Nouveau Recherche PDF, HTML, DOCX et bien d'autres formats.

Prépublication · 2026

Approximate Adjoint Models for Memory and Compute-Efficient Online Learning of Hybrid Models

Emilio González-Zamora, Pierre Tandeo et al. — Monde

Source : arXiv

Résumé

Differentiable programming enables online calibration of numerical models by minimizing errors accumulated along simulated trajectories. This includes estimating physical parameters and training neural sub-models in hybrid configurations. Gradients are typically computed using adjoint methods, or backpropagation through time (BPTT), whose computational and memory costs grow with the length of the training window. This limits the duration and resolution of hybrid-model training, a major bottleneck for realistic geophysical applications. Approaches like truncated BPTT reduces memory use by disca…

Actions

Citation