Neural Network Potential Model Segmentation for Physical Property Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current atomic simulation methods using Neural Network Potentials (NNPs) face challenges in accurately predicting physical properties beyond energy and force, requiring more efficient models for complex calculations.
Innovation Solution
An information processing device is configured with a neural network model that includes layers from an input layer to a predetermined intermediate layer of a pre-trained NNP, with additional layers and training methods like transfer learning to output different physical property values, such as adsorption or activation energy, at a lower computational cost.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a full Neural Network Potential model is trained to predict multiple physical properties, then prediction accuracy improves, but computational cost and model complexity increase
Solution Approach 1:
The patent segments the neural network model into a pre-trained base model (containing layers from input layer to a predetermined intermediate layer) and a separate fine-tuning component. This segmentation allows the base model to be reused for multiple physical property predictions without retraining the entire model, thereby reducing computational cost while maintaining prediction accuracy across different properties like energy, force, and thermodynamic quantities.
Solution Approach 2:
The patent applies preliminary action by pre-training the neural network model on a comprehensive dataset of physical properties before actual use. The pre-trained intermediate layers capture general patterns and relationships that can be leveraged for predicting specific physical properties without requiring full model retraining, thus reducing computational requirements while preserving accuracy.
2Productivity
If a pre-trained NNP model is used for multiple physical property predictions, then computational cost decreases, but the model's ability to accurately predict specific properties may be compromised
Solution Approach 1:
The patent applies local quality by selectively fine-tuning only the output layer or subsequent layers of the pre-trained model for specific physical properties, while keeping the earlier intermediate layers fixed. This allows the model to maintain the general knowledge from pre-training (ensuring computational efficiency) while adapting the local output layer to accurately predict specific properties such as adsorption energy, activation energy, or solubility.
Solution Approach 2:
The patent creates a universal pre-trained NNP model that can serve multiple functions for predicting different physical properties. The pre-trained intermediate layers form a universal feature extractor that works for various properties, and by adding property-specific output layers, the same base model becomes multi-functional, improving computational efficiency without sacrificing accuracy for any single property.
3Measurement precision
If the neural network model includes more layers for better physical property prediction, then prediction accuracy improves, but training time and computational resources increase
Solution Approach 1:
The patent segments the training process into a pre-training phase (completing the base model with intermediate layers) and a fine-tuning phase (adding output layers for specific properties). This segmentation allows the computationally intensive part to be done once during pre-training, and subsequent predictions require only minimal computational resources, reducing training time for new properties while maintaining accuracy.
Solution Approach 2:
The patent performs preliminary action by pre-training the neural network model's intermediate layers on a broad range of physical properties before actual application. This preliminary training captures essential patterns and relationships, so that when the model is applied to specific properties, only minimal additional training is needed, significantly reducing training time compared to training from scratch while maintaining high accuracy.
Data Source
AI summary
An information processing device includes one or more memories and one or more processors. The one or more processors are configured to input information regarding an atom of a substance to a first model; and obtain information regarding the substance from the first model. The first model is a model which includes: layers from an input layer up to a predetermined layer of a second model to which information regarding atoms is input and which outputs at least one of a value of an energy or a value of a force; and another layer, and which is trained to output the information regarding the substance.