Generative structure-property inverse computational co-design of materials
AssigneeSamsung Electronics Co., Ltd.
InventorsGanesh Hegde, Harsono S. Simka
Life of this patent
Estimating the term…
Abstract
A method and a system for material design utilizing machine learning are provided, where the underlying joint distribution p(S,P) of structure (S)-property (P) relationships is explicitly learned simultaneously and is utilized to directly generate samples (S,P) in a single step utilizing generative techniques, without any additional processing steps. The subspace of structures that meet or exceed the target for property P is then identified utilizing conditional generation of the distribution (e.g., p(P)), or through randomly generating a large number of samples (S,P) and filtering (e.g., selecting) those that meet target property criteria.
Claims20
A machine learning system to generate a structure (S) for a target property (P), the system comprising a generative adversarial network (GAN),
- wherein the GAN comprises:
- a latent space vector generator configured to generate random vectors,
- a generator network configured to receive the random vectors as input and to generate a tuple comprising both a simulated structure and a simulated property, and
- a discriminator network configured to be trained from a database comprising real structures and real properties, to distinguish the simulated structure and the simulated property from the real structure and the real property, and to send a loss function to the generator network,
- wherein the latent space vector generator, the generator network, and the discriminator network are arranged to learn an underlying joint probability distribution of structure-property p(S, P) simultaneously.
A method to generate a structure (S) for a target property (P) utilizing machine learning, the method comprising:
- training a discriminator network with a database comprising real structures and real target property,
- generating a simulated structure and a simulated property through a generator network,
- training the generator network so that the discriminator network fails to distinguish a difference between the simulated structure and the real structure from the database and learning an underlying joint probability distribution of structure-property p(S, P) simultaneously, and
- generating the structure for the target property utilizing the underlying joint probability distribution.
A machine learning system to generate a structure (S) for a target property (P), the system comprising a variational autoencoder (VAE),
- wherein the VAE comprises an encoder network, a latent space, and a decoder network,
- wherein the encoder network is configured to receive a database comprising real structures and real properties as an input and compress the input to generate encoded vectors, mean and standard deviation of a distribution of the encoded vectors in the latent space,
- the decoder network is configured to be trained from the database comprising real structures and real properties, to decompress the encoded vectors to create simulated structure and simulated property, and to distinguish a difference between the simulated structure and the simulated property and the real structure and the real property, and to send a loss function to the encoder network,
- wherein the encoder network, the latent space, and the decoder network are arranged to learn an underlying joint probability distribution of structure-property p(S, P) simultaneously.
A method to generate a structure (S) for a target property (P) utilizing machine learning, the method comprising:
- inputting a database comprising real structures and real properties to an encoder network as an input and compressing the input to generate encoded vectors, mean and standard deviation of a distribution of the encoded vectors in a latent space,
- training a decoder network with the database comprising real structures and real properties,
- decompressing the encoded vectors to create simulated structure and simulated property by the decoder network,
- sending a loss function to the encoder network, and
- learning an underlying joint probability distribution of structure-property p(S, P) simultaneously.
Dependencies are inferred from the claim text.