INDEEDopt框架：反应力场参数优化方法

Fig. 1 Schematic representation of the optimization framework.

反应力场对所有化学系统使用固定函数，但不同化学系统中的不同应用需要优化大量参数来调整反应立场，其难度不言而喻。由美国宾夕法尼亚州立大学材料科学与工程系的Mert Y. Sengul教授和机械工程系的Adri C. T. van Duin教授领导的团队，提出了一个“初始设计增强的深度学习优化” (INDEEDopt) 框架，该框架不仅可以加速简单化学体系（如：三元、四元或五元组分）的力场开发，还可以使多组分体系（如：三元或更大组分）的开发成为可能。

Fig. 2 The influence of neural network size to prediction accuracy.

该框架不仅可以加速简单化学体系（如三元、四元或五元组分）的力场开发，还可以实现多组分体系（如六元或更复杂）的开发。随着高熵合金、无机-有机界面等材料的成功设计，机器学习引入材料研究已成为关键。本文提出的机器学习框架，有巨大能力对高度相关的高维数据进行建模，并调整初始设计算法以全面探索高维参数空间。INDEEDopt能够在参数空间中找到几个局部最小值，并产生优化的力场，用于模拟的数据产生阶段。INDEEDopt与传统ReaxFF参数化方法不同。INDEEDopt的一个重要特点是，它使用拉丁超设计（LHD）算法，有效地生成数据集并探索整个参数空间，而不是停留在最初指定的区域。迄今为止开发的大多数优化方法都需要猜测一个初始参数值；然而，并没有一定的规则来判断这些值是否合理。INDEEDopt方法则不需要初始值.事实上，它可以作为一种方法来生成这些初始值，通过与其他优化方法相结合来进一步优化力场参数。另一个关键的区别是，INDEEDopt是一个数据驱动的方法，产生出大量的数据可以用于进一步分析研究。INDEEDopt框架使用的程序是可并行的，并可以与任何ReaxFF-MD软件一起使用，无需修改代码。INDEEDopt作为一个概念，如果需要，将来可以适用于经典和其他反应力场的参数优化。该文近期发表于npj Computational Materials 7: 68 (2021).

Fig. 3 The flow diagram of INDEEDopt framework showing all stages of the algorithm.

Editorial Summary

INDEEDopt Framework: Optimized Reaction Force Field Parameters

The reaction force field （ReaxFF） uses a fixed function for all chemical systems, but different applications in different chemical systems require optimization of a large number of parameters to adjust the reaction position, which exists evidently a big difficult. A team lead by Prof. Mert Y. Sengul from Department of Materials Science and Engineering, The Pennsylvania State University, USA., and Prof. Adri C. T. van Duin from Department of Mechanical Engineering, The Pennsylvania State University, USA, presented an INitial-DEsign Enhanced Deep learning-based OPTimization (INDEEDopt) framework that will not only accelerate the force field development for simple chemical systems (e.g., ternary, quaternary, or quinary component), but also enable the development for multi-component ones (e.g., senary or larger), which have become crucial with the advances in materials discovery (e.g., high entropy alloys, inorganic–organic interfaces).

Fig. 4 The deep learning predictions with respect to corresponding the reference values.

The framework uses machine learning (ML) due to its capability of modeling highly correlated high-dimensional data and adapts an initial design algorithm to explore the high-dimensional parameter space comprehensively. Thus, INDEEDopt is capable of finding several local minima in parameter space and produces optimized force fields to be used in the production stage of simulations. INDEEDopt differs from other interesting ReaxFF parameterization methodologies in several aspects. An important feature of INDEEDopt is that it uses a Latin Hyper Design (LHD) algorithm to efficiently generate a data set and explore whole parameter space instead of being stuck around initially assigned regions. Most of the optimization methods developed to date require initial parameter value guesses; however, there is no certain rule to assign these values. INDEEDopt does not require the initial values; in fact, it can be used as a method to generate these values that can be further optimized by combining with other optimization methods. Another critical distinction is that INDEEDopt is a data-driven method and generates a significant amount of data, which can be used for further investigations.

Fig. 5 The comparison of error values calculated by INDEEDoptand conventional method.

The procedure that INDEEDopt framework uses is fully parallelizable and can be used with any ReaxFF-MD software without modifications to the code. INDEEDopt as a concept can be adapted to the parametrization of classical and other reactive force fields in the future if desired. This article was recently published in npj Computational Materials 7: 68 (2021).

Fig. 6 The comparison of Ni–Cr force field parameters optimized by different methods.

原文Abstract及其翻译

INDEEDopt: a deep learning-based ReaxFF parameterization framework （INDEEDopt：一个基于深度学习的 ReaxFF 参数化框架）

Mert Y. Sengul, Yao Song, Nadire Nayir, Yawei Gao, Ying Hung, Tirthankar Dasgupta & Adri C. T. van Duin

Abstract Empirical interatomic potentials require optimization of force field parameters to tune interatomic interactions to mimic ones obtained by quantum chemistry-based methods. The optimization of the parameters is complex and requires the development of new techniques. Here, we propose an INitial-DEsign Enhanced Deep learning-based OPTimization (INDEEDopt) framework to accelerate and improve the quality of the ReaxFF parameterization. The procedure starts with a Latin Hypercube Design (LHD) algorithm that is used to explore the parameter landscape extensively. The LHD passes the information about explored regions to a deep learning model, which finds the minimum discrepancy regions and eliminates unfeasible regions, and constructs a more comprehensive understanding of physically meaningful parameter space. We demonstrate the procedure here for the parameterization of a nickel–chromium binary force field and a tungsten–sulfide–carbon–oxygen–hydrogen quinary force field. We show that INDEEDopt produces improved accuracies in shorter development time compared to the conventional optimization method.

Fig. 7 Influence of temepature to lattice parameters for Cr (BCC) and Ni (FCC) phases.

摘要 经验性的原子间势需要优化力场参数，以调整原子间的相互作用，模拟基于量子化学方法得到的相互作用。参数的优化很复杂，需要开发新的技术。在此，我们提出了一个“初始设计增强的深度学习优化” (INDEEDopt) 框架来加速和提高 ReaxFF参数化的质量。该程序从拉丁超立方设计（LHD）算法开始，该算法用于广泛地探索参数环境。LHD将探索过的区域的信息传递给深度学习模型，该模型找到最小差异区域并消除不可行的区域，并构建对有物理意义的参数空间的更全面理解。我们在此演示了镍铬二元力场和钨硫碳氧氢五元力场的参数化过程。我们表明，与传统的优化方法相比，INDEEDopt在更短的开发时间内产生了更好的精度。

Fig. 8 The comparison of error values calculated by INDEEDoptand conventional method.

Fig. 9 The comparison of W–S–C–O–H force field parameters optimized by different methods.