fpsl

Fokker-Planck Score Learning (FPSL): Score-based diffusion models for periodic data.

FPSL is a Python package for training and sampling from score-based denoising diffusion models, with specialized support for periodic boundary conditions and force-conditioned generation. The package implements the Fokker-Planck Score Learning approach for learning score functions on toroidal (circular) domains.

This package contains the following main components:

• ddm: This module implements the core FPSL class for learning the equilibrium free energy (PMF) from biased samples. It includes neural network architectures, noise scheduling, prior distributions, and force conditioning schedules for diffusion processes.
• datasets: Collection of one-dimensional potential energy landscapes and biased-force variants for testing and benchmarking diffusion models.
• utils: Utility classes and functions including Gaussian mixture models, numerical integrators for stochastic differential equations, and base classes.

To get started, please have a look at the tutorials.