Thoughts

Paper Review - Discovery Learning for Battery Cycle Life Prediction

Discovery Learning Predicts Battery Cycle Life from Minimal Experiments

DOI: 10.1038/s41586-025-09951-7

The Learner is a Gaussian process regression model. Its goal is to take operating-condition parameters, such as charging rate and temperature, as inputs and estimate the number of cycles it takes for the SOH to reach 90%.

Gaussian process regression learns by repeatedly selecting an input vector for which the current model has the greatest uncertainty about the output, querying its output, and updating itself with the result. To simplify the discussion, suppose the input is a vector with only two elements: temperature and the lithium diffusion coefficient in the negative electrode. For the Learner to learn anything in this setup, there needs to be some entity that can answer a question like:

“If a battery with a negative-electrode lithium diffusion coefficient of xx is cycled at room temperature, after how many cycles will its SOH fall below 90%?”

That entity is the Oracle.

The Oracle represents knowledge about the world. It is a zero-shot model whose goal is to take parameters used in an electrochemical battery model as inputs and accurately predict the number of cycles until the SOH reaches 90%. Because battery degradation depends not only on electrochemical parameters but also on operating conditions, the authors introduce a dual-predictor architecture to separate the effects of operating conditions from those of electrochemical parameters. According to the Supporting Information, the base model that captures the effects of electrochemical parameters uses Elastic Net, the same type of approach used by Severson et al. in their 2019 Nature Energy paper. The meta-model that accounts for the effects of operating conditions uses a support vector machine.

The Interpreter is left yet unexplained. I still need to read further to fully understand why this component is necessary, but so far it looks like the authors essentially introduced another ML module because fitting charge–discharge data directly with a full DFN model to extract electrochemical parameters would be cumbersome and computationally demanding.

Now let us reconsider the Learner–Interpreter–Oracle architecture. The learning process works roughly as follows.

At the beginning, the Learner is clueless, so it picks some point in the input-vector domain - that is, it makes a query. Presumably, a query from the Learner is something along the lines of:

“Let’s take active material A, build a cell according to design B, and cycle it under operating condition C.”

In other words, it chooses a triplet (A,B,C). We then need to actually run the experiment and, in any case, perform RPTs through 50 cycles.

The RPT data are then passed to the Interpreter, whose purpose is to obtain the DFN parameters. With PyBaMM, we can take a set of electrochemical parameters and generate a battery voltage profile. What we want here, however, is the inverse process: given experimental data, infer the electrochemical parameters that could have produced them. The authors use a package called sbi for this purpose. This is a part I still need to study in more detail.

Finally, the input parameter set reconstructed by the Interpreter is passed to the Oracle, which provides the answer we actually want: the number of cycles until the SOH reaches 90%.

Tags: study

Google AI Professional Certificate - Module 1

Three ways to collaborate with AI 1) AI helps you organize and understand information 2) AI helps you create brand-new things 3) AI helps you improve and refine your work.

Tags: study

401k

The 2026 contribution limit for pre-tax or Roth 401(k) is USD 24,500. Mega Backdoor Roth (converting after-tax 401(k) contributions to a Roth 401(k) or Roth IRA) allows up to USD 35,250 in additional contributions, while the standard Backdoor Roth IRA (traditional IRA to Roth IRA) allows up to USD 7,500.

Realistically, however, I won’t have enough cash flow to fully utilize all of these options. Therefore, my primary objective is simply to maximize my employer match.

Since I joined my company in the middle of the year, reaching the full 2026 401(k) contribution limit would require a much higher contribution rate for the remainder of the year, which makes the decision difficult.

Tags: finance

Blog Design Rework

I am working on redesigining the style of this blog. It is a bit pathetic to focusing on the packaging rather than the contents but I cannot help doing this. Designing is more compelling aspect of blogging to me.

Tags: blog

Social Conversation is Rough

Casual conversations in English are still challenging for me. I’m usually excited to start one - and fortunately, my colleagues are proactive about saying hello, so getting the conversation started isn’t the hard part. The challenge is keeping it going. My limited fluency and the lack of conversational topics often cause the conversation to lose momentum.

