Material Property Prediction Using Riemannian Multi-Task Transfer Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing transfer learning techniques struggle with small, complex data sets such as molecular structure data, particularly in non-Euclidean spaces, leading to ineffective prediction performance and stability issues.
Innovation Solution
A method and system using a pre-trained integrated prediction model that applies Riemannian geometry for geometric alignment in an integrated latent space to transfer knowledge across tasks, incorporating multiple loss functions and material property relationships for improved prediction performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If existing Euclidean space-based transfer learning techniques are applied to molecular structure data, then the model can be trained on large-scale data sets, but the prediction performance deteriorates on small, complex data sets due to inability to handle non-Euclidean spaces
Solution Approach 1:
The patent changes the mathematical framework from Euclidean space to Riemannian manifold, transforming how data is represented and processed. This allows the model to handle the curved geometry of molecular data spaces, improving prediction performance on small data sets while maintaining scalability to large data sets through the unified manifold learning approach.
Solution Approach 2:
The patent introduces a new dimensional perspective by mapping data from Euclidean space to a Riemannian manifold with curved geometry. This dimensional transformation enables the model to capture complex relationships in molecular structure data that are invisible in traditional flat spaces, resolving the contradiction between data quantity and prediction reliability.
2Adaptability or versatility
If transfer learning is applied to predict multiple material properties sequentially, then the model can handle each property individually, but the complex coupling relationships between properties are lost
Solution Approach 1:
The patent merges multiple material property prediction tasks into a single unified model that processes all properties simultaneously. By combining the prediction of multiple properties within one model framework that operates on the Riemannian manifold, the model preserves the coupling relationships between properties while maintaining adaptability to predict any individual property.
Solution Approach 2:
The patent creates a universal prediction model that can handle multiple material properties through a single Riemannian manifold framework. This multi-functional model captures the inherent relationships between different properties (such as boiling point, melting point, density) while maintaining the ability to predict each property independently, thus resolving the information loss problem.
3Loss of information
If a single integrated prediction model is trained for multiple material properties, then the coupling relationships between properties are preserved, but the model requires large amounts of data which are difficult to obtain experimentally
Solution Approach 1:
The patent applies preliminary action through pre-training the model on large-scale molecular data before applying it to specific prediction tasks. This pre-training phase allows the model to learn general patterns and coupling relationships from abundant data, and then these learned representations can be transferred to specific tasks with limited experimental data, effectively reducing the data requirement for each individual prediction task.
Solution Approach 2:
The patent uses the concept of copying by creating a pre-trained model that captures general material property relationships, then applying this copied knowledge to specific prediction tasks. The pre-trained model serves as a template that can be adapted to new tasks with minimal additional data, effectively copying the learned patterns rather than requiring complete retraining from scratch.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for predicting a plurality of material properties, according to an embodiment of the present invention, pertains to a method by which a computing system including a memory and a processor predicts characteristics regarding the plurality of material properties, wherein the method includes: obtaining experimental data including characteristics data on the plurality of material properties regarding materials; pre-training an integrated prediction model for a plurality of tasks of predicting characteristics regarding the plurality of material properties from the obtained experimental data; inputting, to the pre-trained integrated prediction model, material information to be predicted; outputting, by the integrated prediction model, a characteristic value for each of the plurality of material properties in regard to the material information; and providing the output characteristic value for each of the plurality of material properties.