Measuring force fields is fundamental to understanding the properties of microscopic particles, including their structure, shape fluctuations and interactions within their environment. These properties can determine, for instance, the elasticity of cells, nanoparticle diffusion in liquid thin films and how exotic particles behave in plasma.
The calibration of these force fields often relies on the trajectory analysis of the particle influenced by Brownian motion. This can be difficult considering the plethora of parameters that researchers have to take into account or when data is scarce, and determining these parameters are often computationally costly.
To streamline the complex nature of force-field calibration, Argun et al. present DeepCalib, a machine-learning method that can outperform standard calibration methods and produce accurate calibrations where no standard exists.
The Python-based software, which can be optimized for specific force fields, is based on recurrent neural networks (RNNs). RNNs address data patterns that change over time based on constant feedback, allowing them to learn from past events to make predictions about the future.
The researchers benchmarked DeepCalib against standard techniques on experiments that included determining the Brownian motion of a polystyrene particle 200 nanometers in diameter, confined in a temperature field inside a nanofabricated structure. In this case, DeepCalib calibrated a range of force fields, including rotational force fields.
Experimental results showed that DeepCalib outperforms standard methods in challenging conditions involving short and/or low frequency measurements and can be applied to non-equilibrium, unsteady force fields for which no simple standard technique exists.
The researchers plan to extend DeepCalib’s functionalities to more-complex biological systems, such as proteins moving through blood.
Source: “Enhanced force-field calibration via machine learning,” by Aykut Argun, Tobias Thalheim, Stefano Bo, Frank Cichos and Giovanni Volpe, Applied Physics Reviews (2020). The article can be accessed at https://doi.org/10.1063/5.0019105.
