Battery X-Ray Computer Vision Ties LFP Speed to Carbon
Battery x-ray computer vision mapped lithium iron phosphate reaction kinetics across 39 particles. Learn how carbon coating limits fast charging.
With Craig Major and Jon Irwin

Estimated reading time: 10 minutes
Key Takeaways
- A model trained on 62 X-ray movies (about 180,000 pixels) reached a 6.8% training error and a 9.6 ± 0.9% cross-validation error. The estimated experimental noise was about 7%.
- In lithium iron phosphate (LFP) particles, spots with a stronger carbon signal reacted more slowly. The pixel-level correlation was −0.40.
- The model was physics-constrained. It used a phase-field model, Bayesian inference and uncertainty estimates, and it did not rely on a black-box image classifier.
- The inferred reaction curve was asymmetric. That contradicts the common porous-electrode assumption that exchange current is proportional to c(1−c).
- No independent source found here shows a commercial product, a quantified manufacturing cost cut or a pack-level capacity gain. For now this is an engineering opportunity.
Table of contents
- How 180,000 Pixels Became a Battery Physics Lesson
- Physics-Informed Battery Machine Learning, Not a Black Box
- What Lithium Iron Phosphate Reaction Kinetics Looked Like Up Close
- Why Carbon Coating on Lithium Iron Phosphate Can Slow Charging
- From Single Particles to Whole Electrodes
- Where Physics-Informed Models Can Still Fail
- What This Means for Startups and Small Businesses
- Frequently Asked Questions
A machine learning model learned how lithium moves inside battery particles from just 180,000 image pixels. Its training error was 6.8%. Its cross-validation error, measured on data held back from training, was 9.6 ± 0.9%. The researchers estimated the experimental noise at about 7%.[1]
This is battery x ray computer vision that does more than label pictures. It reads X-ray movies of lithium iron phosphate (LFP) particles and infers how fast the chemistry runs at each spot. The study was published in Nature in September 2023 by a team from SLAC, Stanford, MIT and Toyota Research Institute.
For battery startups and manufacturing operators, the message is practical. Coating uniformity and interface diagnostics look like possible levers for faster charging, without changing the core LFP chemistry. Physics-informed AI could also cut some of the expensive trial-and-error testing that slows cell development.
How 180,000 Pixels Became a Battery Physics Lesson
LFP is a common battery material. It is valued for low cost, safety and the use of abundant elements, and it is increasingly used in electric vehicles.
During discharge, lithium ions flow into the LFP electrode. During charging, they flow out. This process is called intercalation, which simply means ions sliding into and out of a solid material.
The team built transparent cell batteries so they could watch this happen during operation. They used scanning transmission X-ray microscopy, which makes images by scanning a focused X-ray beam across a sample. The work used beamlines 11.0.2.2 and 5.3.2.1 at the Advanced Light Source.[1]
The data set was specific:
- 39 LFP particles
- 62 charge or discharge half-cycles
- 4–10 time-resolved images per half-cycle
- Cycling rates from 3.0C to 0.2C, meaning a full theoretical charge or discharge in 20 minutes to five hours
An average particle covered about 1.2 square micrometers. That works out to about 490 pixels per particle. Each particle was about 150 nanometers thick along the imaged direction.[1]
Add it up and you get roughly 180,000 pixels of data. That is small by modern AI standards. The small size is why the physics inside the model matters so much.
The 2023 study was not the team's first look at these particles. Chueh and Bazant had worked together for eight years, mixing math models with advanced X-ray microscopy. In 2016 they published nanoscale movies of lithium flow in LFP nanoparticles. With Toyota's support, they also used MIT-built tools to speed up battery testing and improve charging methods.
Source: PubMed Central: Zhao et al., Nature 2023
Physics-Informed Battery Machine Learning, Not a Black Box
Many people hear "computer vision" and picture a model that sorts photos into labels. This study did something different.
The team did not use a black-box image classifier. They built a model where known physics sets the rules, and the images fill in the unknowns. The main ingredients were:
- A reaction-limited Allen–Cahn phase-field model. This describes how lithium-rich and lithium-poor zones form and change inside a particle.
- PDE-constrained optimization. This fits the model to the images while forcing it to obey the governing equations (partial differential equations).
- Bayesian inference and uncertainty quantification. These give a range of likely answers instead of one number.
