Neural Network Potential Training with 2-Body Energy Constraints

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing Molecular Dynamics (MD) simulation methods using Neural Network Potentials (NNPs) do not effectively incorporate 2-body potential functions, leading to inaccuracies in reproducing energy values and stability issues, such as unstable hydrogen molecules at incorrect bond lengths.

Innovation Solution

Incorporating 2-body potential data into the training process of neural network models for NNP, using both compound data and 2-body potential curves to enhance the accuracy of energy and force predictions, allowing for better interpolation and extrapolation capabilities.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If existing MD simulation methods use NNP without 2-body potential functions, then the simulation can be performed with neural network models, but the energy values become inaccurate and stability issues occur

Engineering Contradiction:

Improvestability of simulationVSAvoidaccuracy of energy values

Core Design Contradiction:

ReliabilityVSMeasurement precision

Solution Approach 1:

The patent combines 2-body potential functions with neural network potentials to create a hybrid interatomic potential model. This merging allows the system to leverage both the physical rigor of 2-body potentials and the flexibility of NNP, resolving the contradiction between simulation stability and energy accuracy by integrating both approaches into a unified framework

Inventive Principle:

Principle #5Merging (Combining)

2Measurement precision

If 2-body potential functions are added to the potential function, then the reproducibility of interatomic potential improves, but the complexity of the model increases

Engineering Contradiction:

Improvereproducibility of interatomic potentialVSAvoidcomplexity of potential function model

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the interatomic potential into distinct components: 2-body potential terms and neural network potential terms. This segmentation allows each component to be optimized independently for its specific function, improving overall reproducibility while managing model complexity through modular structure that can be selectively applied

Inventive Principle:

Principle #1Segmentation

3Measurement precision

If neural network models are trained without 2-body potential data, then the training process is simpler, but the accuracy of energy and force predictions decreases

Engineering Contradiction:

Improveaccuracy of energy and force predictionsVSAvoidcomplexity of training process

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies preliminary action by incorporating 2-body potential data into the training process before final model deployment. This preliminary inclusion of physical constraints during training establishes a more accurate energy landscape foundation, improving prediction accuracy while the modular training approach manages complexity through staged processing

Inventive Principle:

Principle #10Preliminary action

Data Source

PatentUS20240111998A1Inferring device, inferring method, and training device

Publication Date: 2024.04.04 PREFERRED NETWORKS INC

AI summary

An inferring device includes one or more memories; and one or more processors. The one or more processors are configured to input information on each atom in an atomic system into a second model to infer a difference between energy based on a first-principles calculation corresponding to the atomic system and energy of an interatomic potential function corresponding to the atomic system.