Neural Network Chemical Structure Generation via Expression Region Modification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for generating chemical structures using neural networks face challenges in efficiently modifying partial structures to meet specific property requirements, particularly in optimizing properties like transmission wavelength and emission wavelength.

Innovation Solution

The method involves using a deep neural network with layer-wise relevance propagation (LRP) and genetics algorithms to determine and modify expression regions in chemical descriptors or images, applying Gaussian noise to pixel values, and iteratively generating new chemical structures until desired property values are achieved.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If neural network is used to generate chemical structures, then chemical structure generation capability is improved, but efficiency in modifying partial structures to meet specific property requirements deteriorates

Engineering Contradiction:
Improvechemical structure generation capabilityVSAvoidefficiency in modifying partial structures
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent segments the chemical structure into partial structures and identifies specific expression regions that contribute to target properties. By using LRP to locate critical substructures and then applying genetic algorithms to modify only those specific regions rather than entire molecules, the system maintains high adaptability while improving modification efficiency.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If LRP technique is applied to determine expression region, then property expression accuracy is improved, but computational complexity increases

Engineering Contradiction:
Improveproperty expression accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

Instead of analyzing the entire chemical structure uniformly, the LRP technique applies local quality assessment by propagating relevance scores to identify specific pixels or substructures that most contribute to the target property. This focused approach improves property expression accuracy while the subsequent genetic algorithm efficiently handles the modification of only these identified critical regions.

Inventive Principle:
Principle #3Local quality

3Manufacturing precision

If genetic algorithm is used to modify partial structure, then property optimization is improved, but number of iterations required increases

Engineering Contradiction:
Improveproperty optimizationVSAvoidnumber of iterations
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by first using LRP to identify and mark the critical expression regions before initiating the genetic algorithm. This pre-identification step ensures that the genetic algorithm only needs to search and modify within these predetermined critical regions rather than the entire molecular structure, thereby achieving property optimization with fewer iterations and reduced computational time.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3614314B1Method and apparatus for generating chemical structure using neural network
Publication Date: 2024.09.04 SAMSUNG ELECTRONICS CO LTD
  • EP3614314B1 patent drawingFigure 1
  • EP3614314B1 patent drawingFigure 2
  • EP3614314B1 patent drawingFigure 3~4

AI summary

Generating a new chemical structure by using a neural network using an expression region that expresses a particular property in a descriptor or an image for a reference chemical structure. The new chemical structure may be generated by changing a partial structure in the reference chemical structure that corresponds to the expression region.