Functional data-driven framework for fast forecasting of electrode slurry rheology simulated by molecular dynamics 
Marc Duquesnoy, Teo Lombardo, Fernando Caro, Florent Haudiquez, Alain C. Ngandjong, Jiahui Xu, Hassan Oularbi & Alejandro A. Franco
npj Computational Materials 8:161 (2022)
doi.org/10.1038/s41524-022-00819-2
Published online:22 july 2022
Abstract| Full Text | PDF OPEN

摘要:基于机理模型的锂离子电池复合电极制造过程的计算模拟可以捕捉到制造参数对电极性能的影响。但是，确保这些性能与实验数据相匹配通常需要付出高昂的计算成本。在这项研究中，我们提出了一个功能数据驱动框架，旨在解决这一昂贵的过程。首先，我们通过检索从分子动力学模拟中计算出的早期数值来预测计算结果是否可能与我们的实验值范围相匹配；接着，在第二步中，我们恢复正在进行中的模拟的附加数值，以预测其最终结果。我们将这种方法应用于计算电极浆料粘度，并在各种电极化学的情况下进行了演示，预期的机理模拟结果可以比完整模拟快11倍，同时准确度可达R2score =0.96。 　　

Abstract:The computational simulation of the manufacturing process of lithium-ion battery composite electrodes based on mechanistic models allows capturing the influence of manufacturing parameters on electrode properties. However, ensuring that these properties match with experimental data is typically computationally expensive. In this work, we tackled this costly procedure by proposing a functional data-driven framework, aiming first to retrieve the early numerical values calculated from a molecular dynamics simulation to predict if the observable being calculated is prone to match with our range of experimental values, and in a second step, recover additional values of the ongoing simulation to predict its final result. We demonstrated this approach in the context of the calculation of electrode slurries viscosities. We report that for various electrode chemistries, the expected mechanistic simulation results can be obtained 11 times faster with respect to the complete simulations, while being accurate with a R2 score equals to 0.96. 

Editorial Summary

Functional data-driven framework 

在人工智能 (AI) 和机器学习 (ML) 应用蓬勃发展的现代世界中，人们可能会认为使用基于物理过程数学描述的传统机理模型已经过时了。然而，事实并非如此，因为机理模型仍然是当今支持复杂系统分析的基本工具，在医学、能源存储、纳米技术、生物学和环境科学等众多领域中无处不在。锂离子电池 (LIBs) 的制造过程包括相互关联的多个步骤和参数，涉及浆料制备、涂覆和干燥、电极的压延、电池组装、电解液渗透和固体电解质界面的形成。尽管用于密切匹配实验数据的数值方法不断改进，但这些方法的执行通常是时间和资源的消耗。机器学习技术可以很容易地与机理模型相结合，帮助研究人员摆脱纯粹的试错实验方法，促进性质计算，并简化机理模型的参数化，以降低其计算成本。来自法国亚眠大学Franco教授领导的研究团队，开发了一种功能性数据驱动框架，使用分子动力学模拟来快速预测电极浆料的流变学特性，从而实现成本的降低。该框架通过执行机理模拟的第一个数值步骤（即时间帧）来检索早期数值，并利用功能主成分分析（FPCA）和K最近邻（KNN）两个算法，压缩时间序列并执行预测任务，避免了完整的模拟过程。该框架分为筛选步骤和预测步骤：筛选步骤的主要目标是识别能够提供与实验数据可比较结果的运行模拟，以验证电极浆料的粗粒度分子动力学（CGMD）机理模型；预测步骤只考虑之前在感兴趣范围内过滤的模拟，快速预测非平衡分子动力学（NEMD）结果。作者将此框架应用于基于NEMD的电极浆料建模，并在三种不同的LIB电极活性材料化学组成上进行仿真，使框架扩展到不同的材料。通过跟踪计算粘度值（η）沿着模拟过程的演变，可以确定剪切-粘度曲线（γ-η曲线）。尽管这项工作主要讨论了LIB电极浆料的情况，但该框架也可以应用于其他领域，例如，在机理模型用于生成时间序列数据的情况下，实现计算成本的显著降低，并加快优化过程。

In a modern world where Artificial Intelligence (AI) and Machine Learning (ML) applications are blooming, one may think that the use of more traditional mechanistic models based on mathematical descriptions of physical processes is becoming obsolete. However, this is not true since mechanistic models still represent nowadays essential tools to support the analysis of complex systems, which are used in numerous domains such as medicine, energy storage, nanotechnology, biology, and environmental sciences. The manufacturing process of lithium-ion batteries (LIBs) encompasses multiple steps and parameters which are interlinked. Such steps concern the slurry preparation, its coating and drying, the calendering of the resulting electrode, the cell assembly, the electrolyte infiltration, and the formation of the solid electrolyte interphase. This complex process has been historically simulated using empirical models with parameters fitted considering experimental trends or by using mechanistic models. Despite the continuous improvement in the numerical methods used to closely match the experimental data, the execution of these methods is usually time and resources consuming. However, ML techniques can be easily combined with mechanistic models and help researchers move away from pure trial-and-error experimental approaches, to facilitate property calculations, and to ease the parameterization of mechanistic models. A team led by Prof. Franco Université de Picardie Jules Verne, France, presented a functional data-driven framework for fast predictions of mechanistic simulation results and tackled the issue of computational cost reduction of 3D-resolved mechanistic models for electrode slurry rheology simulation with molecular dynamics. This framework bases its operation on only executing the first numerical steps (i.e., time frames) of the mechanistic simulation to retrieve early numerical values, and then bypassing the full simulation process by predicting its final results without the need to run it until the end. More precisely, the aforementioned framework proposes first a screening step, whose main goal is to identify running simulations that will end with a result in a range of interest for our manufacturing modeling, e.g., expected to provide results comparable with the experimental data, in order to validate the coarse-grained molecular dynamics (CGMD) mechanistic model of the electrode slurry. Second, it proposes a forecasting step to quickly predict the non-equilibrium molecular dynamics (NEMD) results, considering only the previously filtered simulations within the range of interest. Both steps couple two algorithms: one based on Functional Principal Component Analysis (FPCA) achieving compression of the time series in a low dimensional space (i.e., dimensionality reduction), and another one based on K-Nearest-Neighbors (KNN) performing the predictive task. We applied this framework for electrode slurry modeling based on NEMD after accumulating simulations over three different active materials of LIB electrode, making the framework extensive to different materials. The authors tracked the evolution of the calculated viscosity values (η) along the simulation process to determine the shear-viscosity curve (γ-η curve). It is calculated point by point, i.e., a shear rate is applied through the deformation of the simulated slurry box to define the time series. Despite the particular illustration here for the case of LIB electrode slurries, the proposed framework can be also applied to other fields where mechanistic models are employed to generate time series data providing a significant computational cost reduction, but also making feasible a faster optimization process.