materials-chemistry/ai produced the result/Nature Communications 2020 · v2
Neural network segments 650 battery particles to measure their detachment after cycling
Researchers X-rayed lithium-ion cathodes and used an image-recognition network to outline every active particle in the three-dimensional scans, then measured how far each had pulled away from the conductive matrix around it.
spectrum · one line per step, placed by what the step does · bright lines used AI
Machine-learning-revealed statistics of the particle-carbon/binder detachment in lithium-ion battery cathodes
Nature Communications, 2020
doi:10.1038/s41467-020-16233-5 · record aix-00013 v2 · checked 2026-10-07
- AI was for
- Segmentation, Detection
- Model family
- Convolutional neural network
- Checked by
- Held-out
- Code
- available
The finding the paper is about came from the AI.
What this research was about
A lithium-ion battery cathode is not a solid block. It is a packed mixture of small grains of active material, which store and release lithium, held together by a matrix of carbon and a polymer glue. The carbon carries electrons to and from each grain. If a grain loses contact with that matrix, electrons can no longer reach it easily, and the grain contributes less to the cell. The grains swell and shrink as the battery charges and discharges, which can prise them loose. Seeing this is hard, because the contact between a grain and its surroundings lies buried inside the electrode and is a few millionths of a metre across.
The researchers studied nickel-rich NMC cathodes, a common cathode chemistry. They cycled electrodes at a slow rate and a fast rate, imaged them with hard X-ray phase contrast nano-tomography, which builds a three-dimensional picture of the interior, and set out to measure the degree of detachment of individual grains rather than describing the electrode only as an average. They also modelled how electrical resistance varies across a grain's surface.
Where AI came in
The AI did the outlining. The three-dimensional scan is a stack of image slices in which hundreds of grains touch and overlap. A network called Mask R-CNN, designed to find separate objects in a picture and trace the outline of each, was started from weights learned on a large collection of everyday photographs and then trained further on slices that people had traced by hand. Information about the expected shape of the grains was added as a further constraint. The trained network was then run on the scan of the electrode under study, slice by slice, giving each grain an outline and its own label.
A separate, non-learned step stitched the labels across depth to assemble whole grains in three dimensions. Those labelled grains supplied the input for every per-grain measurement that followed, including the detachment figures for more than 650 grains and the resistance model. The AI stood in for tracing by hand, which the researchers describe as unworkable at that number. A conventional segmentation method, watershed, was run alongside for comparison.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
Researchers imaged Ni-rich NMC composite cathodes with hard X-ray phase contrast nano-tomography and used a Mask R-CNN model, initialised from ImageNet weights and trained on manually annotated tomographic slices, to identify and segment individual active particles in the reconstructed volumes. Over 650 particles from electrodes cycled at 0.1 C and 1 C were then measured for their degree of detachment from the carbon/binder matrix. The distributions indicate that faster-cycled particles show more severe detachment, and that particles smaller than 11 μm in diameter show a broader spread of detachment values. A separate numerical model mapped relative electrical resistance over particle surfaces, and the retrieved electron density of one particle was compared with a Ni K-edge oxidation state map of the same particle, giving a Pearson correlation coefficient of about 0.54.
How AI was used
A Mask R-CNN instance segmentation network with a ResNet-101 feature pyramid network backbone was initialised from ImageNet pre-trained weights and further trained on manual annotations of nano-tomographic slices of thick NMC composite electrodes, with particle shape information incorporated as an additional constraint during optimisation. The trained model was then applied to the reconstructed phase contrast volume of the monolayer electrode dataset, processing each slice separately to produce per-particle bounding boxes and binary masks carrying unique identifiers. Those identifiers were linked across slice depths with the Hungarian maximum matching algorithm, a non-learned step, to assemble 3D particles. The resulting particle labels supplied the inputs for the per-particle quantification of detachment, volume and electron density, and for the numerical model of surface electrical resistance; the tomographic acquisition, phase retrieval, reconstruction and resistance modelling used conventional, non-learned methods. A watershed-based conventional segmentation was run as a comparison baseline.
The shape of the work
Structural · the record, drawn
no AI
Fabricate and electrochemically cycle NMC cathodes
Physical execution, by hand or by robot.
Two cells were both cycled under C/10 for the first cycle and 1 C for the second cycle as an activation process.where the paper describes this · verbatim
no AI
Acquire hard X-ray phase contrast nano-tomography
Obtaining raw data, whether by measurement, download or retrieval.
For every tomography scan, 1500 projections were acquired with 0.2 s exposure time.where the paper describes this · verbatim
no AI
Phase retrieval and tomographic reconstruction
Cleaning, filtering, normalising or labelling data already obtained.
These tomograms were subsequently used for phase retrieval to generate 2D phase maps.where the paper describes this · verbatim
AI
Train Mask R-CNN on manually annotated slices
Fitting model parameters, including fine-tuning an existing model. The AI stood in for manual curation.
Manual annotations on a set of nano-tomographic slices of thick NMC composite electrodes (see Fig. 1a) were used to train a machine-learning modelwhere the paper describes this · verbatim
AI
Segment individual particles slice by slice
Running a trained model over new data to predict, classify or score. The AI stood in for manual curation.
The instance-aware identification and segmentation of NMC particles in each slice are accomplished using a state-of-the-art mask regional convolutional neural network (Mask R-CNN).where the paper describes this · verbatim
no AI
Link masks across depths into 3D particles
Cleaning, filtering, normalising or labelling data already obtained.
Those identifiers are then used for linking slices at different depths of the volume to construct 3D particleswhere the paper describes this · verbatim
no AI
Model local electrical resistance over particle surfaces
Numerical or physics simulation, including where a learned surrogate replaces it.
we developed a numerical model to calculate the spatial distribution of the relative electrical resistance over the surface of the NMC particleswhere the paper describes this · verbatim
no AI
Quantify detachment statistics per particle
Extracting understanding from model behaviour.
we quantified the characteristics of every single NMC particles including their degree of detachmentwhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
The reported statistics of particle detachment rest on the Mask R-CNN identification and segmentation of the >650 particles; the manual alternative is described as infeasible at that scale.
the ResNet-101 feature pyramid network was used and the model was initialized by the weights obtained from the large-scale ImageNet datasetwhere the paper describes this · verbatim
For a more quantitative comparison of the results from the conventional watershed segmentation and the herein developed Mask R-CNN algorithmwhere the paper describes this · verbatim
The source code and detailed instructions are made publicly available at the GitHub repositorywhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- How many were testedThe paper gives no count of what was tested.
- Version of Mask R-CNN (ResNet-101 feature pyramid network backbone)Which version of the model was used is not stated.
About this article
Record aix-00013, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error