Presentation Information
[C-1-03]Evaluation of Initial-Condition Generalization of Leapfrog-RNN in Harmonic Oscillator ODE Prediction
◎Tomohiro Sakai1, Reon Oshio1, Yukihisa Suzuki1 (1. Tokyo Metropolitan University)
Keywords:
Surrogate Model,Neural Network,Ordinary Differential Equation
In surrogate models that replace physical simulations with machine learning, what is required is not only the reproduction of trained trajectories but also stable time-evolution prediction for untrained initial conditions (ICs). The authors have previously proposed the Leapfrog-RNN architecture, which embeds the update structure of the Leapfrog method—a numerical integration scheme—into the input-output relationship of a Recurrent Neural Network (RNN), and demonstrated that it enables more accurate and stable learning than a conventional RNN in the fixed-trajectory prediction of a harmonic oscillator. The aim of this paper is to clarify whether the introduction of this Leapfrog structure improves generalization performance for untrained ICs. To this end, we extend the approach to variable trajectories with differing amplitudes and quantitatively evaluate its effect by comparing long-term prediction accuracy for both ICs within the trained amplitude range (interpolation) and large-amplitude ICs outside the trained range (extrapolation).