What’s even more discouraging is that I can sense, almost in real time, when the other person is also running out of momentum and struggling to stay engaged. The moment I notice that, I become even more reserved, which only reinforces the vicious cycle.

I’m beginning to wonder if I should prepare a small cheat sheet of casual conversation topics until this becomes more natural.

Tags: blog

Thought Starter

This entry isn’t meant to archive anything particularly useful. It’s simply an attempt to keep the momentum of writing alive.

Lately, I’ve found myself trapped in a state of paralysis, where every commitment feels weighed down by its opportunity cost. Instead of making progress on anything meaningful, I end up doing nothing at all, wasting time while knowing exactly what I’m doing. Yet I still haven’t been able to break this vicious cycle.

At times, I even catch myself feeling that the time I devote to my full-time job is taking away from the time I could be spending on side projects or building future passive income. To be clear, I don’t feel this way while I’m actually at work - that’s exactly what that time is meant for. Rather, it’s the feeling that lingers outside of work, as if every free hour carries the burden of what else it could have become. I know it’s an irrational way to think, but the feeling persists nonetheless.

Tags: blog

AI Blacksmith

I’ve been thinking a lot about whether publishing my thoughts on a personal blog still benefits me in the age of AI. Yet I haven’t found a convincing answer, and I’m beginning to lose momentum.

One idea I keep coming back to is that, in the long run, the ability to read, understand, and think for myself may become even more valuable than delegating the entire process to AI assistants.

But then an analogy occurred to me - have I ever worried about losing my arithmetic skills simply because computers have been doing calculations for decades?

Am I just an obsolete blacksmith, hammering away at an iron blade while muttering, “AI has no soul, and I don’t like it”?

Tags: blog

Syncing Obsidian Vault with Github Repo in Iphone (iSH + Obsidian)

Using iSH together with Obsidian, it is possible to write and manage blog posts in Obsidian even on an iPhone. The key is the mount -t ios-unsafe command, which maps an Obsidian Vault stored in the iPhone’s Files app into iSH’s Linux filesystem. This allows the Vault, which would otherwise be accessible only through the Files app and Obsidian, to be accessed directly from iSH as well.

However, repositories mounted into iSH using ios-unsafe may encounter Git ownership verification issues, preventing Git operations from running correctly.

If this happens, you can mark the repository as trusted by running:

git config --global --add safe.directory <repository-path>

This tells Git to treat the repository as a trusted directory despite the ownership mismatch introduced by the ios-unsafe mount.

Tags: blog

Interactive Key Visual Design for Blog Posts

I’m considering using JavaScript-based interactive canvases as the key visual for blog posts instead of static images or animated GIFs. Since AI can now generate high-quality JavaScript code with relatively little effort, the production cost is likely to be much lower than it would have been in the past.

For example, a post about the idea that success requires unseen effort could feature a small interactive game where the reader has to keep a swan afloat by constantly paddling its legs beneath the water. Rather than merely illustrating the concept, the visual would allow readers to experience it firsthand through interaction.

Tags: blog

Added a Page for Accumulating Pieces of Thoughts

I came to believe that publishing unfinished posts and iteratively refining them is not the optimal strategy for maintaining high-quality content over the long term. As the number of incomplete posts - ambitiously thrusted in the beginning - grows, it becomes increasingly difficult to track, revisit, and improve each one. Every post should reach a point where, at the time of publication, there is little or nothing meaningful left to add.

So moving forward, I plan to maintain a separate thoughts collection that serves as a workspace for fragmented ideas, notes, and pieces of information. Once the content around a particular topic has matured sufficiently, it will be developed into a standalone post and published.

In the long term, I plan to introduce an Obsidian-inspired graph view to visualize the evolving network of thoughts, making it easier to explore their connections and observe how ideas gradually converge into larger themes.

Tags: blog