From the X-ray videos, the model inferred three things. The first was the free-energy landscape, which describes how much energy the particle stores at each lithium level. The second was the concentration-dependent exchange current, a measure of how fast the reaction runs at a given lithium level. The third was a particle-specific spatial reaction-rate prefactor, which is a map of how much faster or slower each patch of surface reacts.[1]
The setup also used a 50 nanometer assumed spatial correlation length. That tells the model how smoothly reaction rates can change from one spot to the next.
Two other numbers came from the literature. The LFP binodal parameter delta was around 0.035 at room temperature. The gradient-penalty coefficient was 5.02 × 10^-10 J m^-1.[1]
This design is why the low error counts for something. With only 180,000 pixels, a pure pattern-matcher would be easy to fool. A model locked to physical laws has less room to cheat.
The authors described the approach as image learning. Chueh said it extracts information that was previously inaccessible from the images. The researchers also said the same approach could help study other chemical and biological systems, such as cell division.
Source: PubMed Central: Zhao et al., Nature 2023
Turn Repetitive Lab and Ops Work Into Automated Workflows
Research teams and battery startups often lose hours to copying data between instruments, spreadsheets and reports. Flowgrammer.ca builds business automation that handles that manual work, so a small team can run more experiments, orders and reports without adding headcount.
What Lithium Iron Phosphate Reaction Kinetics Looked Like Up Close
Reaction kinetics is the study of how fast a reaction runs. In a battery, it affects how quickly you can charge without losing capacity or damaging the cell.
The team found that lithium insertion rates varied inside a single LFP nanoparticle. One part of a particle could react faster than another part a few hundred nanometers away. Averages hide that.
The study also checked its inferred exchange-current curve against theory. The curve agreed with coupled ion-electron transfer theory, using a literature reorganization energy of 8.3. The authors described no adjustable fit parameters beyond the theory inputs.[1]
The shape of the curve is the surprising part. It was asymmetric. The commonly used porous-electrode assumption says exchange current is proportional to c(1−c), which is a symmetric curve. The data did not follow that rule.[1]
Why should a founder care about a curve? Because battery simulations are built on assumptions like that one. If the assumption is wrong, the simulation can mislead design choices. Cell designers use these models to pick electrode thickness, charging protocols and materials.
The authors also argued that the electrolyte–electrode interface, and not bulk lithium diffusion alone, controls much of the dynamics they saw. Bazant said interface engineering should be the optimization focus.
That points to the surface of the particle, where ions cross from liquid to solid, as the place where speed is won or lost.
Source: PubMed Central: Zhao et al., Nature 2023
Why Carbon Coating on Lithium Iron Phosphate Can Slow Charging
LFP is a poor electronic conductor. Makers coat the particles with carbon so electrons can reach the surface. The coating is necessary.
The study shows that more carbon is not always better. The researchers compared the inferred surface reaction rate with carbon intensity measured by Auger electron microscopy. This technique maps the elements on a surface. The correlation at the pixel level was −0.40.[1]
In plain terms, regions with a thicker carbon signal reacted more slowly. Excess coating can impede lithium transport.[1]
A correlation of −0.40 is moderate. It does not say carbon explains everything. But it is a measured link between a coating property that manufacturers can influence and a performance property that customers care about.
Here is what that suggests for coating work:
- Uniformity may matter as much as average thickness. A coating that is thick in some spots and thin in others creates fast and slow zones.
- Thick spots are a cost. They add material that can slow lithium, while thin spots may starve the particle of electrons.
- Local measurement beats bulk testing. A whole-cell test cannot tell you which patches of a particle are lagging.
The source article states it directly: the findings suggest that engineering the interface between electrolyte and electrode materials is crucial for improving battery processes. The team's stated goal is to apply these insights to battery design.
The study stops short of a manufacturing recipe. It does not give a target thickness or a process change. It gives a diagnosis and a direction.
Source: PubMed Central: Zhao et al., Nature 2023
From Single Particles to Whole Electrodes
A particle-level finding only matters commercially if it scales up to electrodes and cells. There is early evidence that others are using it that way.
