Material generation apparatus and material generation method

Patent metadata

Field	Value
Publication number	US-10885231-B2
Application number	US-201715622104-A
Country	US
Kind code	B2
Filing date	Jun 14, 2017
Priority date	Jul 12, 2016
Publication date	Jan 5, 2021
Grant date	Jan 5, 2021

Abstract

A genetic algorithm controller that controls respective processes using a genetic algorithm is configured. The processes include generation of a crystal structure of an inorganic material, a mutation operation of a crystal structure, a crossing-over operation of a crystal structure, structural relaxation calculation of a crystal structure, calculation of a predictive value of an objective function, selection and weeding out of a crystal structure based on a predictive value of an objective function, observation of an objective function value of a crystal structure by first-principle calculation, update of a regression model based on a result of observing the objective function value, and end determination for a material generation process.

First claim

What is claimed is:

1. A material generation apparatus comprising:
a genetic algorithm controller that controls respective processes using a genetic algorithm, the processes including generation of a crystal structure of an inorganic material, a mutation operation of a crystal structure, a crossing-over operation of a crystal structure, structural relaxation calculation of a crystal structure, calculation of a predictive value of an objective function, selection and weeding out of a crystal structure based on a predictive value of an objective function, observation of an objective function value of a crystal structure by first-principle calculation, update of a regression model based on a result of observing the objective function value, and end determination for a material generation process; and
a neighborhood set generator that includes
a mutation unit that generates an N-fold crystal structure of a crystal structure, and adds an atom of a randomly selected element to coordinates at which a distance to a nearest neighbor atom is largest,
a mutation unit that generates an N-fold crystal structure of a crystal structure, and deletes an atom of which a distance to a nearest neighbor atom is smallest, and
a crossing-over unit that selects two crystal structures, divides each of the crystal structures by a section determined by random numbers, and combines the two crystal structures by one internal coordinate expression.

2. The material generation apparatus according to claim 1 , wherein
the neighborhood set generator further includes
a mutation unit that selects a replacement target atom of a selected crystal structure by element replacement-weighted roulette selection, selects a replacement element by replacement-weighted roulette selection from the same replacement group as a replacement group of an element of the replacement target atom, and replaces the replacement target atom by an atom of the replacement element.

3. The material generation apparatus according to claim 1 , wherein
the neighborhood set generator further includes
a mutation unit that selects a lattice constant change target crystal structure to randomly select a change target crystal lattice constant parameter p, generates a random number k of −e to f (e and f are real numbers), and changes the lattice constant parameter by setting p′=p*(1+k).

4. The material generation apparatus according to claim 1 , wherein
gene data of a crystal structure used in each process controlled by the genetic algorithm controller includes a lattice vector (a v , b v , c v ), lattice constants (a, b, c, α, β, γ), the number of atoms contained in a crystal lattice, and contained atom information obtained by repeating a combination of an element type and a position vector representing an atom position in a lattice vector expression for each atom contained in the crystal lattice a number of times corresponding to the number of contained atoms.

5. The material generation apparatus according to claim 4 , wherein
the gene data includes a lattice vector, the number of atoms contained in a crystal lattice, and contained atom information obtained by repeating a combination of an element type and a position vector representing an atom position in a lattice vector expression for each atom contained in the crystal lattice a number of times corresponding to the number of contained atoms, excluding lattice constants (a, b, c, α, β, γ).

6. The material generation apparatus according to claim 1 , further comprising:
a structural relaxation calculation unit that performs structural relaxation calculating of calculating a stable position of an atom from a force received by each atom in a crystal structure with respect to a crystal structure of a neighborhood set generated by the mutation unit and the crossing-over unit included in the neighborhood set generator, and correcting gene data of the crystal structure by a calculation result.

7. The material generation apparatus according to claim 1 , wherein
an occurrence probability of an operation by the mutation unit and the crossing-over unit included in the neighborhood set generator varies an occurrence probability of each operation depending on the number of atoms inside a crystal structure selected as an operation target.

8. The material generation apparatus according to claim 1 , wherein
the crossing-over unit included in the neighborhood set generator is a unit that selects two crystal structures, rotates the crystal structures such as aspects of the both crystal structures are matched with each other by sorting orders of lattice lengths or cross-sectional areas of the respective crystal structures to dispose the crystal structures in an analysis space on a computer, generates one or two random numbers of 0 to 1, cuts the respective crystal structures according to one common random number or two random numbers, and combines the two crystal structures by one internal coordinate expression.

9. The material generation apparatus according to claim 6 , further comprising:
a neural network calculation unit that inputs an output of the structural relaxation calculation unit using a deep convolutional neural network configured such that a predictive value of an objective function is output by setting a calculation result of structural relaxation calculation of a crystal structure to an input, and outputs the predictive value of the objective function.

10. The material generation apparatus according to claim 9 , wherein
an N-fold cell of a crystal structure is created and divided by an equally spaced grid, a cube around a grid including each atom which is included in a target crystal structure is formed and converted into 1D arrangement data of each cube in a grid arrangement order, and data of the crystal structure is input to the deep convolutional neural network.

11. A material generation method comprising:
generating a current generation crystal structure set of an inorganic material by user definition or random generation at a time of start according to control of a genetic algorithm;
generating a crystal structure of a new neighborhood set by a mutation operation of a crystal structure and a crossing-over operation of a crystal structure from the current generation crystal structure set;
performing structural relaxation calculation on each crystal structure of the neighborhood set;
calculating a feature amount with respect to each crystal structure after the structural relaxation calculation;
calculating a predictive value of an objective function by inputting the calculated feature amount to a regression model;
performing selection and weeding out of a crystal structure based on the predictive value of the objective function;
repeating a loop from the generating of the crystal structure of the neighborhood set to the performing of the selection and weeding out of the crystal structure predetermined number of times;
observing an objective function value by first-principle calculation or with reference to experimental data with respect to a crystal structure set;
updating a regression model based on a result of observing the objective function value; and
determining whether to end or continue processing by determining whether the result of observing the objective function value satisfies a predetermined condition.

12. The material generation method according to claim 11 , further comprising:
inputting each crystal structure after the structural relaxation calculation to a deep convolutional neural network to obtain a predictive value of an objective function from an output of the neural network,
instead of the calculating of the feature amount and the calculating o

Assignees

• Hitachi Ltd

Inventors

• Matsumura Tadayuki
• ASARI Yusuke

Classifications

• G06F2111/06

  Multi-objective optimisation, e.g. Pareto optimisation using simulated annealing [SA], ant colony algorithms or genetic algorithms [GA] · CPC title

• G06F30/00 Primary

  Computer-aided design [CAD] · CPC title

• G06N3/086 Primary

  using evolutionary algorithms, e.g. genetic algorithms or genetic programming · CPC title

• G16C20/30

  Prediction of properties of chemical compounds, compositions or mixtures · CPC title