At LNE, artificial intelligence is accelerating the observation of nanoparticles

The national metrology and testing laboratory (LNE) chose May 20, the date of World Metrology Day, to unveil a brand new characterization platform dedicated to nanomaterials. Called NanoMétrologIA, this tool developed by the LNE is distinguished by its use of deep learning algorithms to accelerate the dimensional analysis of nanoparticles, from images obtained by scanning electron microscopy (SEM)

The characterization of nanoparticles is indeed a process that is usually long and tedious for researchers. ” Very small particles tend to agglomerate and form clusters with irregular grains “, explains Nicolas Fischer, head of the “data science” department at LNE “ The characterization of a sample requires the analysis of approximately 500 particles, which must be isolated on the image by clipping them. This step was previously done manually. “.

A neural network of Facebook

From 2019, LNE researchers trained the algorithm of the NanoMetrologIA platform to recognize the particles appearing on an SEM image and to separate them. The time saved is spectacular: an operation that previously required half a day can now be carried out in about ten seconds!

NanoMetrologIA’s neural network characterizes the dimensions of several thousand nanoparticles that may be present in an SEM image.

To develop this tool, the LNE data analysis team used an algorithm of deep learning originally developed by Facebook and trained it with SEM images annotated by researchers. The architecture of the neural network has also been reviewed to gain efficiency with a reduced database, says Loïc Coquelin, research engineer at LNE.

A hundred real images were used, as well as 1000 simulated images. The latter were produced by randomly assembling photos of previously cropped particles. A technique that makes it easy to produce agglomerates while knowing the dimensional characteristics of the nanoparticles present in the image.

Adapt to other nanomaterials

At first we focused on the identification of titanium dioxide nanoparticles (TiO2), for which there is a real challenge in the food industry. [auparavant utilisé en tant que colorant, le TiO2 a été interdit dans les aliments en janvier 2020. La particule est donc sous surveillance, ndlr]highlights Loïc Coquelin. But it will be possible to adapt this method in the short term to particles with a geometry close to TiO2, such as silica (SiO2), sodium chloride (NaCl), or even gold and silver. It will not be necessary to completely retrain the network for these materials. »

Beyond the development of this tool, the LNE is working on a method for evaluating uncertainties adapted to the use of deep learning. ” This is an important issue for establishing reference guides and a methodology. The uncertainty of the results provided by the platform must conform to a certain standard, taking into account both the uncertainty of the algorithm used and the uncertainty of the input data. This is important information in metrology. », explains Nicolas Fischer. This work is carried out by a doctoral student, as part of a collaboration with the Ecole Polytechnique.

Selected for you

An ultra-fast tabletop X-ray source for spintronics... And 4 other research advances

We want to give thanks to the author of this short article for this remarkable content

At LNE, artificial intelligence is accelerating the observation of nanoparticles

Take a look at our social media profiles as well as other pages related to them