Molecular Structure Prediction Using Quantum-Trained Neural Potentials
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for predicting molecular structures, such as first-principle computational simulations and artificial neural networks, face challenges in handling large molecular sizes and particle deformations like dissolution, aggregation, or phase transition, requiring high computational loads and lacking efficient learning sets.
Innovation Solution
A method is developed to create a learning data set for an artificial neural network using eigenvector values and quantum mechanics calculations for various structural models, including monoatomic, bulk, slab, nanoparticle, defect, and amorphous models, to predict molecular structures and particle deformations efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If first-principle computational simulation based on quantum mechanics is used to predict molecular properties, then prediction accuracy is improved, but computational load increases significantly with molecular size and complexity
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing quantum mechanical properties (electronic energy, dipole moment, HOMO/LUMO energies) for various molecular structures in a training dataset before actual prediction tasks. This pre-computed data serves as the foundation for training the neural network, enabling fast predictions without repeated quantum calculations. The preliminary quantum mechanical computations are performed once to build the dataset, then the trained model provides rapid predictions for new molecules.
Solution Approach 2:
The patent uses copying by creating a neural network model that replicates the predictive capability of quantum mechanical calculations. Instead of performing actual quantum simulations for each new molecule, the system copies the essential predictive patterns from the training data through the neural network's learned weights and biases. This allows the model to mimic quantum mechanical prediction accuracy while avoiding the high computational costs of actual quantum simulations.
2Productivity
If conventional artificial neural network methods are used for prediction, then calculation speed is improved, but ability to handle particle deformation and dissolution is insufficient
Solution Approach 1:
The patent applies parameter changes by using different descriptor sets for different types of molecular systems. For bulk materials, it uses density-based descriptors; for molecules, it uses constitutional and topological descriptors; for surfaces, it uses coordination number descriptors. This adaptive selection of parameters based on the system type enables the neural network to accurately predict properties across diverse systems including deformed particles, dissolved substances, and phase transitions, while maintaining fast calculation speed.
Solution Approach 2:
The patent achieves universality by developing a single neural network framework that can predict properties of various molecular systems (bulk materials, molecules, surfaces, deformed particles) using appropriately selected descriptors. The same neural network architecture and prediction methodology are applied across different system types, with the flexibility to choose different descriptor sets based on the specific system being analyzed. This multi-functional approach enables the model to handle particle deformation, dissolution, and other complex behaviors while maintaining computational efficiency.
3Measurement precision
If QSPR with thousands of molecular control descriptors is used, then prediction accuracy is improved, but model complexity and computational requirements increase
Solution Approach 1:
The patent applies the taking out principle by extracting only the most relevant and essential descriptors for each type of molecular system, rather than using all possible thousands of QSPR descriptors. For bulk materials, it extracts density-based descriptors; for molecules, it extracts constitutional and topological descriptors; for surfaces, it extracts coordination number descriptors. This selective extraction of critical descriptors maintains prediction accuracy while dramatically reducing model complexity and computational requirements compared to using all available QSPR descriptors.
Data Source
AI summary
A method for predicting a molecular structure includes: preparing a learning data set including first learning data including an eigenvector value and a quantum mechanics calculation value for a monoatomic and molecular structural model, a bulk structural model, a slab structural model, and a nanoparticle structural model of a material including a plurality of elements; learning an artificial neural network using the learning data set to obtain a potential value; and predicting a molecular structure of another material by using the potential value.