微调统一了基础机器学习原子间潜能架构的初始准确性
Jonas Hänseroth1, Aaron Flötotto1, Muhammad Nawaz Qaisrani1
1Theoretical Solid State Physics, Institute of Physics, Technische Universität Ilmenau, 98693 Ilmenau, Germany.
The journal of physical chemistry letters
|March 4, 2026
概括
微调机器学习的原子间潜力 (MLIP) 显著提高了各种化学系统的精度. 这种专门的培训方法确保在各种MLIP架构中实现一致的,接近ab initio的预测.
科学领域:
- 计算化学计算化学
- 材料科学 材料科学 材料科学
- 机器学习 机器学习
背景情况:
- 机器学习的原子间潜力 (MLIP) 提供了高效的原子模拟.
- 一般用途的MLIP显示了与高精度初始方法的架构依赖偏差.
- 需要普遍准确且计算效率高的MLIP.
研究的目的:
- 为了证明微调可以改变基本的MLIP以达到接近ab initio的准确性.
- 在各种MLIP架构和化学化合物中对微调的有效性进行基准测试.
- 引入一个可重复的微调工作流程的工具包.
主要方法:
- 在七个不同的化合物上对五个领先的MLIP框架 (MACE,GRACE,SevenNet,MatterSim,ORB) 进行基准测试.
- 使用来自ab initio分子动力学轨迹的数据集进行微调.
- 评估力和能量预测与初始参考数据对比.
主要成果:
- 微调可以使力预测提高5-15倍,能量精度提高2-4倍.
- 专门的系统特定微调消除了依赖于架构的偏差.
- 微调可以将力误差减少一个数量级,并协调跨架构的性能.
结论:
- 微调是实现MLIP系统特定精度的通用方法.
- 这种方法保留了MLIP的计算效率.
- aMACEing工具包有助于广泛采用微调工作流程.
更多相关视频
05:37Rapid in-silico Battery Electrolyte Electrochemical Reaction Generation using 3T-VASP Multi-Scale Energy Minimization
Published on: August 22, 2025
742
13:56Probe Type II Band Alignment in One-Dimensional Van Der Waals Heterostructures Using First-Principles Calculations
Published on: October 12, 2019
8.4K
相关概念视频
Electronic Structure of Atoms
29.6K
An atom comprises protons and neutrons, which are contained inside the dense, central core called the nucleus, with electrons present around the nucleus. Taking into account the wave–particle duality of electrons and the uncertainty in position around the nucleus, quantum mechanics provides a more accurate model for the atomic structure. It describes atomic orbitals as the regions around the nucleus where electrons of discrete energy exist, characterized by four quantum...
29.6K
Predicting Molecular Geometry
46.6K
VSEPR Theory for Determination of Electron Pair Geometries
46.6K
Hybridization of Atomic Orbitals I
68.7K
The mathematical expression known as the wave function, ψ, contains information about each orbital and the wavelike properties of electrons in an isolated atom. When atoms are bound together in a molecule, the wave functions combine to produce new mathematical descriptions that have different shapes. This process of combining the wave functions for atomic orbitals is called hybridization and is mathematically accomplished by the linear combination of atomic orbitals. The new orbitals that...
68.7K
Atomic Orbitals
46.1K
An atomic orbital represents the three-dimensional regions in an atom where an electron has the highest probability to reside. The radial distribution function indicates the total probability of finding an electron within the thin shell at a distance r from the nucleus. The atomic orbitals have distinct shapes which are determined by l, the angular momentum quantum number. The orbitals are often drawn with a boundary surface, enclosing densest regions of the cloud.
46.1K
Hybridization of Atomic Orbitals II
49.9K
sp3d and sp3d 2 Hybridization
49.9K
Atomic Structure
89.0K
The Greek philosopher Democritus proposed that everything on Earth is made up of tiny particles called atomos, Greek for "indivisible," from which the modern term "atom" is derived. In the 19th century, John Dalton proposed the atomic theory that is still largely correct today. He put forth five postulates to explain how atoms made up the world around us. (1) All matter is composed of infinitely small particles or atoms. (2) All atoms of a given element are identical to one...
89.0K
