Physically Organized Latent Spaces in Unsupervised Autoencoders: Evidence from Aerodynamic Databases
Abstract
Using aerodynamic databases of progressively increasing complexity - from linear analytical theory to nonlinear separated RANS flows - we demonstrate that an autoencoder, trained solely by reconstruction, can spontaneously organize the latent representation according to physically meaningful variables and aerodynamic laws. Although the latent coordinates themselves vary under random initialization, the underlying physical organization is reproducible, even quantitatively. In addition, we present a three-dimensional latent space representation of the airfoil flow obtained by an unsupervised autoencoder with sequential training in which the inviscid field is decoupled from the boundary layer effects with the third latent variable strongly correlated with the Reynolds number and aerodynamic drag. Cross-validated regression and affine-alignment measures show that physical parameters and responses are strongly and linearly encoded in the learned coordinates, consistently across independent trainings.
Related
- On the aerodynamic forces on heaving and pitching airfoils at low Reynolds number
- Three-dimensional effects on the aerodynamic performance of flapping wings in tandem configuration
- Influence of Wind Unsteadiness and Inhomogeneity on the Mean and Fluctuating Aerodynamic Force Acting on a Wingsail
- Geometry Parameterisation via a Structural Modal Basis for Aerodynamic Shape Optimisation
- Aerodynamic Shape Optimization of unsteady flapping wings
BibTeX
@misc{tognaccini2026physically,
title = {{Physically Organized Latent Spaces in Unsupervised Autoencoders: Evidence from Aerodynamic Databases}},
author = {Renato Tognaccini and Ettore Saetta and Gianluca Iaccarino},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.08451},
eprint = {2610.08451},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.08451},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Physically Organized Latent Spaces in Unsupervised Autoencoders: Evidence from Aerodynamic Databases
AU - Renato Tognaccini
AU - Ettore Saetta
AU - Gianluca Iaccarino
PY - 2026
DA - 2026/10/06
PB - arXiv
DO - 10.48550/arxiv.2610.08451
UR - https://arxiv.org/abs/2610.08451
AB - Using aerodynamic databases of progressively increasing complexity - from linear analytical theory to nonlinear separated RANS flows - we demonstrate that an autoencoder, trained solely by reconstruction, can spontaneously organize the latent representation according to physically meaningful variables and aerodynamic laws. Although the latent coordinates themselves vary under random initialization, the underlying physical organization is reproducible, even quantitatively. In addition, we present a three-dimensional latent space representation of the airfoil flow obtained by an unsupervised autoencoder with sequential training in which the inviscid field is decoupled from the boundary layer effects with the third latent variable strongly correlated with the Reynolds number and aerodynamic drag. Cross-validated regression and affine-alignment measures show that physical parameters and responses are strongly and linearly encoded in the learned coordinates, consistently across independent trainings.
ER -