Generative Structure-Property Inverse Co-Design for Material Discovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current material discovery processes using machine learning are limited by their reliance on prior knowledge and intuition, which can overlook suitable materials outside expert intuition or dissimilar to known materials, and are constrained by time requirements, allowing only a fraction of the material universe to be simulated.
Innovation Solution
A machine learning system that learns the underlying joint distribution of structure-property relationships using generative techniques, such as generative adversarial networks (GANs) or variational autoencoders (VAEs), to directly generate samples that meet target properties without additional processing steps, enabling the identification of suitable structures through conditional generation or filtering of randomly generated samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If forward design is used to simulate material properties, then materials similar to known materials can be identified, but materials outside expert intuition or dissimilar to known materials are overlooked
Solution Approach 1:
The patent applies inverse design by inverting the traditional forward design workflow. Instead of starting from material structure and predicting properties, the system starts from desired properties and generates candidate structures that could achieve those properties. This inversion allows the search to begin with the target property space and work backwards, enabling discovery of materials that may be dissimilar to known materials while still meeting the desired property criteria.
Solution Approach 2:
The patent utilizes generative models that can explore the material space by changing structural parameters systematically. The system generates diverse material structures by varying compositional and structural parameters, allowing it to escape from local optima and discover novel materials that differ significantly from known materials while maintaining the desired properties.
2Reliability
If comprehensive material simulation is performed to ensure property targets are met, then accurate material identification is achieved, but computational time requirements increase significantly
Solution Approach 1:
The patent employs a two-stage approach where generative models first pre-screen and generate candidate structures that are likely to meet property targets based on learned patterns from training data. This preliminary generation step filters out unlikely candidates before they undergo expensive detailed simulations, allowing comprehensive property verification to be performed only on a small subset of promising candidates rather than the entire material space.
Solution Approach 2:
The patent introduces generative models as an intermediary between the desired property specifications and the detailed material simulation process. These models act as a bridge that translates property requirements into candidate structural descriptions, enabling the system to prioritize which materials warrant detailed simulation based on predicted property matches, thus reducing the overall computational burden.
3Productivity
If only a fraction of the material universe is simulated due to time constraints, then computational resources are conserved, but the search space is limited and suitable materials may be missed
Solution Approach 1:
The patent employs generative models capable of exploring diverse regions of the material space by systematically varying compositional and structural parameters. The model can generate materials with different stoichiometries, crystal structures, and atomic arrangements, enabling comprehensive coverage of the material universe without requiring exhaustive simulation of every possible material configuration.
Solution Approach 2:
The patent transforms the material search problem from a high-dimensional exhaustive search into a guided exploration by introducing latent space dimensions through generative models. By mapping materials to and from a compressed latent representation, the system can efficiently navigate the material space, exploring diverse regions systematically while maintaining the ability to recover detailed material structures when needed.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
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.