| Author | Rodion Akinzhala |
| Consultant | Alexander Terentyev |
| Advisors | Yury Maximov, DSc |
| Vadim Strijov, DSc |
Operator learning methods typically approximate the solution map of a physical system with a single global model — an assumption that fails when solutions change character fundamentally across the parameter space.
We propose TEMPO: a framework that identifies latent dynamical regimes in snapshot data via Expectation-Maximization, builds a dedicated Neural-POD basis for each, and learns to route new inputs to the right regime at inference time. On multi-regime benchmarks, TEMPO achieves substantially lower errors than POD-DeepONet and vanilla DeepONet, with interpretable structure that reflects the underlying physics.
Operator learning, DeepONet, Proper Orthogonal Decomposition, Expectation-Maximization, Gaussian mixture models, Regime discovery