Machine Learning Chemical Structure Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for designing chemical structures with multiple intended chemical and physical properties are inefficient, relying on trial-and-error experimentation and limited machine learning techniques that require pre-defined material structures or are restricted to inorganic materials, making it difficult to explore the vast parameter space of undiscovered materials.
Innovation Solution
A method and system that utilize machine learning to predict chemical structures by receiving intended structural and chemical property values, constructing a prediction model using dimension reduction and regression methods, and automatically configuring chemical structure candidates, enabling the prediction of structures with multiple intended properties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional trial-and-error experimentation and chemical simulation are used for material design, then new materials with new physical and chemical properties can be produced, but the process becomes very time-consuming and not conductive to exploring the vast parameter space of undiscovered materials
Solution Approach 1:
The patent replaces conventional mechanical trial-and-error experimentation with machine learning-based prediction systems. The system uses trained models to predict material properties from compositional data, substituting physical experimentation with computational prediction, thereby dramatically reducing time consumption while maintaining design accuracy.
Solution Approach 2:
The patent creates virtual copies of material systems through computational models. Instead of physically synthesizing and testing materials, the system uses machine learning models to simulate and predict material properties, allowing exploration of the vast parameter space without time-consuming physical experiments.
2Productivity
If machine learning methods base predictions on structural information of materials, then prediction efficiency improves, but the structures of materials must first be determined before machine learning can take place
Solution Approach 1:
The patent inverts the conventional approach by training machine learning models to predict material structure from desired properties, rather than predicting properties from known structures. This inversion allows the system to work backwards from target properties to identify candidate material compositions, eliminating the need for pre-determined structures and improving overall process efficiency.
3Measurement precision
If regression methods are used for chemical structure prediction, then the system receives structural feature vectors and predicts chemical features, but the output is restricted to scalar values and cannot predict multiple chemical features simultaneously
Solution Approach 1:
The patent implements a universal prediction framework that can simultaneously predict multiple chemical and physical properties from a single material composition input. The system uses ensemble methods and multi-output regression models that can handle multiple target properties concurrently, making the system versatile for designing materials with multiple desired characteristics.
4Reliability
If kernel methods based on similarity search are used for organic material structure prediction, then existing material structures can be analyzed, but the prediction is limited to structures with chemical values that range only in existing materials
Solution Approach 1:
The patent employs parameter change strategies that allow the system to explore beyond the range of existing material data. By using techniques such as data augmentation, transfer learning, and exploration-exploitation balancing in the prediction algorithm, the system can reliably predict novel material structures with properties extending beyond the training data range, enabling discovery of new material compositions.
Data Source
AI summary
A method and system are provided for predicting chemical structures. The method includes receiving, at a user interface, intended structural feature values and intended chemical property values, as vectors. The method further includes constructing, by a hardware processor, a prediction model, wherein the prediction model predicts other structural feature values from the intended structural feature values and the intended chemical property values, and automatically configuring, by the hardware processor, at least one chemical structure candidate from the other structural feature vectors. The method additionally includes evaluating the at least one chemical structure candidate to determine structural feature values and chemical property values of the at least one chemical structure candidate and performing, by the hardware processor, machine learning of a chemical structure predicting system based on the evaluated structural feature values and the evaluated chemical property values of the at least one chemical structure candidate.