A 2025 electrochemical-engineering study took the Zhao et al. kinetics and put them into a network model. The model represents an electrode as many single-crystal particles that are electrically connected. That suggests the particle-scale findings are being used as inputs for electrode-scale simulations, and are not just a microscopy result.[2]
Two 2026 reviews place the work in a wider trend. They describe a move from purely predictive battery AI toward hybrid and physics-informed systems. One is in MRS Energy & Sustainability and the other is a mini-review in EES Batteries.[3][4]
The reviews say hybrid models that mix machine learning with physical constraints can improve:
- Interpretability, so engineers can see why a model gives an answer
- Data efficiency, so useful models need fewer expensive experiments
- Extrapolation, so predictions hold up better outside the training range
Data efficiency may be the most relevant for startups. Battery data is slow and costly to collect. An approach that gets more from each experiment is a real advantage.
The work was supported by Toyota Research Institute through its Accelerated Materials Design and Discovery program. The X-ray work used the Advanced Light Source, a U.S. Department of Energy Office of Science user facility supported under contract DE-AC02-05CH11231.[1]
Source: RSC Energy & Environmental Science (2025); Springer, MRS Energy & Sustainability (2026); RSC EES Batteries (2026)
Where Physics-Informed Models Can Still Fail
Adding physics to AI does not make it automatically right. The 2026 reviews are direct about this.[3][4]
They flag two main risks:
- Incomplete assumptions can add systematic bias. If the physics in the model is missing something, the model can be confidently wrong in a consistent way.
- Narrow lab data can fail to carry over. Models trained on small laboratory data sets can break when the chemistry, cell design, manufacturing variation or operating protocol changes.
Apply that to this study. The data came from 39 particles of one material, imaged in transparent lab cells. Real cells use different electrolytes, binders, electrode thicknesses and charging profiles. The method may still work there, but this study did not prove it.
The commercial picture is also thin. No independent source retrieved for this article shows that the 2023 method has produced a commercial battery product. None shows a quantified manufacturing cost reduction or a demonstrated pack-level capacity gain.
So the market significance is an engineering opportunity, and it has not been validated as a commercial outcome.
That honest framing is useful. It tells operators where to place a small bet and where to wait for proof.
Source: Springer, MRS Energy & Sustainability (2026); RSC EES Batteries (2026)
What This Means for Startups and Small Businesses
You do not need a synchrotron to use the lessons here. Several takeaways apply to battery startups, materials suppliers, contract manufacturers and teams that buy cells.
For battery and materials startups
- Treat coating uniformity as a product metric. The −0.40 link between carbon signal and reaction rate suggests that spread matters, not only average thickness.
- Ask for local data from suppliers. Bulk numbers can hide fast and slow regions inside particles.
- Test the assumptions inside your simulations. If your model uses a symmetric exchange-current rule, this study gives you a reason to check it.
For manufacturing operators
- Look at interface diagnostics. The authors point to the electrolyte–electrode interface as the key place to optimize.
- Plan for variation. The reviews warn that manufacturing variation can break models trained on narrow data. Log your process changes so you can tell when a model is out of range.
For founders evaluating AI tools
- Prefer models with physics and uncertainty built in. This study reports a range of error, not just a single score.
- Compare error to noise. A model that matches the experimental noise (9.6 ± 0.9% versus about 7% here) is more credible than one that claims near-perfect accuracy.
- Do not promise commercial gains yet. Right now there is no evidence of cost savings or pack-level capacity gains from this method. Investor decks should say so.
For small teams without a research lab
- Follow the electrode-scale work. The 2025 network model shows that particle findings are already feeding larger simulations. Open literature may give you usable inputs before you could produce them yourself.
- Automate your data handling. Physics-informed AI needs clean, well-labeled data. Teams that organize their test results now will be ready to use these methods later.
Source: PubMed Central: Zhao et al., Nature 2023; RSC Energy & Environmental Science (2025); Springer, MRS Energy & Sustainability (2026)
Frequently Asked Questions
Q: What is battery x ray computer vision?
A: It means using computer vision to analyze X-ray images of battery materials, pixel by pixel. In this study, it was used on X-ray movies of lithium moving in and out of LFP particles. The goal was to measure how fast the reaction runs at each spot.
Q: How did the researchers use X-ray videos to study LFP batteries?
A: They built transparent cell batteries and filmed lithium moving through LFP particles with scanning transmission X-ray microscopy. The data covered 39 particles and 62 half-cycles. They then fit a physics-based model to those images.
Q: What does lithium iron phosphate reaction kinetics mean?
