Machine Learning Material Discovery for Target Properties
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current material development techniques rely heavily on intuition and experience for determining search ranges in high-throughput computational screening, making it difficult to efficiently find new material candidates with target performance, especially due to the complexity of large calculation data sets.
Innovation Solution
A method and apparatus utilizing machine learning to generate structure candidates of new materials with target physical properties by analyzing descriptors, physical properties, and structures of existing materials, determining factors and structural factors, and applying encoding and decoding functions to predict and generate new structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If high-throughput computational screening is performed using conventional methods, then material candidates can be evaluated faster than direct synthesis, but the search range determination relies on researcher experience and intuition making it difficult to find new material candidates with target performance
Solution Approach 1:
The patent introduces machine learning models as an intermediary between the computational screening system and the material database. The ML models learn from existing material data and guide the search process, enabling the system to identify promising material candidates with target performance more accurately while maintaining high evaluation speed through automated screening.
Solution Approach 2:
The patent transforms the search strategy by changing from experience-based parameter selection to data-driven parameter optimization. Machine learning algorithms analyze patterns in material descriptors and properties to automatically determine optimal search ranges and screening criteria, improving both the speed and accuracy of material discovery.
2Productivity
If the search range is determined based on researcher experience and intuition, then the screening process can be performed, but it requires repeatedly performing search range setting and computational screening multiple times to obtain target performance
Solution Approach 1:
The patent performs preliminary actions by pre-training machine learning models on extensive material databases before the actual screening process. The ML models learn optimal search strategies and material-property relationships in advance, enabling single-pass or minimal-iteration screening that achieves target performance without repeated search range adjustments.
Solution Approach 2:
The patent implements feedback mechanisms where machine learning models continuously learn from screening results and refine their predictions. The system analyzes outcomes from computational screening and uses this feedback to improve subsequent searches, reducing the number of iterations needed to discover materials with desired properties.
3Quantity of substance
If a large amount of calculation data is generated from computational screening, then comprehensive material evaluation is achieved, but it is difficult to directly analyze and determine the next search region
Solution Approach 1:
The patent extracts meaningful patterns and features from large computational data sets using machine learning techniques. The ML models identify key descriptors and relationships that are most relevant to target material properties, extracting essential information from vast amounts of calculation data without requiring manual analysis of the entire data set.
Solution Approach 2:
The patent replaces manual data analysis with automated machine learning systems. Instead of researchers directly analyzing large computational data sets to determine next search regions, ML algorithms automatically process the data, identify patterns, and suggest optimal search directions, significantly reducing analysis complexity while maintaining data completeness.
Data Source
AI summary
A structure-generating method for generating a structure candidate of a new material including: by a structure-generating processor: performing machine learning on a machine learning model, wherein the machine learning model is configured to provide a result based on a descriptor of a material, a physical property of the material, and a structure of the material; and generating a structure candidate of the new material based on the result of the machine learning, wherein the new material has a target physical property, and wherein the descriptor of the material, the physical property of the material, and the structure of the material are stored in a database.