Machine learning reveals complex behaviours in optically trapped particles

Abstract

Since their invention in the 1980s, optical tweezers have found a wide range of applications, from biophotonics and mechanobiology to microscopy and optomechanics. Simulations of the motion of microscopic particles held by optical tweezers are often required to explore complex phenomena and to interpret experimental data. For the sake of computational efficiency, these simulations usually model the optical tweezers as an harmonic potential. However, more physically-accurate optical-scattering models are required to accurately model more onerous systems; this is especially true for optical traps generated with complex fields. Although accurate, these models tend to be prohibitively slow for problems with more than one or two degrees of freedom (DoF), which has limited their broad adoption. Here, we demonstrate that machine learning permits one to combine the speed of the harmonic model with the accuracy of optical-scattering models. Specifically, we show that a neural network can be trained to rapidly and accurately predict the optical forces acting on a microscopic particle. We demonstrate the utility of this approach on two phenomena that are prohibitively slow to accurately simulate otherwise: the escape dynamics of swelling microparticles in an optical trap, and the rotation rates of particles in a superposition of beams with opposite orbital angular momenta. Thanks to its high speed and accuracy, this method can greatly enhance the range of phenomena that can be efficiently simulated and studied.

Recent Advances of Optical Tweezers–Based Dynamic Force Spectroscopy and Mechanical Measurement Assays for Live-Cell Mechanobiology

“…Alternatively, the throughput can be improved by introducing automation and artificial intelligence to complete multiple tasks and minimize human input. Examples include bead detection, bead classification, and DFS measurement [84,85]. Furthermore, neural networks could also be trained to address one of the remained obstacles, the quantification of the Brownian motion, by measuring the values of the trap stiffness and diffusion coefficient, thereby reducing OT system uncertainty [86].…”

Faster and More Accurate Geometrical-Optics Optical Force Calculation Using Neural Networks

“…We have employed fully connected NNs because they have already proved successful in similar situations. 20 The NNs have been trained using data generated with GO using the toolbox OTGO. 14 Even though the training data come with artifacts due to the finite number of rays, both the NN architecture and the training process are designed to obtain NN predictions that get rid of these artifacts.…”

“…Recently, neural networks (NNs) have been demonstrated to be a promising approach to improve the speed of optical force calculation for spheres using the T-matrix method. 20 NNs are able to use data to adapt their solutions to specific problems. 21 These algorithms have proved to improve on the performance of conventional ones in tasks such as determining the scattering of nanoscopic particles, 22 enhancing microscopy, 23 tracking particles from digital video microscopy 24 or even epidemics containment.…”

Optical Tweezers Exploring Neuroscience

“…Simulating particles near walls, at the tip of optical fibers (i.e., FOT) or near plasmonic structures is often more complicated, and tools such as finite difference time domain (Yee, 1966;Benito et al, 2008;Lenton et al, 2017), discrete dipole approximation (Oskooi et al, 2010;Loke et al, 2011;Yurkin and Hoekstra, 2011), surface integral methods (Ji et al, 2014), or commercial packages such as COMSOL (Zhang et al, 2016;2020a) and Lumerical (David et al, 2018;2020b) are often used. Recent developments in machine learning have led to faster methods of simulating particles in OT (Lenton I. C. D. et al, 2020c), and faster hybrid algorithms for optimizing and simulating light scattering (Jiang et al, 2020). These advances could be useful in designing optical potentials that optimize certain OT properties such as trap stiffness and particle orientation.…”