Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper introduces a minimal normalizing flow (MNF) approach to accelerate chemical potential calculations in molecular simulations, particularly for challenging systems like electrolytes. By intentionally limiting the expressivity of the flow and employing low-dimensional, physically informed transformations, the MNF can be trained rapidly, achieving significant improvements in sampling efficiency. The authors demonstrate that their method can accelerate chemical potential calculations by at least 10 times and effectively increase sample sizes for charging free energy calculations by 3 times and for force field perturbations by 8 times.
Accelerating chemical potential calculations by over 10 times with a novel minimal normalizing flow approach could revolutionize how we simulate complex molecular systems.
Chemical potentials are among the most important properties that can be obtained from a molecular simulation since they define many technologically relevant collective properties. The chemical potential of a species in solution is obtained by computing the free energy change of adding that species into a bulk system, a calculation typically very expensive for systems such as electrolytes, due to the lack of phase space overlap between "not-inserted" and "inserted" states. Recently, normalizing flows have been introduced as a way to accelerate free energy computations by learning a bijective function, constructed to be as expressive as possible, that maps the configuration space of one Boltzmann distribution onto another. This expressivity makes them difficult to train, limiting their ability to be generated "on-the-fly" for any new system, and in practice these mappings have shown only modest sampling improvements for liquids. We address these issues by introducing a "minimal" normalizing flow (MNF). This is a trainable bijective mapping that is intentionally limited in expressivity, and instead applies low-dimensional, physically informed transformations. Useful MNFs can be trained in 1 minute of GPU time due to their simplicity and our introduction of a novel training strategy. We show how calculations of chemical potentials of Lennard-Jones particle systems can be accelerated by at least 10 times with a simple radial mapping. We also apply a radial and orientational mapping to ion solvation in water, showing that MNFs can increase the effective sample size by 3 times for charging free energy calculations and 8 times for calculating free energy changes due to force field perturbations. This provides the foundation for the development of physically-informed mappings that can accelerate complex free energy calculations while retaining low training costs.