International outlook
Practical applicability
Realism
Imagine you need to understand how a car engine works, but all you have are the speedometer and fuel consumption readings. You can't take the engine apart, and you can't stop it. This is roughly the situation engineers face when trying to figure out what's happening inside a lithium battery.
In the world of electric vehicles and energy storage, this is a critical task. A battery isn't just a «big AA battery.» Inside, complex processes unfold, changing with temperature, age, and state of charge. And if you don't understand these processes, you can't predict how long it will last, charge it correctly, or notice in time that it's starting to degrade.
The Problem No One Notices
When the average person buys a smartphone or an electric car, they think of the battery as something simple: if it has a charge, it works; if not, it doesn't. But for an engineer, a battery is a system of many interconnected elements, each influencing the behavior of the whole.
Inside a lithium-ion battery, fast processes lasting seconds and slow ones stretching over hours occur simultaneously. It's as if your house had both an instant hot water system and underfloor heating with multi-hour inertia. To manage this whole setup properly, you need to understand the principles behind each system.
Traditionally, engineers studied batteries like this: they took a new battery, connected it to a lab bench, and ran special tests. They would fully charge the battery, then discharge it with a tiny current over tens of hours, measuring the voltage along the way. Then they would plot a graph of voltage versus state of charge – the so-called OCV (Open-Circuit Voltage) curve.
This sounds reasonable, but there are several problems. First, such tests take a very long time – sometimes several days for a single battery. Second, lab conditions are very different from real life. Third, a battery ages and its characteristics change, and no one is going to run multi-day tests every month.
Math vs. Reality
But that's not all. The main problem lies in how engineers process data from the battery. The real world is continuous: current flows smoothly, temperature changes gradually, and chemical processes develop over time without jumps. But computers work discretely: they measure and process data in chunks, at a specific frequency.
Imagine trying to understand how a watch mechanism works, but you can only look at it through a strobe light flashing a couple of times per second. A lot goes unnoticed between the flashes. This is exactly what happens when you move from the continuous processes in a battery to discrete data in a computer.
Classical analysis methods force you to first build a mathematical model in continuous time and then «slice» it into discrete pieces to work with the measured data. In this slicing process, information is inevitably lost, especially about fast processes. It's like trying to understand a melody from individual notes played with long pauses in between.
In batteries, this problem is particularly acute because the fast and slow processes differ so much in speed. If you tune the measurement system for fast processes, the slow ones get «smeared out» and become indistinguishable. If you tune it for the slow ones, the fast ones turn into unintelligible noise.
The Siberian Approach
A group of researchers decided to approach the problem from a different angle. Instead of translating a continuous model into a discrete one, they developed a method that works directly with the continuous description while still using discrete measurements.
Picture this analogy. You have a complex borscht recipe where some ingredients must be added in a specific sequence and cooked for different times, with the final taste depending on many factors. The classic approach is like trying to replicate the recipe from photos of the finished borscht taken at regular intervals. The new approach is like analyzing the aroma coming from the pot, which contains information about all the processes at once.
In the case of a battery, the «aroma» is a set of special mathematical filters called Laguerre filters. These filters have a remarkable property: they can «extract» information about the system's internal processes from external observations, without losing details of either the fast or the slow processes.
Laguerre filters work like a set of special «sieves» with different mesh sizes. Each «sieve» lets through information of a certain «size»: fast changes pass through fine sieves, while slow ones are caught by coarse ones. By analyzing what passes through each sieve, you can reconstruct the complete picture of what's happening inside the battery.
The Smart Curve That Knows Everything
But there's another challenge. The relationship between a battery's voltage and its state of charge is not a simple straight line or even a smooth curve. It has sections where the voltage changes very sharply, smooth areas, bends, and quirks. Things get especially tricky at the edges – at very low and very high states of charge.
Traditional mathematical methods struggle with such «tricky» curves. They either oversimplify the curve beyond recognition or create numerous artifacts – false features that don't exist in reality.
The researchers used a mathematical tool called «cubic B-splines» to describe this curve. To continue the analogy, a standard approximation is like trying to draw a complex shape with a compass and a ruler. A B-spline, however, is like a flexible wire that can take almost any shape while remaining smooth and not creating unnecessary loops or kinks.
A B-spline consists of several pieces, each being a third-degree polynomial. At the junction points, smoothness is ensured – there are no sharp transitions or breaks. At the same time, each piece can «adjust» to the local features of the curve independently of the others.
