Home

Calibrate-Emulate-Sample

Purpose

Construct a registered (?) package for black-box uncertainty quantification of parameters in noisy, expensive and non-differentiable models.

Short term

• Refactor the code with better data structures as detailed below
• Separate out Calibrate into EnsembleKalmanProcesses.jl and have CalibrateEmulateSample.jl depend upon this
• Explore an end-to-end example

Longer term

• Working with user model instability. (We could also do nothing here, I'm not a fan of modifying priors, but perhaps it needs to be done)
• Public Availability?
• Explore examples from CliMA users to aid development,
• Build example use-case library from CliMA applications?)
• Documentation goals (i recall this good talk in particular that Simon referenced a while back on Slack. We could get some flavour from here perhaps to break down the task - or learn from CliMA experience people have )

Current Points of action (Assignee), 27 Jan

1. Working Example: Investigate a simple CliMA-based example (Ignacio/Mike)
2. Refactor: Move Calibration to EnsembleKalmanProcesses.jl (-)
3. Refactor:Interface with EnsembleKalmanProcesses.jl (-)

5 Latest features / Developments (writing a good PR)

1. (PR #104) Adds DataContainers for consistent dimensions, tagline "data are columns"
2. (PR #101) Adds example for Lorenz96 model, learning periodic forcing function
3. (PR #100) Adds EKP examples for Loss minimization
4. (PR #89, #94) Adds ParameterDistributions to deal with priors and posteriors

Functionality

Overall Code structure

Mind-map form of project.

Data Structures

Here we include the data structures we use in the project

ParameterDistribution

ParameterDistribution(...)

Module contains the additional functions

set_distribution()
get_distribution()
sample_distribution()
transform_constrained_to_unconstrained()
transform_unconstrained_to_constrained()
get_mean()
get_cov()
batch
...

ProcessedData Storage

DataContainer(...)
PairedDataContainer(...)

Module contains the additional functions

set_data()
get_data()
get_inputs() (paired data only)
get_outputs() (paired data only)
size()

Calibrate: Interface

Here we include the interface with the EnsembleKalmanProcesses.jl module.

EnsembleKalmanProcess(...)

We extract input-output pairs PairedDataContainer from this object using get_training_points(...)

Workflow 1

GModel(...)

Workflow 2

ModelInterface(...)

Emulate

GaussianProcessEmulator(...)

Sample

MarkovChainMonteCarlo(...)

Visualization:

This will be performed through the vizCES.jl module

Docs(?)