Compound Structure Generation With Latent Property Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional virtual screening methods are limited to interpolative searches within training data ranges and struggle to generate novel materials with physical property values exceeding known materials, requiring large datasets for learning grammar rules and property relationships.
Innovation Solution
A system utilizing a structure model and structure-property relationship model, including variational auto-encoders, to generate novel compound structures with desired physical properties using a smaller number of data sets, by converting chemical structure representations to real number vectors and extracting latent variables.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a physical property estimation model using a neural network is applied to randomly generated chemical structure formulas, then candidates with desired physical property values can be identified, but the method is limited to interpolative search within training data ranges and cannot generate novel materials with physical property values exceeding known materials
Solution Approach 1:
The system segments the learning task into two distinct models: a structure model that learns grammar rules and representation forms of chemical structures, and a structure-property relationship model that learns the mapping between structures and physical properties. This segmentation allows each model to specialize, enabling the structure-property relationship model to perform extrapolative searches beyond training data ranges while maintaining reliability through the structure model's grammar validation.
Solution Approach 2:
The structure model acts as an intermediary between the generation process and the property estimation. It converts randomly generated chemical structure formulas into properly formatted representations that satisfy grammar rules, thereby enabling reliable property estimation for novel structures that extend beyond the training data distribution.
2Adaptability or versatility
If both grammar rules for chemical structure representation and physical property relationships are learned by a single model, then comprehensive capability is achieved, but a large amount of data including sets of chemical structure formulas and physical property values is required
Solution Approach 1:
The learning task is divided into two separate models with distinct training data requirements. The structure model learns only grammar rules from chemical structure data, while the structure-property relationship model learns physical property relationships. This segmentation reduces the data burden for each individual model, as neither needs to learn both grammar and property relationships simultaneously.
Solution Approach 2:
The structure model serves multiple functions: it validates grammar rules for chemical structures and generates properly formatted representations for property estimation. This multi-functionality allows the system to achieve comprehensive capability without requiring a single model to learn everything, thereby reducing overall data requirements.
3Productivity
If conventional virtual screening methods are used, then chemical structure formulas with desired physical property values can be identified, but the efficiency of novel material discovery is reduced due to the need for large datasets and limited search capability
Solution Approach 1:
The system architecture is segmented into a structure model and a structure-property relationship model, each with specialized functions. This segmentation improves productivity by enabling efficient novel material discovery through extrapolative search, while the modular design manages complexity through clear separation of concerns and independent training processes.
Data Source
AI summary
A structure model includes a first encoder that converts a 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. A structure-property relationship model includes a second encoder that converts, to a real number vector, an extended vector including, as components, the real number vector generated by the first encoder and a target value vector including target values of one or more physical property values and a second decoder that estimates the extended vector from the real number vector generated by the second encoder.