Christopher Jarzynski – University of Maryland
Wednesday, July 29, 2026
Zoom opens: 10:30AM EDT
Seminar begins: 10:45AM EDT
Estimating free energy differences with virtually escorted trajectories
The convergence of numerical free energy estimation methods can be accelerated using artificial fields that “escort” simulated trajectories along near-equilibrium paths. Unfortunately, designing such fields is not easy. Taking a cue from the mathematics behind diffusion models – a class of generative models in machine learning – we introduce a method based on virtual escorting. This method adopts a post-processing approach. Given a fixed set of nonequilibrium trajectories, a parameter-dependent virtual escorting field is constructed, possibly using a neural network. This field is used in combination with the trajectories to produce an estimate of the desired free energy difference. The parameters are then adjusted to optimize the convergence of the estimate. I will describe the method, and will discuss conditions under which it produces a zero-variance estimator of the free energy difference.