Traditional Molecular Dynamics (MD) simulators deal with large numbers of particles at a very high scale. MD simulations usually start by placing the points in a structured topology such that some preconditions are respected, for example a grid-based approach. In this work we develop a generative component that creates a first candidate for an initial configuration based on a conditional diffusion model, along with a corrective procedure that replaces the points such that a relaxation constraint is respected in the entirety of the initial pointcloud. The corrector procedure is attention-based and scales linearly to larger pointclouds. In contrast to other relaxation approaches, like force fields, which are iterative and conditioned on a soft margin. Our model is able to displace the points correctly in a single try, while at the same time following a hard constraint.     «

Traditional Molecular Dynamics (MD) simulators deal with large numbers of particles at a very high scale. MD simulations usually start by placing the points in a structured topology such that some preconditions are respected, for example a grid-based approach. In this work we develop a generative component that creates a first candidate for an initial configuration based on a conditional diffusion model, along with a corrective procedure that replaces the points such that a relaxation constraint...     »