Artificial Intelligence Investigation of NMC Cathode Manufacturing Parameters Interdependencies

Abstract

The number of parameters involved in lithium-ion battery electrode manufacturing and the complexity of the physicochemical interactions throughout the associated processes make highly complex to find interdependencies between the final electrode characteristics and the fabrication parameters. In this work, we have analyzed three different machine-learning algorithms (decision tree, support vector machine, and deep neural network) in order to find the best one to uncover the interdependencies between the slurry manufacturing parame-ters and the final properties of NMC-based cathodes. The results revealed that the support vector machine method shows high accuracy and the possibility to predict the influence of manufacturing parameters on themass loading and porosity of the electrodes in a straightforward graphical way. Furthermore, we report for the first time this new approach and a case study that, by comparing the trends observed experimentally and from the model, demonstrates the validity and the quality of the proposed approach.[a] R.

How Machine Learning Will Revolutionize Electrochemical Sciences

“…(b) Example of a classification machine learning algorithm (Support Vector Machine) able to predict the impact of the percentage of NMC active material, solid-to-liquid ratio, and viscosity of the slurry on the final porosity of a lithium ion battery positive electrode. [Reprinted with permission from ref ( 38 ). Copyright 2019 Wiley-VCH GmbH.]…”

Accelerating Battery Manufacturing Optimization by Combining Experiments, In Silico Electrodes Generation and Machine Learning

“…This is of major importance for the electrode manufacturing process, as during the coating and drying stages the process parameters should be controlled to reach an optimal porosity. As we showed in our previous work, 32 this can be modulated not only by the slurry AM/CB/binder relative composition but also with the amount of solvent used during the slurry preparation. In general, a sufficient 14 amount of the latter is needed for a proper dispersion of the solid particles, dissolution of the polymer and processability of the slurry during the coating stage.…”

“…The blooming Artificial Intelligence (AI) field promises to accelerate the manufacturing optimization by revealing patterns hardly recognizable by "classical" analysis methods. 14,[32][33][34][35][36][37] As they do not rely on physical models, the feasibility of this approach depends on the capability to generate high quality datasets (from experiments, physical models or both of them simultaneously) complete enough to describe the battery manufacturing, which most likely represents the limiting step to develop AI models. Takagishi et al recently reported a machine learning (ML) approach in which the datasets were built by randomly generating in silico electrode mesostructures composed of only AM particles coupled with a zerodimensional electrochemical model to calculate the charge/discharge specific resistance.…”

“…27,28,39 ML algorithms combined with experimental data were recently proved by us to be suitable for discovering and assessing the interdependencies between manufacturing parameters and electrode properties for the case of electrode slurry features and coating conditions. 32 An ideal strategy for a systematic use of ML-based approaches should combine the intrinsic reliability of experiments and physics-based modelling with the high throughput of stochastically generated electrodes. This work aims to introduce a novel methodology (Figure 1), hereafter called hybrid methodology, encompassing experiments and/or physics-based modelling together with data-driven electrode mesostructures generation and ML algorithms to speed up manufacturing optimization of either wellknown or novel LIBs chemistries.…”

Materials Informatics Screening of Li‐Rich Layered Oxide Cathode Materials with Enhanced Characteristics Using Synthesis Data

“…Albeit this assumption can be seen as unreasonable it allowed us to significantly enhance the results of modeling. The descriptors describing the synthesis process are efficiently used in virtual screening applications in materials science . Heat‐treatment information including the temperature and time required for the calcination and sintering processes has been normalized to a scale of zero to one accepting the maximal and minimal known temperatures and time limits.…”