发表机构
Lawrence Livermore National Laboratory; The University of Utah; Universidad de Buenos Aires（劳伦斯利弗莫尔国家实验室; 犹他大学; 布宜诺斯艾利斯大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文提出用元胞自动机从高保真模拟数据学习规则，生成任意大的格点起始构型，以桥接分子动力学与动力学蒙特卡洛的模拟长度尺度，并在阴离子交换膜上验证其性能。
AI 中文摘要
多尺度模拟需要耦合在不同特征长度和时间尺度上运行的物理模型，因为没有任何单一方法能够覆盖大多数感兴趣问题所需的范围。转向更粗粒化的表示可以解锁更长的长度和时间尺度，但它丢弃了构建形态的微观相互作用。细粒度模型可以从任意堆积初始化，并让其自组装成具有物理意义的结构；而低分辨率模型则不能，它必须从更高保真度的模拟中继承其起始构型。因此，粗粒化模型可访问的长度尺度并非由粗粒化方法本身决定，而是由能够经济地平衡的最大细粒度构型决定。一个代表性例子是从基于粒子的分子动力学（MD）升级到基于格点的表示，如动力学蒙特卡洛（kMC）。在这项工作中，我们提出了一种元胞自动机（CA）方法，用于生成任意大的格点起始构型。CA非常适合这项任务：短程局部规则驱动格点的演化，其重复应用产生涌现的长程有序，这在主题上反映了MD中短程相互作用如何产生自组装形态。我们使用来自更高保真度模拟的构型作为训练集，并通过逻辑回归从中学习CA规则。作为原理验证，我们为水合阴离子交换膜（AEM）开发了这些规则，生成新的起始构型，并在介观kMC模拟中将其性能与MD衍生的“地面真值”进行基准比较。然后，我们展示了生成更大格点的能力，并表明它们在kMC中的行为与较小的CA基准构型一致。
英文摘要
Multiscale simulation requires coupling physics models that operate at different characteristic length and time scales, because no single method spans the range needed for most problems of interest. Moving to a coarser-grained representation unlocks longer length and time scales, but it discards the microscopic interactions that build morphology. A fine-grained model can be initialized from an arbitrary packing and left to self-assemble into a physically meaningful structure; a lower-resolution model cannot, and must inherit its starting configuration from a higher-fidelity simulation. The length scales accessible to the coarse-grained model are therefore set not by the coarse-grained method itself, but by the largest fine-grained configuration that can be affordably equilibrated. A representative example is the scale-up from particle-based molecular dynamics (MD) to a lattice-based representation such as kinetic Monte Carlo (kMC). In this work, we present a cellular automata (CA) approach for generating arbitrarily large lattice starting configurations. CA is a natural fit for this task: short-ranged local rules drive the evolution of a lattice, and their repeated application gives rise to emergent long-range order, thematically mirroring how short-ranged interactions in MD produce self-assembled morphology. We use a configuration from a higher-fidelity simulation as a training set and learn the CA rules from it via logistic regression. As a proof of principle, we develop these rules for a hydrated anion exchange membrane (AEM), generate new starting configurations, and benchmark their performance in mesoscale kMC simulations against an MD-derived "ground truth." We then demonstrate the ability to generate substantially larger lattices and show that their behavior in kMC is consistent with that of the smaller CA benchmark configurations.
