Numerical methods for unraveling inter-particle potentials in colloidal suspensions: A comparative study for...

We compare three model-free numerical methods for inverting structural data to obtain interaction potentials, namely, iterative Boltzmann inversion (IBI), test-particle insertion (TPI), and a machine-learning (ML) approach called ActiveNet. Three archetypal models of two-dimensional colloidal systems are used as test cases: Weeks-Chandler-Anderson short-ranged repulsion, the Lennard-Jones potential, and a repulsive shoulder interaction with two length scales. Additionally, data on an experimental suspension of colloidal spheres are acquired by optical microscopy and used to test the inversion methods. The methods have different merits. IBI is the only choice when the radial distribution function is known but particle coordinates are unavailable. TPI requires snapshots with particle positions and can extract both pair- and higher-body potentials without the need for simulation. The ML approach can only be used when particles can be tracked in time and it returns the force rather than the potential. However, it can unravel pair interactions from any one-body forces (such as drag or propulsion) and does not rely on equilibrium distributions for its derivation. Our results may serve as a guide when a numerical method is needed for application to experimental data and as a reference for further development of the methodology itself.