Convolutional neural networks applied to differential dynamic microscopy reduces noise when quantifying heterogeneous...

Area of Science:

• Soft matter physics
• Biophysics
• Materials science

Background:

• Differential dynamic microscopy (DDM) requires extensive imaging data for accurate motion quantification in soft matter.
• Short or dynamic imaging sequences yield noisy and unreliable DDM results, limiting applications.

Purpose of the Study:

• To develop and validate a machine learning-based denoising method for DDM.
• To improve DDM accuracy with limited frames and capture fast-evolving dynamics.

Main Methods:

• A convolutional neural network encoder-decoder (CNN-ED) model was employed to denoise the intermediate scattering function computed via DDM.
• The method was tested on colloidal particle suspensions undergoing gelation and exhibiting viscosity gradients.

Main Results:

• The CNN-ED model effectively reduced noise in DDM analysis, even with limited imaging frames.
• Accurate quantification of particle diffusivity was achieved during fluid gelation and across viscosity gradients.
• The denoising approach successfully captured time-varying and spatially varying particle dynamics.

Conclusions:

• Machine learning-enhanced DDM provides a robust solution for analyzing complex dynamics in soft matter systems.
• This method expands the applicability of DDM to scenarios with limited data or non-equilibrium conditions.
• The approach holds promise for high-throughput screening and studying dynamic processes.