The key advantage of this approach is that the B-spline parameters can be determined simultaneously with the parameters of the battery's electrical model. You don't need to first plot the curve in a lab and then separately determine the model's parameters. Everything is done in one go, based on data obtained under real operating conditions.
Regularization vs. Overfitting
When you have many parameters to determine, a classic machine learning problem arises: overfitting. The system might find a set of parameters that perfectly describes the measured data but is completely useless in new situations.
It's like a student who has memorized every problem in the textbook, including all the intermediate steps, but hasn't understood the underlying principles. He'll get a perfect score on an exam with the same problems, but any new problem will leave him stumped.
To avoid this problem, the algorithm incorporates two regularization mechanisms – mathematical «penalties» for overly complex solutions.
The first mechanism is called nuclear norm regularization. It's based on the fact that in real physical systems, the number of truly independent processes is usually small. Most observable effects are combinations of a few basic mechanisms. Nuclear norm regularization «nudges» the algorithm toward finding such simple explanations.
The second mechanism is L1 regularization of the B-spline's third derivative. It sounds complicated, but the idea is simple. The third derivative characterizes the «jaggedness» of the curve – how many small bends and features it has. L1 regularization penalizes excessive jaggedness, forcing the algorithm to choose smoother, yet still accurate, solutions.
Imagine you're trying to build a road through a mountainous area. You could lay it out to follow the terrain exactly, replicating every dip and bump. Such a road would match the landscape perfectly, but it would be impossible to drive on. Alternatively, you could choose a route that accounts for the main features of the terrain while remaining smooth enough for practical use. L1 regularization provides exactly that balance.
From Theory to Practice
Any theory is only as good as its practical application. Therefore, the researchers thoroughly tested their method on both synthetic data and results from real experiments.
First, a computer model of an ideal battery with known parameters was created. Noise was added to this model's data to simulate the inaccuracies of real measurements. The task was to reconstruct the original model parameters from the noisy data.
The testing used a standard FUDS (Federal Urban Driving Schedule) load profile – a sequence of charge and discharge currents that mimics city driving with stops, accelerations, and braking. This profile is useful because it contains both rapid changes (sharp accelerations) and slow periods (long traffic jams), allowing for a test of the algorithm's ability to handle processes on different timescales.
The results exceeded expectations. The voltage reconstruction error was less than one millivolt – a level of accuracy sufficient for practical use in battery management systems. The electrical model parameters were recovered with over 95% accuracy, and the voltage vs. state of charge curve was nearly indistinguishable from the true one.
Crucially, the algorithm demonstrated stability: repeated runs with different noise implementations yielded consistent results. This indicates that the method genuinely extracts useful information rather than just fitting to random features in a specific dataset.
Field Test
But the real test began when the method was applied to data from an actual battery. Open-source data from the CALCE (University of Maryland) database – one of the world's leading battery research centers – was used.
The experiment was conducted on a lithium-ion battery at 25 degrees Celsius. The initial state of charge was 80%, which is typical for practical applications, as batteries are usually operated between 20% and 80% to extend their lifespan. The load again simulated an urban driving cycle.
The results were impressive. The model reproduced the measured voltage with over 99% accuracy, and the average error was just 16 millivolts. For comparison, the nominal voltage of a lithium-ion battery is about 3.7 volts, so the relative error is less than 0.5%.
The identified battery parameters revealed two characteristic time constants: about 1.5 seconds for fast processes and approximately 31 seconds for slow ones. This aligns well with the physical understanding of processes in lithium-ion batteries. The fast time constant corresponds to charge transfer processes at the electrode-electrolyte interface, while the slow one relates to lithium diffusion within the bulk electrode material.
The Secrets a Battery Hides
But the most interesting part came from analyzing the reconstructed curve of voltage versus state of charge. The resulting curve not only matched the results of a traditional lab test well but also provided more detailed information.
In the areas between the experimental points of the standard test, the new method revealed subtle features of the curve that had previously gone unnoticed. Particularly important was the discovery of a sharp voltage drop below a 10% state of charge – a region almost inaccessible during standard stationary tests due to the risk of damaging the battery.
This information is critically important for battery management systems. Knowing the precise shape of the curve in these extreme regions allows for more accurate determination of the remaining capacity and helps prevent critical operating conditions.
