On the road to optimized battery cells with artificial intelligence
The use of machine learning has great potential to improve the development and production of battery cell sustainably. That’s why at CustomCells we are testing smart technology to optimize production processes continuously.
Our world is moving at high speed toward an electrified future. By the end of the year, electric cars could account for a third of all newly registered passenger cars on Germany’s roads, according to estimates by management consultants EY. By 2031, the share of new electric vehicles is expected to be as high as 66 percent. In addition, the majority of the EU Parliament has voted to phase out the internal combustion engine in 2035, and the massive increase in fuel costs since the beginning of the year and the outbreak of the Ukraine war are likely to give this development additional momentum.
But it is not only mobility on the road that is being electrified. Our cities are growing. There’s a need for more sustainability in aviation. There’s a need for more sustainability on and under water. The digitization of more and more areas of everyday life and the economy ensures the widespread use of ever more new electrical solutions. The development and production of special lithium-ion battery cells are setting the pace for this comprehensive technological transformation. Whether it is the development of air cabs and uncrewed research submarines, the operation of novel medical solutions, or photovoltaic systems — none of these solutions would be possible without battery cells that are precisely adapted to their respective application-specific requirements.
Making complexity manageable
Battery cells are incredibly complex, and the number of possible parameters that determine the qualities of a cell is high. Anodes, cathodes, separators, electrolytes, and housing materials influence the chemical reactions inside the cell and determine, among other things, the charging properties and service life. Artificial intelligence, specifically machine learning, and other technologies can contribute to optimization and make complexity manageable, at least in part. For example, by independently recognizing which process parameters achieve desirable results, production systems help to improve manufacturing quality and increase production. In these cases, the system continuously compares the properties of the cell and its subcomponents with the parameter sets of the individual process steps.
What is essential to understand in the context of process optimization in battery cell manufacturing is that high scrap is still one of the biggest problems in the industry. The average first-time yield in battery cell manufacturing is sometimes put at 15 to 16 percent. End-of-line testing, which results are often only available days or weeks after manufacturing, is not suitable as the sole solution. Seamless monitoring of the entire value chain is therefore needed. Inline inspections using computer monitoring can make a significant contribution in this area. Combined with sensor data, statistical algorithms, and deep learning, machine monitoring can detect anomalies in individual production steps such as the coating and stacking process and allows predictive quality analytics approaches to make early statements about the subsequent performance of the cell. As part of a cyber-physical system, such solutions can be used in such a way that the machine itself continuously and automatically optimizes all individual processes. At CustomCells, we are currently testing for various individual processes and in cooperation with our partners in related research projects, which improvements can be achieved by using machine learning.
However, this is preceded by the development of an extensive data basis. After all, machine learning and other innovative technologies can only be as successful as a large, reliable database allows.
Depending on the specific process under consideration, CustomCells tests various machine learning approaches on the software side, which either rely on supervised learning methods such as Random Forest or use combinations of supervised and unsupervised methods. Because of the experimental character currently inherent to many of the optimization steps, using Python with the Scikit-Learn library, among others, is often a good way to quickly achieve initial results, based on which further concrete development steps can be planned. In addition, relevant deep learning frameworks such as TensorFlow or PyTorch also play a role.
Great potential for development
While attention is currently focused on machine learning for process optimization in production, the technology also offers enormous potential in development — especially in combination with a digital twin. For example, instead of developing application-specific battery cells using a broad matrix of experiments, various settings can be adjusted in a digital environment controlled by AI. This could be the correct configuration of the mixing unit for the electrode slurries or the discovery of innovative material combinations, for example.
However, the road to fully automated development is a long one, especially one which provides the suitable materials and cell design for the application-specific purpose at the push of a button and is constantly evolving by reinforcement learning. Numerous constraints must be taken into account during development. For example, an AI could use machine learning to identify promising electrode configurations. Appropriate approaches are already being pursued in research. In these cases, the AI learns to make predictions about the electrochemical properties of a wide range of materials. In the future, this should help speed up development processes. AI could then help, for example, pre-select materials or focus on materials that have received little attention to date.
However, such selection processes via machine learning are followed by a series of follow-up questions that such systems have so far mostly ignored: Are the relevant materials available at all? Is the use of the materials desirable? What is the cost-benefit ratio? Can production be scaled up in the event of subsequent series production? Can the materials be processed with the existing machines? A highly toxic material may be particularly suitable in the opinion of the AI, but it may still not be possible to use it.
The quality of models stands and falls with the availability of data
Another challenge in using machine learning models in development arises from the question of which data the companies and their respective AIs can draw on. Such approaches are particularly powerful when they do not work with isolated data sets of a company or research lab, but rather have various data and data sets at their disposal. The transfer of operational data, as is standard in many software applications today, would also be important in the field of mechanical engineering to accelerate further developments and enable innovations.
In reality, however, legal requirements impose strict limits on the exchange of data — especially when data is exchanged across different continents. It’s not only about the often-discussed legal differences between the EU and the USA, but also about the often-missing exchange between Asian buyers and European producers. A European supply chain, such as the one utilized to a large extent at CustomCells, must therefore always be viewed from the perspective of data availability.
The future view of the entire life cycle of a cell
As CustomCells, we are both developers and manufacturers of battery cells optimally suited to their respective applications. Marginal and borderline areas of physics, as well as chemistry are regularly exhausted for this purpose. However, this is one of the main challenges in the industry. It is not only a matter of developing solutions, but also of being able to scale them later in production. In this context, machine learning can significantly contribute to optimizing and stabilizing processes in the future. However, optimizing sub-processes and individual development steps through machine learning is just the beginning.
In the future, machine learning, combined with advanced statistical analysis methods, is likely to shape the entire lifecycle of a battery cell — from development, production, and subsequent primary use to second-life application and later recycling. The extent to which such a lifecycle view succeeds lies with the answer to two questions. First, is the data available? Secondly, is it possible to understand the complexity that arises along the entire lifecycle in its entirety so that it can then be reduced appropriately and mapped in algorithms? Suppose both questions can be answered positively in the future. In that case, this should contribute significantly to the pace of development and production of battery cells and thus pave the way even more forcefully to an electrified future.
Master of Batteries is a publication by CustomCells, one of the leading companies in the development and series production of state-of-the-art lithium-ion battery cells.