A: It describes how fast lithium reacts and moves in and out of LFP. Faster kinetics generally means the battery can charge and discharge more quickly. The study found that these rates vary inside a single particle.
Q: What is physics-informed battery machine learning?
A: It is machine learning that has known physical laws built into the model. This study combined a phase-field model, PDE-constrained optimization and Bayesian inference. The aim is to get answers that are more interpretable and need less data.
Q: How accurate was the model?
A: It had a 6.8% training error and a 9.6 ± 0.9% threefold cross-validation error. The estimated experimental error was about 7%. That means the model's error was in the range of the measurement noise.
Q: How does carbon coating on lithium iron phosphate affect battery performance?
A: LFP conducts electricity poorly, so a carbon coating is needed. But the study found that regions with a thicker carbon signal reacted more slowly. The pixel-level correlation was −0.40.
Q: Does thicker carbon coating always slow LFP charging?
A: No. Some coating is necessary because LFP is intrinsically a poor electronic conductor. The finding is that excess coating can impede lithium transport, and the correlation was moderate.
Q: Did this research produce a new commercial battery?
A: No evidence found for this article shows that it has. No independent source reports a commercial product, a quantified manufacturing cost reduction or a pack-level capacity gain. The market value remains an engineering opportunity.
Q: Who funded and carried out the study?
A: It came from a team at SLAC, Stanford, MIT and Toyota Research Institute. It was supported by Toyota Research Institute through its Accelerated Materials Design and Discovery program. The X-ray work used the Advanced Light Source, a DOE Office of Science user facility.
Q: What are the limits of physics-informed AI for batteries?
A: Two 2026 reviews warn that incomplete assumptions can introduce systematic bias. They also warn that models trained on narrow lab data may fail when chemistry, cell design, manufacturing or operating conditions change. Physics-informed models are not automatically better.
Additional Links
Flowgrammer.ca – 💡 Unlock AI-Powered Automation! Let flowgrammer.ca handle complex workflows so you can focus on growing your business.
Flowgrammers.ca – 💡 Join Toronto's Fastest Growing AI and Business Automation Community.
- SRED Funding Solutions – Maximize your R&D tax credits and secure funding for your innovation.
- Silicon Scoop Podcast – Stay Informed, Stay Ahead!
- Startup Talk Podcast
- Flowgrammer Youtube
- Startup Investor Drinks – Join Toronto’s Startup Community and connect with fellow startup enthusiasts and investors.
- TorontoStarts Startup and Entrepreneur Community – Become a part of Toronto’s thriving startup ecosystem.
- Workhaus – Discover flexible workspace solutions for your business needs.
- Work with The Startup Coach – Get expert guidance and mentorship for your startup journey.
- Reach out to Jon Irwin on LinkedIn – Connect with our co-host and technology expert.
Episode transcript
Read the full transcript
Craig: here's a twist. the carbon coating that makes many lfp batteries work can also slow them down. researchers just watched it happen pixel by pixel using x-ray movies, and a physics based model. lesson is simple. more carbon is not always better. so meet lithium iron phosphate or lfp. it's cheap. it's safe. it's built from abundant elements, and it's increasingly used in electric fields... vehicles. one catch, lfp is a poor electronic conductor. so makers coat the particles with carbon. some coating is necessary. but how much and how evenly a team from slac, stanford, mit, and toyota research institute went looking inside the particles to find out? what happened here, john, when they looked?
Jon: yeah. so what they saw was, nano... so what they they reanalyzed nanoscale x-ray movies pixel by pixel. right? is... what they saw was that differences in the carbon coating surrounding lithium ion phosphate particles appear to strongly influence how quickly the lithium moves in and out of different areas of those particles. so in other words, seemingly tiny manufacturing differences at the nano scale can determine how efficiently a battery charges and discharges. so what this suggests is that, battery makers may be able to improve performance not simply by inventing new materials by... but by engineering existing materials much more precisely, which is... i think that's good because we've got this technology. we've got this lithium iron phosphate or life power lfp. right? and as you said, it's very important. it avoids expensive nickel and cobalt. it has good thermal stability. it has a strong safety record and relatively long cycle life. and so, the important reaction here is... it's called lithium intercalation. and so during discharge, lithium ions move through the electrolyte and insert themselves into the crystal structure of the lfp cathode. and so during charging, that process reverses. and i think we covered this before in another... in... we've done multiple battery things. so the useful analogy that i can tell you about for batteries is a parking garage? right? so lithium ions occupy available spaces within the solid material, but the spaces do not fill and empty uniformly. and so different parts of a particle can be at different states of charge at the same time. and so that microscopic non uniformity turns out to be extremely important because it can influence charging speed, internal stresses, and eventually degradation. so this is really cool. the thing that i like the most about this article is that essentially, up until now, they were only able to make, we'll call them sophisticated x-ray movies showing these changes. but now they're able to actually... by using computer vision and a bunch of different things, they're actually able to analyze these movies and pixel by pixel to see the... and what they came out with this. i think this is really cool. what do you think about this? do you think this is cool?
