closure-challenge 0.3.0

Evaluation metrics for The Closure Challenge RANS benchmark

Closure Challenge

A benchmark challenge for evaluating machine learning-augmented RANS closure models. The main challenge page is https://github.com/rmcconke/closure-challenge-benchmark. This repo is for the evaluation source code.

Installation

pip install closure-challenge

Quick Start

You need to interpolate your data onto the evaluation points:

from closure_challenge import evaluation_points
xyz_eval = evaluation_points(case) #(1000,3) numpy array for each case

Then, there are 2 ways to get your score from the package:

Method 1: Direct scoring, using python

from closure_challenge import score, evaluation_points, case_names
from scipy.interpolate import griddata # interpolation method/library is up to you

your_grid_points = {...}   # Dict[case_name -> (N, 3) array of xyz coordinates]
your_model_results = {...} # Dict[case_name -> (N, 3) array of velocity vectors]

predictions = {}
for case in case_names():
    # your model predictions on your grid
    xyz_your_points = your_grid_points[case]
    U_your_points = your_model_results[case]  

    # evaluation_points returns shape (1000, 3) where columns are [x, y, z]
    xyz_eval = evaluation_points(case)       

    # again, interpolation method/library is up to you
    U_pred = griddata(xyz_your_points, U_your_points, xyz_eval, method='nearest') 
    
    predictions[case] = U_pred


final_score = score(predictions)

Method 2: Loading from CSV files

This is what will be used when updating the main challenge github with your submitted files.

You can pass in either:

1. The path to a folder containing your interpolated predictions on the test cases, which should be stored in the files:

• alpha_15_13929_4048.csv
• alpha_15_13929_2024.csv
• alpha_05_4071_4048.csv
• alpha_05_4071_2024.csv
• AR_1_Ret_360.csv
• AR_3_Ret_360.csv
• AR_14_Ret_180.csv
• NASA_2DWMH.csv

from closure_challenge import score_from_csv
final_score = score_from_csv("path/to/predictions/")

1. A dictionary containing paths to your interpolated predictions for each of the test cases.

from closure_challenge import score_from_csv
final_score = score_from_csv({
    "alpha_15_13929_4048": "path/to/your/alpha_15_13929_4048/predictions.csv",
    "alpha_15_13929_2024": "path/to/your/alpha_15_13929_2024/predictions.csv",
    "alpha_05_4071_4048": "path/to/your/alpha_05_4071_4048/predictions.csv",
    "alpha_05_4071_2024": "path/to/your/alpha_05_4071_2024/predictions.csv",
    "AR_1_Ret_360": "path/to/your/AR_1_Ret_360/predictions.csv",
    "AR_3_Ret_360": "path/to/your/AR_3_Ret_360/predictions.csv",
    "AR_14_Ret_180": "path/to/your/AR_14_Ret_180/predictions.csv",
    "NASA_2DWMH": "path/to/your/NASA_2DWMH/predictions.csv"
})

Scoring

The benchmark uses scaled Mean Absolute Error (MAE). Each case's error is normalized by the mean magnitude of the true velocity field, then summed across all cases.

Available functions

i.e., you can do from closure_challenge import:

• case_names() - List of test cases
• evaluation_points(case) - Get evaluation (x, y, z) coordinates for a case
• score(predictions) - Calculate final benchmark score (see Method 1 above)
• evaluate_by_case(predictions) - Get individual case scores returned (Method 1)
• score_from_csv(path) - Calculate final benchmark score from CSV files (see Method 2 above)
• evaluate_from_csv_by_case(path) - Get individual case scores returned (Method 2)

Test Cases

The current scoring set includes these cases, which you can get by

• alpha_15_13929_4048
• alpha_15_13929_2024
• alpha_05_4071_4048
• alpha_05_4071_2024
• AR_1_Ret_360
• AR_3_Ret_360
• AR_14_Ret_180
• NASA_2DWMH

You can get these names using

from closure_challenge import case_names
print(case_names())

Reminder: DO NOT train or validate on any of these cases. In other words, you cannot use them during training in any way. This violates the challenge rules.