Fig. 2: Order of convergence of the proposed Euler gradient approximations. | Communications Physics

Fig. 2: Order of convergence of the proposed Euler gradient approximations.

From: Online calibration of deep learning sub-models for hybrid numerical modeling systems

Fig. 2: Order of convergence of the proposed Euler gradient approximations.

We provide experimental validation of the gradient based on both the EGA (17) and Static-EGA (19) formulas. The error is computed by comparing these gradients to the exact ones, returned using automatic differentiation of the solver Ψ. We recall that in the EGA, the Jacobian of the flow is computed exactly using automatic differentiation. The dashed line correspond to the h2 slop and indicates the order of convergence of the proposed methods.

Back to article page