romtools.workflows.inverse#
Inverse workflows estimate unknown model parameters from observed data.
An inverse problem starts from a forward model, experimental or synthetic observations, and a parameterized uncertainty model. The goal is to identify parameter values, or a posterior distribution over parameter values, that make the model predictions consistent with the observations.
romtools currently supports:
Ensemble Kalman inversion (EKI) for derivative-free parameter calibration.
Multifidelity EKI with control variates and adaptive reduced-order model refresh strategies.
Variational inference (VI) with gradient, Adam, and Newton optimizers for Gaussian variational families, including diagonal and full-covariance VI.
Multifidelity VI with control variates and adaptive reduced-order model updates, including full-covariance natural-coordinate Newton optimization.
Modules
Single-fidelity efficient global optimization drivers. |
|
Optimization methods and configs used by inverse EGO workflows. |
|
Bound-handling adapters for EKI and MF-EKI. |
|
Single-fidelity ensemble Kalman inversion drivers. |
|
Multifidelity ensemble Kalman inversion drivers. |
|
Gaussian variational inference workflows and optimizer configurations. |