Craig: yeah. i like the science behind this, and i like how we're getting more detailed. we generally knew it worked, but didn't know how, like, the maximum mo... most efficient level, and now we start looking at this. so excess coating can actually impede lithium transport and reaction speeds buried inside single particles, which hide that when you start looking through this. so the authors of the study argue that interface between electrolyte and electrode controls much of the action, not just lithium diffusion on its own. so this, i really like. you know? this is not a black box image classifier. the team combined reaction, limit, allen, khan phase field model, pde constraint optimization, and bayesian inter... bayesian inference and uncertainty quantification. so i definitely do not... i'm not going in the physics field. there's too many words here that i have a hard time talking about.
Jon: you know what, though? the thing... the clever part that i liked about this is the computer vision and why this matters. so, like, so traditional image analysis, it might identify a particle and measure its average state of change, but they took this much further. they treated every pixel as an individual experimental measurement. and so they... approximately, i think, a 180,000 pixels across 62 movies. and so the computational framework, what they did was they analyzed how lithium concentration at each location changed over time. now the clever part about this is that they weren't merely training an ai to classify an image. they were trying to infer the underlying physical laws and parameters responsible for producing the image. and it's important because this is much closer to more, like, scientific machine learning or physics informed inverse modeling than conventional object recognition ai. and the reason why i wanted to point that out is because i think there's a lot there could be, and i haven't thought of other applications. but, like, the way that they've approached this, there's unlimited, applications for this kind of, approach in the future, i think.
Craig: yeah. and being able to optimize these types of codings that do different things for, you know, faster charging, you know yeah. it's a lever, not a product for this. so let me ask the chat. i won't... if you... if a battery got a better coating tomorrow, what would you want the most? would you want faster charging, cheaper batteries, or safer batteries? i think the answer might depend on where we're putting the battery, i guess.
Jon: i would say faster charging for me because i don't... i don't find safety to be a problem. but what do you... what about you? what would you want,
Craig: craig? i don't know what the battery inside my phone costs. so i... so inside the phone, i would think i'd want faster charging or cheaper batteries depending on the cost, percentage cost of the battery over the phone because i just don't know. my iphone, what percentage of the cost is it back? don't know. yeah. other than that, outside, you know, the regular handheld devices that go through regular testing already, we've seen so many times, those ebikes and the scooter batteries and all that kinda catching fire and cause. so if we could have safer batteries for those types of batteries, i'd be all for that. so that's why i'm saying it depends if somehow we could have a coating that stop that from, those things from actually exploding, i would be all for it.
Jon: yeah. or we could just go with one of the newer battery technologies that you and i have covered that are much safer. right?
Craig: absolutely. so here's the takeaway. lfp needs its carbon coating, but too much of it slows the lithium down. controlling that coating is a promising path to faster charging. for now, it's an opportunity, not a product. leave... let us know your comments in the chat below. we'd love to hear from you.
More from Silicon Scoop
All episodes
silicon-scoop · November 18, 2026
AI Agents' Canada Hack Attempt Sent 13 Failed Probes
AI agents Canada hack claims explained: the Transluce AI incident report found 13 attack probes in 899 requests. Learn what failed and why it matters.
View episode

silicon-scoop · November 16, 2026
OpenAI Model Misalignment Reports Now Have Deadlines
OpenAI model misalignment now has deadlines. Learn what the OpenAI misalignment framework requires, the six first incidents and what founders should do.
View episode

silicon-scoop · November 13, 2026
Methane Graphene Oxide Production Reaches 5 Grams a Day
Methane graphene oxide production gets a plasma route at Texas A&M: 5 grams a day plus hydrogen. Learn what it means for graphene oxide battery materials.
View episode
