Scalable Differentiable Programming via Trajectory Reweighting

Differentiable programming is an emerging paradigm in scientific computing. However, its application to practical systems is inhibited by (i) numerical instabilities and high memory costs, and (ii) the necessity to reimplement an existing codebase in an automatic differentiation framework. In this talk, we demonstrate how a simple physics-informed reweighting scheme enables gradient calculation with respect to existing scientific computing codes. This method can be applied generally to any oracle which samples states from a known probability distribution (e.g. Monte Carlo schemes, enhanced sampling routines, quantum simulations, Focker-Planck). We will here illustrate proof-of-concept applications alongside two popular simulation packages, namely oxDNA and HOOMD-Blue, in order to design DNA sequences and assemble a honeycomb lattice, respectively. In future work, we aim to extend this method to non-equilibrium settings and further validate it in experimentally-relevant systems.