Active Learning Sampling for Structure-Energy Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional active learning methods struggle to sample training data that are similar in structure but different in energy, leading to reduced accuracy in energy prediction values for neural networks.
Innovation Solution
An active learning program that determines training data by extracting structural features and energy features, using a neural network to select data points with similar structures but varying energies, thereby enhancing the accuracy of energy predictions while reducing the number of data pieces.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional active learning using geometric distances is used to sample training data, then the diversity of material structures is considered, but data that is similar in structure but different in energy cannot be sampled effectively
Solution Approach 1:
The patent extends the sampling criterion from considering only structural features to a combined feature space that includes both structural features and energy prediction values. This dimensional extension allows the system to capture data points that are similar in structure but different in energy, which are critical for improving energy prediction accuracy while maintaining structural diversity coverage.
Solution Approach 2:
The patent changes the sampling parameters by incorporating energy prediction values as an additional dimension in the feature space. This parameter change transforms the sampling strategy from purely structure-based to a hybrid structure-energy based approach, enabling effective selection of training data that improves energy prediction accuracy.
2Measurement precision
If a large amount of labeled training data is prepared for neural network training, then the accuracy of energy prediction values is improved, but the computation workload and cost of DFT simulations increase significantly
Solution Approach 1:
The system uses the trained neural network to predict energy values for unsampled data points, and these predictions are fed back into the sampling criterion to identify the next most valuable training samples. This self-service mechanism allows the system to intelligently select a minimal subset of data points that will most effectively improve model accuracy, avoiding the need for large-scale DFT simulations.
Solution Approach 2:
The patent implements a feedback loop where energy prediction values from the neural network are used to refine the sampling criterion. This feedback mechanism enables the system to iteratively improve model accuracy by selecting training data points that maximize information gain, thereby reducing the overall computation workload required compared to traditional approaches.
3Quantity of substance
If the number of training data pieces is reduced to lower computation workload, then the computational cost is reduced, but the accuracy of energy prediction values deteriorates
Solution Approach 1:
The patent changes the selection parameters by incorporating energy prediction values as an additional dimension in the feature space. This allows the sampling criterion to identify data points that are similar in structure but different in energy, ensuring that even with a limited number of training samples, the selected data maximally improves energy prediction accuracy.
Solution Approach 2:
By extending the sampling criterion to a combined feature space of structural features and energy prediction values, the system can select training data that covers both structural diversity and energy variation. This dimensional extension ensures that fewer training samples are needed to achieve high accuracy, as each selected point provides maximum information gain in the combined feature space.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
An active learning program causes a computer to execute a process including: extracting a first feature related to a structure of each of a plurality of materials by inputting a plurality of structure data associated one with each of the plurality of materials into an active learning neural network; obtaining a second feature related to energy of the structure of each of the plurality of the materials, using the active learning neural network, the second feature being based on the first feature; and determining, based on the first feature and the second feature of each of the plurality of materials, one or more structure data to be training data (2) for training an energy prediction neural network for predicting energy of a material from among the plurality of structure data.