神经网络对分子晶体潜力的知识蒸
1Center for Data Science, Waseda University, 1-6-1 Nishiwaseda, Shinjuku-ku, Tokyo 169-8050, Japan. takuya.taniguchi@aoni.waseda.jp.
Faraday discussions
|September 18, 2024
概括
这项研究通过使用知识蒸来增强有机分子晶体的神经网络潜力 (NNP). 用软目标进行微调可以提高预测晶体性质的准确性.
科学领域:
- 材料科学 材料科学 材料科学
- 计算化学计算化学
- 机器学习 机器学习
背景情况:
- 有机分子晶体提供了多样化的功能,但通过计算来探索它们是具有挑战性的.
- 量子化学计算是准确但缓慢的;神经网络潜力 (NNP) 提供速度,但风险偏差来自无机训练数据.
研究的目的:
- 为了提高有机分子晶体的预训练NNP的准确性.
- 调查知识蒸策略,从教师模型转移信息.
主要方法:
- 微调预先训练的NPP,使用从教师模型中提炼知识.
- 比较软目标 (教师推断值) 与硬目标 (基本真相) 对NNP准确性的影响.
- 评估蒸NNP在预测有机晶体弹性特性方面的表现.
主要成果:
- 仅使用软目标的知识蒸是提高NNP准确性的最有效方法.
- 增加硬目标损失的比例导致知识传输效率下降和过度装配.
- 与原始预训练模型相比,蒸的NNP在预测弹性特性方面表现出更好的准确性.
结论:
- 知识蒸,特别是软目标,是适应NNP用于有机分子晶体的可行策略.
- 这种方法减轻了偏差,并提高了对材料性质的预测准确度.
- 开发的NNP显示了加速发现新型有机晶体材料的前景.
更多相关视频
08:55Methods of Ex Situ and In Situ Investigations of Structural Transformations: The Case of Crystallization of Metallic Glasses
Published on: June 7, 2018
8.5K
09:15Growing Protein Crystals with Distinct Dimensions Using Automated Crystallization Coupled with In Situ Dynamic Light Scattering
Published on: August 14, 2018
10.5K
相关概念视频
Molecular and Ionic Solids
17.0K
Crystalline solids are divided into four types: molecular, ionic, metallic, and covalent network based on the type of constituent units and their interparticle interactions.
Molecular Solids
Molecular crystalline solids, such as ice, sucrose (table sugar), and iodine, are solids that are composed of neutral molecules as their constituent units. These molecules are held together by weak intermolecular forces such as London dispersion forces, dipole-dipole interactions, or hydrogen bonds, which...
Molecular Solids
Molecular crystalline solids, such as ice, sucrose (table sugar), and iodine, are solids that are composed of neutral molecules as their constituent units. These molecules are held together by weak intermolecular forces such as London dispersion forces, dipole-dipole interactions, or hydrogen bonds, which...
17.0K
Recrystallization: Solid–Solution Equilibria
1.1K
Recrystallization is a purification technique used to separate impurities from solid compounds. In this technique, no chemical reactions occur. Instead, it exploits physical properties only, specifically, the solubility differences between the desired compound and impurities, either at a single temperature or at different temperatures, and under other selected conditions. The solid-solution equilibrium (solubility equilibrium) of each component in the solution represents a binary phase...
1.1K
Crystal Growth: Principles of Crystallization
1.7K
Crystallization is a phase transformation process in which crystals are precipitated from a supersaturated solution or formed from other sources. During crystallization, atoms or molecules arrange themselves into a well-defined, rigid crystal lattice to minimize energy.
Initiating crystallization involves manipulating the concentration of the solute and the temperature of the solution. Since crystal growth occurs when the ratio of concentration and solubility of the solute in the solvent...
Initiating crystallization involves manipulating the concentration of the solute and the temperature of the solution. Since crystal growth occurs when the ratio of concentration and solubility of the solute in the solvent...
1.7K
Polymer Classification: Crystallinity
2.8K
Unlike ionic or small covalent molecules, polymers do not form crystalline solids due to the diffusion limitations of their long-chain structures. However, polymers contain microscopic crystalline domains separated by amorphous domains.
Crystalline domains are the regions where polymer chains are aligned in an orderly manner and held together in proximity by intermolecular forces. For example, chains in the crystalline domains of polyethylene and nylon are bound together by van der Waals...
Crystalline domains are the regions where polymer chains are aligned in an orderly manner and held together in proximity by intermolecular forces. For example, chains in the crystalline domains of polyethylene and nylon are bound together by van der Waals...
2.8K
Molecular Models
38.0K
Physical models representing molecular architectures of chemical compounds play essential roles in understanding chemistry. The use of molecular models makes it easier to visualize the structures and shapes of atoms and molecules.
38.0K
