The performance of a battery is highly dependent on the properties of the individual components; primarily the cathode, anode, and electrolyte. For example, the practical specific capacity and cyclability of a cathode material depend on the composition, amount of doping (if applicable), particle size and distribution, and the thickness and porosity of the electrode. The properties of the cathode powder in turn are influenced by the type of precursors used and processing conditions, such as heat treatment temperature, soak time, and atmosphere, among other factors. Battery materials researchers, users and developers deal with the issue of having to optimize a large number of synthesis and process variables in order to maximize the performance of the components of the cell. As shown in Figures A and B, it is easy to optimize the performance when only one or two parameters need to be varied, but it is not straightforward when several variables are involved. One can use a design of experiments approach, however it requires significant resources to run these experiments, which translates to a lot of time and money. An alternative approach is Machine Learning.
The value of machine learning is rooted in its ability to create accurate models to guide future actions and to discover patterns that we’ve never seen before – Wired Magazine
Advances in computational algorithms and the significant increase in computational resources currently available has led to the new and exciting field of Artificial Intelligence, and Machine Learning in particular. This has reduced the need to perform a large number of experiments in order to obtain the optimized process parameters. For example, consider a scenario where the manufacturing and assembly of a battery component is dependent on more than a dozen parameters. Using Machine Learning (ML) algorithms, such as Gaussian Processes, the situation can be expressed as:
ŷ = ƒ(x1,x2,.…….,xn)
Where, ŷ is the characteristic of the battery or battery component we need to optimize (e.g., specific energy density) and xi are different composition and process parameters (e.g., type and amount of precursor for the cathode material, heat treatment temperature, thickness of the cathode tape, charge/discharge rates), and ƒ is a function of xi. The ML algorithms can determine the shape of ƒ with a limited number of experiments (called training set) so that ƒ can then be used to predict the value of the desired characteristic ŷ with given inputs. Machine Learning can also be used to determine which parameters are more important (and need to be observed more closely during manufacturing) and which parameters are less important.
The Machine Learning algorithm is applicable to both battery materials and assembled cells. Contact us today using the form below to explore how the ML approach can be adapted for your specific need in a cost effective manner.