With the widespread use of rechargeable batteries like lead-acid, nickel-cadmium, nickel-metal hydride, and lithium-ion batteries (LIBs) in portable electronics and electric vehicles, battery researchers are increasingly interested in improving the performance of lithium-ion battery materials and accurately predicting battery state.
Machine learning, the core technology of artificial intelligence (AI), has emerged as a new technique for addressing current battery challenges, such as battery materials design and accurate battery state prediction. Machine learning is thought to help accelerate the adoption and improvement of lithium-ion batteries on a large scale.
The research and development of traditional materials are mainly based on experience and repeated trial and error experiments, consisting of seven stages: discovery, development, property optimization, system design and integration, certification, manufacturing, and deployment. Interdisciplinary research teams would carry out these different stages. They usually take 10-20 years, which is quite a long time, and cannot meet the rapidly developing demand of the market. Besides, the cost of experiments and calculations can be extremely high.
Thanks to AI, ML offers new ways to solve problems in battery research, although at present, the application of ML in the field of battery is in the preliminary stage. This article provides a brief introduction to ML and the general process of its algorithmic implementation in LIBs.
ML in battery applications
1. Screening the battery materials
All-solid-state lithium-ion batteries (ASSLBs) are considered the next generation of energy storage devices with advantages such as high safety and high energy density. They are currently a popular research direction in the field of LIBs. Solid-state electrolytes are a key part of ASSLBs, for which they need to meet: high ionic conductivity, stable interfacial properties, and other requirements. Selecting suitable materials is currently the main technical bottleneck in developing solid-state battery technology.
ML can rapidly screen out suitable candidates from large material databases based on the required electrolyte properties, greatly reducing experimental cycles and costs, and can effectively solve the problem of difficult material selection.
Many open-source material databases such as the Inorganic Crystal Structure Database, Materials Project, and Total Materia have been established by major research institutes, including energy bands, energy gaps, crystal structures, and fundamental physical properties data. ML models can identify materials 1000 times faster, greatly speeding up the high ionic conductivity material selection process.
2. Prediction of material properties
For battery materials, there are potentially thousands of compounds that can be synthesized by inserting metal ions (Li, Na, etc.) or changing the molar fraction of the constituent elements, many of which have not yet been synthesized. It is even more difficult to characterize the materials, so battery research needs to combine previous experimental data to predict the properties of new materials more accurately.
ML models can predict fitting material properties such as thermodynamic stability of arbitrary chemicals, binding energies of multiple compounds, melting temperatures of simple component solids, coordination energies of alkali group metals in battery electrolytes, and ionic conductivity of electrolyte materials, and the electrode voltage of lithium-ion batteries.
3. Calculating the optimum composition of composite battery materials
Pure component battery materials are generally difficult to meet the needs of battery design, so they need to be used in the form of multiple components or additives, which improves the electrochemical performance of the material. Experiments and DFT calculations are difficult to search for the multicomponent ratio or the number of additives with the best performance from a complex composition space. The introduction of ML can significantly reduce the number of calculation iterations and effectively solve the material improvement problem. The combination of high-throughput experiments and ML offers the possibility to explore the optimal solution for the design space of complex systems.
4. Material microstructure analysis
The microstructure of materials, such as crystal structure and polymer molecular weight, also influences materials’ physical and chemical properties. Using ML to assist in the statistical analysis of the patterns between microstructure and material physical and chemical properties can further guide the design of batteries and improve battery performance.
Common ML algorithms used in battery applications
- Naïve Bayes – Stable classification efficiency, good performance for small data sizes, less sensitive to real data, simpler algorithms
- Logistic regression – Simple implementation, low computational effort, high speed, and low storage resources
- Linear regression – Simple to implement, simple to calculate
- K-nearest neighbor – No assumptions on data, high accuracy, can be used for nonlinear classification; mature theory
- Decision trees – Computationally simple, easy to understand, and highly interpretable; able to handle uncorrelated features and better suited to handle uncorrelated features
- Support vector machine – Can solve high-dimensional problems; can handle interactions of nonlinear features; can improve generalization
- Artificial neural networks – High classification accuracy, robust and fault-tolerant to noisy nerves, and able to adequately approximate complex nonlinear relationships
- Random forests – Regression and classification can be performed. No need to adjust parameters repeatedly. No scaling of data required
- Convolutional neural networks – Parameter sharing and sparse connections result in a significant reduction in training parameters. With translational invariance
- Recurrent neural networks – Deep models in the time dimension can model sequence content.