Molecular Dynamics (MD) is a powerful computational tool that lets scientists and engineers study chemical, biological, or material systems at a micro-or nano-scale. In particular, we target a materials science application of molecular self-assembly in which the goal is to model the dynamics and structure of bulk systems containing many particles that interact with one another via a specified potential energy function. By simulating the motion and interaction of particles in a molecular system, material properties can be measured from the resulting equilibrated particle structures. While MD undoubtedly provides engineers with the capacity to perform high-fidelity material simulations, it is not without its own limitations, namely computational expense. For one, very large systems (i.e. with many particles) are required to emulate the properties of a bulk material as accurately as possible.

Learning pair interactions from experimental or simulation data is of great interest for molecular simulations. We propose a general stochastic method for learning pair interactions from data using differentiable simulations (DiffSim). DiffSim defines a loss function based on structural observables, such as the radial distribution function, through molecular dynamics (MD) simulations. The interaction potentials are then learned directly by stochastic gradient descent, using backpropagation to calculate the gradient of the structural loss metric with respect to the interaction potential through the MD simulation. This gradient-based method is flexible and can be configured to simulate and optimize multiple systems simultaneously. For example, it is possible to simultaneously learn potentials for different temperatures or for different compositions. We demonstrate the approach by recovering simple pair potentials, such as Lennard-Jones systems, from radial distribution functions. We find that DiffSim can be used to probe a wider functional space of pair potentials compared to traditional methods like Iterative Boltzmann Inversion. We show that our methods can be used to simultaneously fit potentials for simulations at different compositions and temperatures to improve the transferability of the learned potentials.