Data Science Approaches for Electrochemical Engineers: An Introduction through Surrogate Model Development for Lithium-Ion Batteries

Abstract

Data science, hailed as the fourth paradigm of science, is a rapidly growing field which has served to revolutionize the fields of bio-informatics and climate science and can provide significant speed improvements in the discovery of new materials, mechanisms, and simulations. Data science techniques are often used to analyze and predict experimental data, but they can also be used with simulated data to create surrogate models. Chief among the data science techniques in this application is machine learning (ML), which is an effective means for creating a predictive relationship between input and output vector pairs. Physics-based battery models, like the comprehensive pseudo-two-dimensional (P2D) model, offer increased physical insight, increased predictability, and an opportunity for optimization of battery performance which is not possible with equivalent circuit (EC) models. In this work, ML-based surrogate models are created and analyzed for accuracy and execution time. Decision trees (DTs), random forests (RFs), and gradient boosted machines (GBMs) are shown to offer trade-offs between training time, execution time, and accuracy. Their ability to predict the dynamic behavior of the physics-based model are examined and the corresponding execution times are extremely encouraging for use in time-critical applications while still maintaining very high (∼99%) accuracy. Data science, also known as data-intensive scientific discovery, is hailed as the fourth paradigm of science.1 A field focused on extracting knowledge or understanding from data, it includes the subdomains of machine learning, classification, data mining, databases, and data visualization. In the age of internet-scale data, these techniques are not only powerful, but also necessary to extract the signal from the noise and to have the throughput to do so in a reasonable amount of time. It has revolutionized the fields of bio-informatics, climate science, word recognition, advertising, medicine, and is finding more applications daily. In Google's Translate application, substantial improvements over previous methods were achieved using artificial neural network (ANN) structures, making 60% fewer errors than the previous state-ofthe-art algorithm.2 In climate science, where models are sophisticated and numerous, data science techniques are used to determine which of 20 models will give the best prediction on future and historical data, the accuracy of which surpasses the accuracy of the average of all models, the current benchmark.3 As chemical engineers are increasingly tasked with the analysis of more complex data sets, these same data science tools which have revolutionized other fields become more relevant. 4 When data sets grow, they must be managed intentionally in order to be useful. Data management, a subfield of data science, fills this role and gives the tools to be able to correct for missing data points, ensure consistency of the data, and transform the content of the data such that it is suitable for use in other aspects of data science....

On the Creation of a Chess-AI-Inspired Problem-Specific Optimizer for the Pseudo Two-Dimensional Battery Model Using Neural Networks

“…In order to reduce this large dimensionality to something more reasonable, a simple one-at-a-time sensitivity analysis was performed. A range for each parameter was established based on a combination of literature values 9 and model solution success, and an initial parameter set was chosen. For each of the parameters, the values were permuted from the initial value to the upper and lower bounds.…”

Machine Learning Approaches for Designing Mesoscale Structure of Li-Ion Battery Electrodes

“…Recently, parametric studies and sensitivity analyses of the process parameters, using artificial neural networks (ANNs) combined with physico-chemical models, have been carried out [18,19]. ANNs are brain-inspired systems, which are one of the main machine learning tools.…”

Machine Learning Approaches for Designing Meso-scale Structure of Li-ion battery Electrode

“…Although the calculation load is lower in these models, the mesoscale porous structures of the electrodes are not reflected because these structures are approximated as homogeneous solid and liquid state. On the other hand, the three-dimensional extended simulations using 3D-SEM tomography data [12,13] or artificial structure [14,15] Recently, parametric studies and sensitivity analyses of the process parameters using Artificial Neural Networks (ANNs) combined with the Physico-chemical models have been carried out [16,17]. In these papers, it has been proved that ANNs are useful tool for regression of the battery performance and analysis of the importance features.…”