Furthermore, the method allowed for the identification of certain asymmetries in the battery's behavior during charging and discharging, which are not captured by standard quasi-static testing. These asymmetries are related to different kinetic processes during the forward and reverse electrochemical reactions.
The Practical Value of the Discovery
The developed method opens up new possibilities for practical application. Its main advantage is the ability to determine a battery's characteristics directly during operation, without the need for lengthy lab tests.
This is especially important for electric vehicles, where the battery operates under constantly changing conditions. Temperature, humidity, driving style, and battery age all affect its characteristics. Traditional methods can't account for this diversity because they are based on measurements under standard laboratory conditions.
The new approach allows the battery management system to «learn» on the fly, continuously refining its model based on accumulating data. It's the difference between a doctor who makes a diagnosis based on a textbook of symptoms and one who observes a patient over time and adjusts treatment based on the body's response.
Another important application is the health diagnostics of batteries during operation. Changes in the model's parameters can signal the onset of degradation long before it becomes noticeable through external signs. Early detection of problems allows for measures to be taken to extend the service life or to schedule a battery replacement before it fails.
A Look to the Future
Although the developed method has shown impressive results, this is just the beginning. The researchers see several directions for future development.
First, the model can be expanded to account for the temperature dependence of its parameters. In real-world conditions, a battery operates over a wide range of temperatures, and its characteristics change significantly. Incorporating temperature dependence into the model will improve prediction accuracy in various climates.
Second, it would be interesting to adapt the method for analyzing long-term battery degradation. The current model assumes that parameters are constant during a single charge-discharge cycle. But on longer timescales, parameters change due to aging. Tracking these changes will help predict the remaining useful life and optimize operating modes.
A third direction is applying the method to new types of battery systems. Lithium-ion batteries are gradually being supplemented and replaced by other technologies: lithium iron phosphate, sodium-ion, and solid-state. Each technology has its own dynamic characteristics, and the method will need to be adapted for them.
The Philosophy of the Approach
Behind the technical details lies an important philosophical idea: it's better to study a system in its natural operating conditions rather than in an artificial lab environment. A battery in a car behaves very differently from one on a test bench. It is subjected to vibrations, temperature fluctuations, and irregular loads. All these factors influence its behavior.
The traditional approach tries to isolate the object of study from external influences to obtain «pure» results. But often, it is the interaction with the real environment that holds the key to understanding the system's true behavior.
This is similar to studying animal behavior. You can observe an animal in a zoo and draw certain conclusions about its habits. But only observation in its natural habitat – where the animal interacts with its own kind, forages for food, and defends itself from predators – provides the full picture.
Similarly, a battery in a real system «communicates» with other components, reacts to changes in load, and adapts to operating conditions. This «social» aspect of its behavior is no less important than its internal electrochemical processes.
What This Means for the Industry
The development of battery analysis methods is of strategic importance for the entire energy storage industry. As the share of renewable energy sources grows, batteries are becoming a critical component of the power grid. The success of the energy transition depends on their reliability and efficiency.
More accurate battery models will enable the creation of more effective management systems, which will directly impact the cost and reliability of electric vehicles, uninterruptible power supplies, and industrial energy storage systems. Every percentage point of improvement in battery utilization translates into millions of dollars in savings on an industry-wide scale.
Furthermore, a better understanding of degradation processes will help developers create more durable batteries and optimize their operating algorithms. This is especially important in the context of growing requirements for disposal and recycling – the longer a battery lasts, the smaller the environmental footprint of its life cycle.
This method shows that modern mathematical tools can extract far more information from ordinary measurement data than was thought possible just a short time ago. This opens up prospects not only for battery research but also for the analysis of other complex technical systems.
In the age of digitalization and the Internet of Things, every device generates vast amounts of data. The ability to analyze this data correctly is becoming a key competitive advantage. Approaches like the one described here show the way to turn «big data» into useful knowledge about the behavior of technical systems.
When engineers learn to make equipment tell the truth about its condition, it will change our approaches to maintenance, diagnostics, and design. Instead of scheduled maintenance, we will have condition-based maintenance. Instead of conservative safety margins, we will have precise assessments of remaining useful life.
Ultimately, this is the path to creating truly smart technical systems that understand themselves and can independently optimize their behavior. A battery that knows its own capabilities and limitations will become a far more valuable component than just a container for storing energy.
