With 80% accuracy, deep learning identifies nanoparticle shapes in Brownian motion

Editor|Pothos

With the practical application of nanoparticles in the medical, pharmaceutical and industrial fields, it is necessary to understand the properties and properties of each nanoparticle from the perspective of materials. Evaluation of agglomeration status and quality control

One way to evaluate nanoparticles in liquids is to analyze the trajectories of Brownian motion. Although nano-tracking analysis (NTA) is a simple method for measuring individual particles from micro to nanoscale, its inability to assess the shape of nanoparticles has been a long-standing problem. NTA always assumes sphericity when quantifying particle size using the Stokes-Einstein equation, but cannot verify whether the measured particles are actually spherical.

A research team from the University of Tokyo has proposed a new method for evaluating the shape anisotropic properties of nanoparticles, solving a problem of nanoparticle evaluation that has existed since Einstein's time.

The researchers built a deep learning (DL) model to predict the shape of nanoparticles using time-series trajectory data of Brownian motion obtained from NTA measurements. By using the ensemble model for trajectory analysis, the deep learning model was able to achieve about 80% single-particle classification accuracy for two gold nanoparticles of approximately the same size but different shapes, which traditional NTA cannot distinguish individually.

In addition, studies have shown that the mixing ratio of spherical nanoparticles and rod-shaped nanoparticles can be quantitatively estimated based on measurement data of nanoparticle mixed samples. This result demonstrates that by applying dynamic light scattering analysis (DL) to nanoparticle tracking analysis (NTA) measurements, it is possible to evaluate particle shape, which was previously thought to be impossible. Published on "APL Machine Learning", the theme is "Using Deep Learning to Analyze Brownian Motion Trajectories of Non-spherical Nanoparticles"

Using the characteristics of light scattering and Brownian motion, NTA ( Nanoparticle tracking analysis) is a detection method used to measure the particle size distribution of samples in liquid suspensions and has been widely used commercially. NTA used a theoretical formula proposed by Einstein more than 100 years ago to calculate the diameter of a particle. The trajectory of Brownian motion reflects the influence of the shape of the particle, but in practice it is difficult to measure extremely fast motion. Furthermore, even if the particles are non-spherical, traditional analysis methods are inaccurate because they unconditionally assume that the particles are spherical and use the Stokes-Einstein equations for analysis.

However, using deep learning, which is good at finding hidden correlations in large-scale data, it is possible to detect differences caused by shape differences even if the measurement data is average or contains inseparable errors.

A research team led by Professor Takanori Ichiki of the University of Tokyo successfully developed a deep learning model. This model can identify shapes from measured Brownian motion trajectory data without requiring changes to experimental methods. In order to simultaneously consider the time series changes of data and the correlation with the surrounding environment, they used a one-dimensional convolutional neural network (1D CNN) model to extract local features and combined it with a bidirectional LSTM model with temporal dynamic aggregation capabilities

The development of shape estimation models usually includes three stages: first, NTA measurements of raw data collection, then creation of datasets and models for deep learning, and finally deep learning training

Study using time series data with different trajectory lengths (20, 40, 60, 80 and 100 frames) by changing four models (MLP, LSTM, 1D CNN and 1D CNN Bi-LSTM ) to verify the convergence of learning.

The accuracy of both LSTM and 1D CNN models at 100 frames is above 80%, which shows that local features and temporal dynamics are extracted through convolution Accumulation is an effective method for extracting shape features. At the same time, the high accuracy indicates that the shape classification of nanoparticles in liquids has reached the realistic level of single particle analysis with NTA and DL.

After deep learning analysis, we successfully classified individual nanoparticles in liquids according to their shapes, and the accuracy was very high, reaching a practical level. Meanwhile, in this study, we also established a calibration curve for determining the mixing ratio of the mixed solution of spherical and rod-shaped nanoparticles. Considering the currently known types of nanoparticle shapes, we believe that this method can effectively detect the shape of nanoparticles

Illustration: Particle analysis system using microcapillary chip , using NTA technology to obtain the particle size distribution of the mixture from the measurement results of Brownian motion. (Quote from: paper)

The traditional NTA method cannot directly observe the shape of the particles, and the characteristic information obtained is limited. Using the DL method, even particles of different shapes with the same hydrated diameter can be distinguished from mixtures based on their trajectories.

In the study, they sought to determine the shape of both particles, but given the types of shapes of commercially available nanoparticles, they believe this method could be used for practical applications, such as detecting nanoparticles in homogeneous systems. foreign body. The extension of NTA can be applied not only in research but also in industry, such as evaluating the properties, agglomeration state and homogeneity of nonspherical nanoparticles, and quality control.

The researchers said: "It will be an interesting research direction to extend the measurement objects of particles to various shapes and materials, and future research topics will be to test the applicability of the DL NTA method."

In particular, it is expected to be a solution for evaluating the properties of various biological nanoparticles, such as extracellular vesicles, in an environment similar to that of living organisms. It also has the potential to become an innovative method for fundamental research on the Brownian motion of nonspherical particles in liquids.

Paper link: https://doi.org/10.1063/5.0160979

Reference content: https://phys.org/news/2023-10-deep-long-standing-identification -nanoparticle.html

The above is the detailed content of With 80% accuracy, deep learning identifies nanoparticle shapes in Brownian motion. For more information, please follow other related articles on the PHP Chinese website!