Neural Network Adsorption Energy Training via Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge lies in accurately learning and distinguishing adsorption energy, which has a small value relative to the prediction error of the whole system, making it difficult to execute learning using machine learning methods like Neural Network Potential (NNP), especially due to the small magnitude of adsorption energy values.
Innovation Solution
A training device and inferring device are developed, utilizing a neural network model to input atomic structures of surfaces and adsorbed molecules, optimizing parameters based on errors calculated from DFT data, and employing periodic boundary conditions to enhance the learning of adsorption energy by differentiating it from system energy errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning methods like Neural Network Potential (NNP) are used to learn adsorption energy, then the ability to predict system energy is improved, but the small magnitude of adsorption energy (around 1 eV) becomes indistinguishable from the prediction error of the whole system, making learning difficult
Solution Approach 1:
The patent segments the energy calculation into multiple components: total system energy, reference energy (from DFT calculations), and adsorption energy. By separating the adsorption energy calculation from the total system energy and using DFT-calculated reference values, the method isolates the small adsorption energy signal from the large total energy, making it distinguishable from prediction errors and enabling effective machine learning training.
2Measurement precision
If DFT calculation data is used for training, then the ground truth energy values are obtained, but the computational cost and time required increase significantly
Solution Approach 1:
The patent performs preliminary DFT calculations to obtain accurate reference energy values for training data preparation before conducting machine learning training. By pre-calculating these reference values using high-accuracy DFT methods, the patent creates a robust training dataset that enables subsequent machine learning models to achieve high prediction accuracy without requiring repeated expensive DFT calculations during the training process.
Data Source
AI summary
A training device includes processor. The processor inputs a first atomic structure including a surface and an adsorbed molecule close to the surface into a model to obtain an energy outputted from the model in response to the input, and obtains a first error based on the outputted energy of the first atomic structure and a ground truth value of the energy of the first atomic structure, input a fourth atomic structure including a cluster and an adsorbed molecule close to the cluster into the model to obtain an energy outputted from the model in response to the input, and obtains a fourth error based on the outputted energy of the fourth atomic structure and a ground truth value of the energy of the fourth atomic structure, and update a parameter of the model by the first and the fourth error. The surface and the cluster include the same atom.