About HydroBayesCal

What is HydroBayesCal?

HydroBayesCal is a Python 3 package for the optimisation and calibration of hydrodynamic and morphodynamic models using a surrogate-assisted Bayesian Active Learning approach. It trains a Gaussian Process Emulator (GPE) as a fast surrogate (metamodel) of a full-complexity numerical model, and evaluates that surrogate using Bayesian model evidence (BME) and relative entropy (RE), following Oladyshkin et al. (2020).

Because every parameter update must be propagated back into the numerical model, HydroBayesCal couples to fully open-source modelling software through a common binding layer:

  • TELEMAC (2D and 3D) — fully supported.

  • OpenFOAM (interFoam) — binding under active development.

  • Delft3D-FLOW — binding planned (placeholder interface in place).

The architecture is deliberately solver-agnostic: a binding only needs to implement how the model is run and how outputs are extracted (see Code Structure), so support for additional solvers can be added with limited effort.

Good to know

To work with HydroBayesCal and TELEMAC, familiarise yourself with the TELEMAC software. Useful starting points:

Purpose and motivation

Stochastic calibration techniques require a large number of full-complexity model realisations to perform statistical analysis. This is unfeasible when a single realisation takes hours or days. HydroBayesCal makes it tractable by first constructing a surrogate model from only a few realisations at carefully chosen initial collocation points, sampled with advanced design-of-experiments methods.

The package builds on BayesValidRox for the design of experiments, and uses Gaussian Process Regression (GPR) to build single-output and multi-output surrogate models. These predict the model outputs for any parameter combination, and Bayesian inference then quantifies the uncertainty of the calibration parameters.

Bayesian Active Learning (BAL) is used to iteratively add new training points (parameter combinations) where the expected information gain — measured by relative entropy and Bayesian model evidence — is highest, increasing the surrogate’s accuracy precisely in the regions of parameter space that matter most for the calibration.

Where to go next