	US 12,406,752 B2
	System for generating compound structure representation
Yoshihiro Osakabe, Tokyo (JP); and Akinori Asahara, Tokyo (JP)
Assigned to HITACHI, LTD., Tokyo (JP)
Appl. No. 17/919,804
Filed by Hitachi, Ltd., Tokyo (JP)
PCT Filed Apr. 9, 2021, PCT No. PCT/JP2021/015042 § 371(c)(1), (2) Date Oct. 19, 2022, PCT Pub. No. WO2021/220774, PCT Pub. Date Nov. 4, 2021.
Claims priority of application No. 2020-079790 (JP), filed on Apr. 28, 2020.
Prior Publication US 2023/0117325 A1, Apr. 20, 2023
Int. Cl. G16C 20/90 (2019.01); G16C 20/70 (2019.01); G16C 20/80 (2019.01)

CPC G16C 20/90 (2019.02) [G16C 20/70 (2019.02); G16C 20/80 (2019.02)]

8 Claims

1. A system for generating a compound structure representation, the system comprising:

one or more processors; and

one or more storage devices,

wherein each of the one or more storage devices stores a structure model, a structure-property relationship model, a compound structure representation of each of one or more known materials, and one or more target values of each of one or more types of physical property values,

wherein the structure model includes:

a first encoder that converts the compound structure representation to a real number vector; and

a first decoder that estimates the compound structure representation from the real number vector resulting from the conversion by the first encoder,

wherein the structure-property relationship model includes:

a second encoder that converts, to a real number vector, an input vector including, as components, the real number vector generated by the first encoder and a target value vector including the target values of the one or more types of physical property values; and

a second decoder that estimates the input vector from the real number vector generated by the second encoder,

wherein each of the one or more processors generates, using the first encoder of the structure model, one or more structure generation vectors on the basis of each of the compound structure representation of each of the one or more known materials and the one or more target values of each of the one or more types of physical property values,

wherein each of the one or more structure generation vectors includes, as components, the real number vector of the compound structure representation of one of the known materials generated by the first encoder and the target value vector including the target values of each of the one or more types of physical property values,

wherein each of the one or more processors inputs, to the structure-property relationship model, each of the one or more structure generation vectors, extracts, from an output of the second decoder of the structure-property relationship model, the real number vector corresponding to the compound structure representation, and inputs the extracted real number vector to the first decoder of the structure model and generates a novel compound structure representation,

wherein each of the one or more storage devices includes, in addition to the one or more target values of each of the one or more types of physical property values, one or more target values of each of one or more other types of physical property values,

wherein the structure-property relationship model includes a plurality of auto-encoders,

wherein the first auto-encoder included in the plurality of auto-encoders includes the second encoder and the second decoder,

wherein each of encoders included in the plurality of auto-encoders and other then the first auto-encoder is interposed between the encoder and a decoder of another of the auto-encoders,

wherein an input to each of the auto-encoders other than the first auto-encoder includes, as components, a real number vector from the decoder of the other auto-encoder and the target value vector including the target values of each of the one or more types of physical property values selected from among the one or more other types of physical property values,

wherein each of the one or more storage devices stores training data for the structure-property relationship model,

wherein each of the one or more processors trains the structure-property relationship model by using the training data,

wherein the training data includes a plurality of groups to be used to train the plurality of respective auto-encoders,

wherein each of the plurality of groups associates each of a plurality of compound structure representations with measurement values of one or more predetermined types of physical property values,

wherein, of two groups included in the plurality of groups, one group having a larger number of types of physical property values includes all the types of physical property values and all the compound structure representations of another group having a smaller number of types of physical property values, and

wherein the group having the larger number of types of the physical property values is used to train the inner auto-encoder in the structure-property relationship model.