2022: a year in battery modeling. Progresses and challenges.

This article is a summary of the open discussion that I hosted at the Battery Pub in January. Its purpose is to outline the main conclusions but also the questions that remained unanswered. The discussion is summarized from my point of view, thus, it contains personal opinions and perspectives.?The language of this article is colloquial, as in the style of the Battery Pub, and sometimes frank.

Before starting the event, I had in mind to dedicate enough time to all the macro areas in the battery sector that are positively affected by modeling. Defining a macro area is challenging because the boundaries between different phenomena and descriptions of the cell are grey, but for clarity, I list the following ones:?

• Materials discovery and cell design
• Pack design?
• Manufacturing
• Diagnostic and prognosis
• Recycling and second-life

These areas are very interconnected, being descriptions of batteries at different time and space scales, and at different steps of the battery life. To fully exploit their synergy, in the last two decades, scientists and engineers have had the common goal of integrating them in a multi-scale approach.

We were lucky to host people from all different walks of life, from academia to industry and startups, and the discussion was enriched by this heterogeneity. In my opinion, there was a stark contrast between the goals that are usually reported at the end of the research papers and the problems that people on the field experience every day.

It was clear from the beginning that diagnostic and prognosis raised the highest interest among the participants, and almost eclipsed all the other areas of discussion. It was no surprise given that most of the literature is focused on them and most of the battery modelers in the industry are working on them. The main conclusions of this part were:

• While research seems to focus more on electrochemical models (e.g. DFN, SPM), equivalent circuit models are still the option to go within many players in the industry
• The complexity of doing diagnostic and prognosis on battery packs is one or two orders of magnitude higher than at the cell level, and electrochemical models don’t scale as well as equivalent circuit models
• The technical and economical feasibility of the sensors is one of the main contributors to the stark difference between research papers and real-life applications?goals
• Parametrization is hard. In my opinion, it is harder with equivalent circuit models because of their weaker connection with physical phenomena. Nevertheless, it is a pain point regardless of the model of choice. Moreover, extensive and very accurate parametrizations that are presented in papers have the risk to be an overkill for real-life applications
• Temperature matters and it is sometimes forgotten
• It was suggested that too often the battery is considered and modeled as an isolated component, while, especially in automotive applications, it should be considered in relation to the whole vehicle, improving its safety and lifetime

Arrived at this point I just realized that I put “progresses and challenges” in the title, but I just talked about challenges so far. Well, progresses came mainly from the integration of machine learning into more traditional approaches and from a better understanding of how to model different degradation mechanisms both in commercial lithium-ion batteries and in the upcoming new chemistries.

Personal parenthesis about machine learning: it seems like forever ago when in 2019 I started to study the topic on my own, while it was not considered so useful by many of my colleagues.

Machine learning and neural networks are reducing the complexity of tuning more traditional models to on-field applications. For instance, real-life battery degradation is so convoluted and it depends on many factors that any model involving a system of partial differential equations will be either too expensive to run or too simple to capture it. Machine learning avoids this modeling problem by leveraging real-life battery data, and if it works, who cares that we don’t understand the physical meaning of the 10k+ parameters (weights) that are then used in the neural network? (I know it sounds like a rhetorical question but it is an open one actually).

Despite what I've just written, I don’t think reducing the complexity is the best describer, rather, machine learning is shifting the complexity toward the dataset. Data was a big topic of discussion as well and it comes with its additional challenges. The following problems were highlighted:

• Data availability; not enough open data
• Data skewness; the available data are not heterogeneous enough and centered around few cell models and applications
• Data quality; even if heterogeneous data are available they could be biased by the sensors that are used or by anthropocentric biases (more of this later)

On a good note, to avoid these problems, it has to be registered the rise of open source communities/projects and a call for more open data and standardized methodology across the whole battery sector.

We had a few minutes to dedicate to materials discovery and cell design. I think we all agreed that the classic trial-and-error approach is too slow and costly for modern battery challenges. On this topic, models, especially if they are able to be data-driven, can accelerate the research process significantly. Coming back to anthropocentric biases; I was told once, I wasn’t born yet, that there was an idyllic time when scientists used to publish every result, good and bad. Nowadays, either you discover a new material with good properties or you don’t publish anything. In a data-driven world, this represents a huge obstacle that can hold back the research of new materials for at least one decade. For instance, how can you find a new good material with a data-driven approach if all the materials in your dataset are fairly good? Imagine ChatGPT being trained only on Pulitzer-awarded novels. Will it be this useful?

Recycling and second-life are being discussed more and more, fortunately, and I think that modeling can help here too, especially connected to diagnostics; while manufacturing is still going under the radars. On the other hand, there is a lot of innovation in pack design, but to the best of my knowledge, I think the models used in this sector are already well-established and tested.

Lastly, one open question that we did not have time to answer was: what is a digital twin? It seems a stupid question but in the last year I’ve seen the following things being called digital twins:

• Battery's materials passport
• Datalog of SOx from on-field batteries
• DFN models that track everything is going on with the battery
• Even more complicated models
• Batteries registered on the blockchain

So, what is a digital twin?

In the end, I want to make some not obvious predictions for the following year or two (better to have a bit of room to avoid failure):

• ML/AI basic knowledge will be required in every battery modeling job post
• Manufacturing will be the sector that will be disrupted more by the use of models
• New ventures that want to develop new battery chemistries cannot consider anymore modeling as an afterthought
• At least one open-source project in the battery space will become closed and for-profit
• Digital Twin will become synonymous with Model