Original version of FFJORD does not have a support of conditional input. To address this issue we rewrote one of the base layers that were used in FFJORD library. The example of such statistics is presented in Fig 6. Figure 6: An example of monitored statistics during surrogate training for one iteration of optimization. The gradient bias is calculated per component of the gradient vector, i.e., if C.1 Procedure For Mixing Matrix Generation 10-dimensional mixing matrix A could be generated with the following Python code: C.3 Numerical Derivatives To obtain numerical derivatives of R we are using central difference scheme: f Muons are bent by the magnetic field and simultaneously experience stochastic scattering as they pass through the magnet which causes random variations in their trajectories. Color represents number of the hits in a bin.

We propose a novel method for gradient-based optimization of black-box simulators using differentiable local surrogate models. In fields such as physics and engineering, many processes are modeled with non-differentiable simulators with intractable likelihoods. Optimization of these forward models is particularly challenging, especially when the simulator is stochastic. To address such cases, we introduce the use of deep generative models to iteratively approximate the simulator in local neighborhoods of the parameter space. We demonstrate that these local surrogates can be used to approximate the gradient of the simulator, and thus enable gradient-based optimization of simulator parameters. In cases where the dependence of the simulator on the parameter space is constrained to a low dimensional submanifold, we observe that our method attains minima faster than all baseline methods, including Bayesian optimization, numerical optimization, and REINFORCE driven approaches.