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…