Charge-Transfer Interatomic Potentials for Scalable Hamiltonians
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning interatomic potentials (MLIPs) face challenges in scaling to large systems, capturing long-range effects, and accounting for discontinuities and transitions, leading to inaccuracies and computational inefficiencies in simulations.
Innovation Solution
A machine learning model is trained to learn charge transfers within atomic systems, which are then used to construct atomic charges, ensuring local and global charge neutrality, and determine properties such as total energy, enabling scalable and accurate simulations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning models are used to predict interaction energies and forces, then accuracy of property prediction is improved, but scalability to large-scale systems deteriorates
Solution Approach 1:
The patent segments the charge distribution problem by introducing atomic charges as intermediate representations that can be computed locally for each atom based on its environment. This allows the system to handle large-scale systems by computing charges independently for each atom rather than requiring global calculations, thus improving scalability while maintaining prediction accuracy through the use of these localized charge representations.
Solution Approach 2:
The patent introduces atomic charges as an intermediary representation between the atomic structure and the interaction energies/forces. These atomic charges serve as a mediator that captures essential electronic structure information in a computationally efficient manner, enabling accurate property predictions for large systems without requiring expensive ab-initio calculations for every interaction.
2Measurement precision
If machine learning models capture short-range effects, then local accuracy is improved, but ability to capture long-range effects deteriorates
Solution Approach 1:
The patent makes the atomic charge representation universal by training the machine learning model to capture both short-range and long-range effects through a single unified approach. The atomic charges computed for each atom automatically incorporate information about both nearby and distant atoms through the training data, allowing the same model to accurately describe both local and long-range interactions without requiring separate models for different interaction ranges.
3Productivity
If atomic charges are used to determine total energy, then computational efficiency is improved, but charge neutrality constraints deteriorate
Solution Approach 1:
The patent implements feedback mechanisms to ensure charge neutrality is maintained throughout the computation. The machine learning model is trained with explicit constraints that enforce charge neutrality, and the atomic charges are adjusted based on feedback from the total energy calculation to maintain this constraint. This feedback approach allows the system to achieve both computational efficiency through the use of atomic charges and reliability through maintained charge neutrality.
Data Source
AI summary
Methods for a machine learning network that trains and subsequently executes one or more machine learning (ML) models are disclosed. The system described herein is configured to embed atomic positions and species of a given atomic system and apply those to ML model(s) to learn charge transfer properties and local energies. By constructing atomic charges from learned charge transfer properties, both local and global charge neutrality is ensured. The atomic charges are then used to generate an auxiliary Hamiltonian description. By combining both the auxiliary Hamiltonian description and the learned local energies, properties such as total energy of the atomic system are determined. By determining total energy from the auxiliary Hamiltonian description and the learned local energies, such methods ensure that long and short range effects are accounted for, while also appropriately enabling for realistic discontinuities and/or transitions within the potential energy surface.