MULTIF Run Database

This git repository contains a number of input ouput pairs from MULTIF.

File Format

Input and outputs to MULTIF are encoded in files named *.input and *.output respectively where each row corresponds to one run. Output files also specify the fidelity level (0-13) and version in the format *level6_v13.output.

The input contains 175 variables corresponding to the 3D parameterization of the nozzle as specified in general-3d.cfg. The output contains either 71 or 72 variables depending on the version, where the last column is the log10 of the residual decrease in SU2.

Levels

MULTIF supports a variety of fidelity levels involving different models (1D, 2D, 3D) as well as different mesh refinements. Here we use the table of levels specified below and specified in general-3d.cfg:

NONIDEALNOZZLE, 1e-8, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
NONIDEALNOZZLE, 1e-8, AEROTHERMOSTRUCTURAL, NONLINEAR, 0.5
EULER, 2D, COARSE, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
EULER, 2D, MEDIUM, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
EULER, 2D, FINE, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
EULER, 3D, COARSE, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
EULER, 3D, MEDIUM, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
EULER, 3D, FINE, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
RANS, 2D, COARSE, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
RANS, 2D, MEDIUM, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
RANS, 2D, FINE, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
RANS, 3D, COARSE, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
RANS, 3D, MEDIUM, AEROTHERMOSTRUCTURAL, LINEAR, 0.5
RANS, 3D, FINE, AEROTHERMOSTRUCTURAL, LINEAR, 0.5

Versions

MULTIF is under active development and the output is known the vary between versions. Here, version numbers correspond to the docker container builds of MULTIF featured on docker hub.

Campaigns

Input sets have been generated in a number of campaigns, described below. Numbers denote additional attemps using the same approach

baseline: baseline design plus random samples in the random variables
design: designs that approximately solve the design under uncertainty problem; similarly fixed design plus random samples of random variables.
rand: randomly sampled points on the measure associated with the domain
sweep: parameter sweeps starting at a random point and pointed in a random direction; each sweep contains 100 points.
stretch: evenly spaced samples along identified ridge directions and randomly sampled in the orthogonal complement. Typically 10 points per ridge and 5 samples at each point. Letters denote that the ridge directions were determined using rand and the preceeding stretches; for example, stretch2c_l0_v13 samples ridges built from level 0 version 13 data in rand2, stretch2a_l0_v13, and stretch2b_l0_v13.
uniform: samples the domain using Mitchell's best candidate to attempt to evenly sample the 175 dimensional input space.
extend: a greedy approach that attempts to sample each of the ridge directions with an equal number of points using Mitchell's best candidate. Similar to stretch, letters denote samples built by refining the domain; e.g., extend1c_l0_v13 was built using level 0 version 13 data for uniform1, extend1a_l0_v13, and extend1b_l0_v13.

Computational Effort

In addition to the input output mappings, I've also recorded the total run time for each simulation I've performed. In this directory is a file named total_runtime.txt that contains a record of every simulation run since November 2017. Each row describes a single run of MULTI-F with each column encoding