Extended Data Fig. 4: Binary-black-hole modelling problem.
From: State estimation of a physical system with unknown governing equations
a, Orbital trajectory estimates. b, Trajectory estimates versus time. In this experiment, we approximate the dynamics using a neural network. This is a useful model in situations in which it is not clear what dictionary of basis functions is appropriate. The left figure shows the predicted orbital trajectories in the plane of the orbit and the right figure shows the predicted orbital trajectories as a function of time. The black lines indicate the data windows, the orange lines indicate the testing data and the blue intervals indicate probabilistic predictions from our model. Note that we only have access to the waveform observations, w(t), up to time 0.6 × 105 in training. We are able to accurately estimate the states and provide probabilistic estimates for the orbital trajectories well into the future, despite not having access to the underlying governing equations.
