Exploring Downscaling in High-Dimensional Lorenz Models Using the Transformer Decoder
Original article: https://doi.org/10.3390/make6040107
Summary
This study tests whether machine-learning models can recover small-scale variables from large-scale variables in high-dimensional Lorenz systems. Linear regression, feedforward neural networks, and transformer-decoder models are trained on generalized Lorenz model output to estimate secondary-scale variables from primary-scale inputs. The work frames downscaling as a bridge between nonlinear dynamics and AI, using an idealized system where the relationship between scales is known well enough to evaluate model behavior.
