计算机科学
机器学习
人工智能
纳米孔
采样(信号处理)
从头算
分子动力学
先验与后验
样品(材料)
量子
集成学习
罕见事件
蒙特卡罗方法
虚拟筛选
能量(信号处理)
渲染(计算机图形)
物理系统
统计物理学
灵活性(工程)
瞬态(计算机编程)
势能
离解(化学)
作者
Sudheesh Kumar Ethirajan,Ambarish R. Kulkarni
标识
DOI:10.1021/acs.jctc.5c01191
摘要
Machine learning potentials (MLPs) can help bridge the length- and time-scale gaps required to study diverse physicochemical phenomena in nanoporous materials with ab initio accuracy. These MLPs are typically trained on quantum chemical data obtained from traditional molecular dynamics (MD) simulations that predominantly sample near-equilibrium configurations. This often limits the model's ability to describe high-energy, off-equilibrium states required to study rare events. To address this capability gap, we introduce a novel active learning curriculum using the On-the-fly Probability Enhanced Sampling (OPES) method. Using imidazole diffusion in the SALEM-2 metal-organic framework (MOF) as a prototypical example, we apply time-dependent OPES biases with temperature- and distance-based collective variables to systematically sample the most relevant regions of the potential energy surface. This strategy enables nanosecond-scale MD simulations with near-DFT accuracy, capturing both the expected diffusion pathway across the 6-membered window and a previously unreported ring-opening mechanism through the more constrained 4-membered window. The latter process, which involves the transient dissociation of a Zn-N bond, cannot be captured using classical force fields and is prohibitively expensive with DFT. Thus, this study highlights how enhanced sampling methods can overcome data scarcity challenges associated with training MLPs for studying rare events in nanoporous materials.
科研通智能强力驱动
Strongly Powered by AbleSci AI