a data set based on calculation data. (Supplementary Note 3) The machine learning device according to Supplementary Note 2, wherein the structure includes at least one of a chemical formula, a reaction formula, a composition formula, a structural formula, an elemental species, an atomic configuration, a molecular graph, a SMILES (Simplified Molecular Input Line Entry System) character string, and a descriptor created using a portion of any of these as input, and the object includes at least one of energy, a wave function, a band structure, a density of states, a spectral density, a density matrix, an energy difference between two states, a free energy, or a quantity obtained as a function thereof, and the learning result output from the output unit is a result of aligning at least one of the structure, the object, and the generation means in the second data set with the first data set. (Supplementary Note 4) The machine learning device according to Supplementary Note 2, wherein the second data set uses three or more heterogeneous data. (Supplementary Note 5) The machine learning device according to Supplementary Note 2, wherein the conversion unit has a mechanism for projecting a plurality of data in the first data set and the second data set into a common data space. (Supplementary Note 6) The machine learning device according to Supplementary Note 2, further comprising an estimation unit that constructs a function used in the conversion unit by a machine learning model and converts data. (Supplementary Note 7) The machine learning device according to Supplementary Note 2, further comprising a feedback mechanism that uses the results obtained by the prediction model to propose promising compositions in new measurements.(Supplementary Note 8) A machine learning method comprising: acquiring a first data set and a second data set that differs from the first data set in at least one of structure, expression format, target, and generation means; transforming the second data set into a data space to which the first data set belongs; learning a predictive model that predicts material performance using the transformed second data set and the first data set; and outputting a learning result using the predictive model. (Supplementary Note 9) The machine learning method according to Supplementary Note 8, wherein the first data set is a data set based on experimental data, and the second data set is a data set based on calculation data. (Supplementary Note 10) A program causing a computer provided in a machine learning device to execute the following steps: acquiring a first data set and a second data set that differs from the first data set in at least one of structure, target, and generation means; transforming the second data set into a data space to which the first data set belongs; learning a predictive model that predicts material performance using the transformed second data set and the first data set; and outputting a learning result using the predictive model. (Supplementary Note 11) The program according to Supplementary Note 10, wherein the first data set is a data set based on experimental data, and the second data set is a data set based on calculation data.

[0048] Some or all of the elements (e.g., configurations and functions) described in Supplements 2 to 7 that are dependent on the machine learning device described in Supplement 1 may also be dependent in a similar dependency relationship on the machine learning method described in Supplement 8 and the program described in Supplement 10. Some or all of the elements described in any of the Supplements may be applied to various hardware, software, recording means for recording software, systems, and methods.

[0049] Although the present invention has been described above with reference to the embodiments, the present invention is not limited to the above. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the invention.

[0050] This application claims priority based on Japanese Patent Application No. 2023-170239, filed September 29, 2023, the disclosure of which is incorporated herein by reference in its entirety.

[0051] 10 Machine learning device 100 Server 101 Input unit 102 Conversion unit 103 Learning unit 104 Output unit 105 Estimation unit

Claims

1. An input unit that acquires a first data set and a second data set that differs from the first data set in at least one of its structure, object, and generation means, A conversion unit that converts the second data set into the data space to which the first data set belongs, A learning unit that learns a predictive model for predicting material performance using the converted second data set and the first data set, The system includes an output unit that outputs learning results using the aforementioned prediction model. Machine learning device.

2. The aforementioned first data set is a data set based on experimental data, The second data set is a data set based on computational data. The machine learning apparatus according to claim 1.

3. The aforementioned structure includes at least one of the following: a chemical formula, a reaction formula, an empirical formula, a structural formula, element species, atomic arrangement, a molecular graph, a SMILES (simplified molecular input line entry system) string, and a descriptor created using a portion of these as input. The aforementioned object includes at least one of the following: energy, wave function, band structure, density of states, spectral density, density matrix, energy difference between two states, free energy, or quantities obtained as functions thereof. The learning result output from the output unit is obtained by aligning at least one of the structure, the target, and the generation means in the second data set with that of the first data set. The machine learning apparatus according to claim 2.

4. The second data set utilizes three or more heterogeneous data sets. The machine learning apparatus according to claim 2.

5. The conversion unit has a mechanism for projecting multiple data from the first data set and the second data set into a common data space. The machine learning apparatus according to claim 2.

6. Furthermore, the system includes an estimation unit that performs data transformation by constructing the function used in the transformation unit using a machine learning model. The machine learning apparatus according to claim 2.

7. Furthermore, it includes a feedback mechanism that proposes promising compositions in new measurements based on the results obtained by the predictive model. The machine learning apparatus according to claim 2.

8. A first data set and a second data set that differs from the first data set in at least one of its structure, object, and generation means are obtained. The second data set is transformed into the data space to which the first data set belongs. Using the transformed second data set and the first data set, a predictive model for predicting material performance is trained. Output the learning results using the aforementioned predictive model. Machine learning methods.

9. The transformation unit transforms the second data set into the data space to which the first data set belongs, so that it can be used as source data in transfer learning for the purpose of training the prediction model of the first data set. The machine learning apparatus according to claim 1.

10. The experimental data is measurement data relating to a chemical reaction represented by a chemical formula consisting of the relationship between the target reactants, products, and catalyst. The aforementioned calculation data corresponds to the elementary processes of a catalytic reaction. The machine learning apparatus according to claim 3.