How Fawkes Models Batteries

A battery is always telling you something about itself. You can see it in the way its voltage sags under load, how quickly it warms up, and how its capacity quietly erodes over hundreds of cycles. The trouble is that most of the industry reads that information secondhand, through numbers a battery management system (BMS) has already interpreted and often distorted. At Fawkes Energy, we built our modeling stack to read it firsthand.

This post explains how our models actually work. What goes in, how the pieces fit together, and why we built it the way we did.

A hybrid, not a black box

There are two classical ways to model a battery, and each one fails in a different direction.

Physics-based models describe what is happening inside the cell using electrochemical first principles. They are accurate and explainable, but they are hard to calibrate for every chemistry and every real-world condition, and they get brittle when the world stops behaving like the lab.

Machine-learning models learn patterns directly from data. They are flexible and powerful, but they are only as trustworthy as the data they were trained on, and they can produce confident nonsense the moment they drift outside what they have seen before.

Fawkes uses neither one on its own. Our approach is a physics-informed, data-driven, sometimes called a grey-box or digital-twin framework. We use physics to set the rules of what is possible, and we use machine learning to capture the messy reality of how batteries actually behave in the field. The physics keeps the machine learning realistic, and the machine learning keeps the physics honest.

What you get is an algorithm that predicts battery behaviour, including capacity fade, state of health, and emerging safety anomalies, from incoming field data. It carries the explainability of physics and the adaptability of modern machine learning at the same time.

Here is how it comes together, in four stages.

Stage 1: Laboratory baselining

Everything starts with real batteries on real test benches.

We cycle cells of varying chemistries inside controlled environmental chambers and put them through precisely known conditions of temperature, current, and depth of discharge. Because the environment is controlled, we can isolate clean measurements of how each chemistry behaves. Think of it as the pure electrochemical fingerprint of the cell.

Those measurements become the foundation for our physics-based and empirical models. They are the baseline curves that describe how a healthy cell of a given chemistry should behave under normal conditions, and they are the ground truth that everything downstream is anchored to.

Stage 2: Rejecting OEM bias and working from raw signals

This is where our approach parts ways with most of the industry.

When a battery operates in the field, its BMS constantly reports interpreted metrics, most notably State of Charge (SoC) and State of Health (SoH). It is tempting to build analytics on top of these numbers because they are convenient. We deliberately don't.

Manufacturer-calculated metrics like SoC and SoH are downstream estimates. They carry the assumptions, rounding conventions, sensor drift, and software biases of whatever hardware produced them, and those biases vary from one OEM to the next. If you build your intelligence on top of them, you inherit every error baked in beneath.

Instead, our models work from the raw signals a battery genuinely produces: voltage, current, and temperature measured over time. Because we start from raw ground-truth data rather than someone else's interpretation of it, every insight we generate is our own and traces back to physics rather than to a third-party estimate.

Stage 3: Hybrid machine-learning training

Lab data tells you how a cell behaves in ideal conditions. The real world is never ideal, and that gap is exactly where our machine-learning layer lives.

We combine the physics and empirical parameters from Stage 1 with large volumes of historical, real-world field data. On top of the raw voltage, current, and temperature streams, the models add contextual signals that capture how a battery is actually used. That includes ambient temperature at the point of operation and usage patterns that differ sharply by application, such as the erratic acceleration and charging of an EV versus the irregular daily cycling of a solar storage system.

The machine-learning layer uses this real-world context to refine and correct the boundaries of the physics-derived curves. Where the idealized model and the field diverge, the machine learning learns the correction. Because the physics still defines the outer limits of what is plausible, the model adapts to reality without ever losing its grounding. It fills in how a battery really behaves without contradicting how it physically can behave.

That is the core of what makes the algorithm both accurate and defensible. It is tuned to the real world but bounded by physical law.

Stage 4: Closed-loop lab validation

An AI prediction is only worth as much as its proof. So we close the loop by sending our predictions back to where they started, the lab.

When the algorithm flags a particular behaviour, degradation trajectory, or safety condition, we can recreate those exact conditions on physical cells in our chambers and test whether real batteries do what the software said they would. The predictions get checked against physical reality, not just against held-out data.

This closed loop does two jobs. It confirms accuracy before any insight reaches a customer, and it feeds fresh ground truth back into the models so the system keeps getting sharper instead of drifting.

The models compound over time

Most software only gets better when engineers sit down and improve it. Battery intelligence works differently. Ours gets better on its own, simply by being used.

Every charging event, every diagnostic session, every degradation pathway and field failure adds a new piece of evidence about how batteries behave. That evidence flows back into the models automatically, recalibrating them, tightening their uncertainty, and testing their assumptions against what actually happened in the field. Each new chemistry, pack design, or operating environment we encounter widens the range of conditions the models understand.

The effect builds on itself. The more batteries the system observes, the more accurate it becomes, and the more accurate it becomes, the more valuable each new observation is. Over time this creates a body of battery knowledge that no fixed dataset can match, because it is not really a dataset at all. It is a continuously improving understanding of battery behaviour across chemistries, manufacturers, and applications.

Why AI alone can't replicate this

It is a fair question to ask in an era of powerful AI. Couldn't a competitor just train a model to do the same thing?

The short answer is no, and the reason is worth understanding. Modern AI has made writing software far easier, which means code is no longer the bottleneck it once was, and it is no longer where the real advantage lives. Anyone can build a model architecture. What they cannot build overnight is everything that has to sit underneath it.

Our advantage is not a single clever algorithm. It is the way we bring together things that are hard on their own and much harder as a set: raw, high-fidelity measurements captured from real batteries in the field rather than borrowed from biased OEM software, a foundation of controlled lab data across many chemistries, physics that keeps the machine learning honest, and a validation loop that turns every deployment into new ground truth. A competitor starting today has none of that accumulated evidence and no shortcut to acquiring it, because it can only be earned battery by battery and cycle by cycle, over time.

The moat, in other words, is not the model. It is the proprietary combination of hardware-grade data, battery science, and a system that has been learning from real assets longer than anyone else's. That head start compounds, and no amount of AI can prompt it into existence.

Why this matters

Put the four stages together and you get a modeling framework with a set of advantages that are genuinely hard to replicate.

• It is unbiased at the source. By refusing OEM-calculated metrics and working from raw signals, we do not inherit the drift and error of third-party hardware. Our insights are our own.
• It is explainable. Because physics underpins every prediction, we can explain why the model expects a given outcome rather than simply asserting it, which matters enormously for safety-critical decisions.
• It is context-aware. The same cell behaves differently in an EV than in a grid installation, and our models account for application and usage rather than assuming one size fits all.
• It is proven, not just predicted. The closed-loop lab validation means every insight stays grounded in physical reality.
• And it compounds. The system improves with every battery it observes, building an advantage that widens over time rather than one a competitor can simply copy.

Taken together, physics gives us rigour, machine learning gives us real-world adaptability, raw signals give us integrity, and lab validation gives us proof. That is what we mean when we call the Fawkes Energy modeling stack a proprietary framework. The individual ingredients draw on established science, but the way we fuse them, the data we train on, and the validation pipeline we built around them are uniquely ours.

Batteries are becoming the backbone of how the world stores and moves energy. Understanding them accurately, rather than approximately, is the difference between guessing and knowing. That is what Fawkes was built to do.