MINAS — Machine-learning INtegrated Analysis with photometric Astronomical Surveys

MINAS is a Python package for the complete Machine Learning workflow applied to photometric astronomical surveys. It integrates all stages — from preprocessing to final model application — in a single, modular interface.

Fun fact: MINAS is also the name of a Brazilian state (Minas Gerais), the home state of Icaro Meidem, the package creator. As a proud mineiro, the name represents both the astronomical focus and personal heritage.

---

Installation

pip install minas

---

Quick Start — Full ML Workflow

import minas as mg

# 1. Load catalog
catalog = mg.read_csv('my_catalog.csv')

# 2. Assemble feature DataFrame (magnitudes + pairwise colors)
work_df = mg.preprocess.assemble_work_df(
    df=catalog,
    filters=mg.FILTERS['JPLUS'],
    correction_pairs=dict(zip(mg.FILTERS['JPLUS'], mg.CORRECTIONS['JPLUS'])),
    add_colors=True,
)

# 3. Select most important features
features, df_importance = mg.evaluation.get_important_features(
    X=work_df,
    y=catalog['Teff'],
    n_features_to_save=20,
)
work_df = work_df[features]

# 4. Tune hyperparameters
param_dist = {
    'selectkbest__k'                         : [10, 15, 20],
    'randomforestregressor__n_estimators'    : [100, 300, 500],
    'randomforestregressor__min_samples_leaf': [1, 5, 10],
    'randomforestregressor__max_features'    : ['sqrt', 'log2'],
    'randomforestregressor__bootstrap'       : [True, False],
}
best_pipeline, search = mg.hyperparameter_search(
    X=work_df,
    Y=catalog['Teff'],
    model_type='RF',
    param_dist=param_dist,
    tuning_id='teff_rf',
    n_iter=30,
    save_dir='pipelines/',
)

# 5. Apply model with Monte Carlo error propagation
predictor = mg.models.Predictor(
    id_col='ID',
    mag_cols=mg.FILTERS['JPLUS'],
    err_cols=mg.ERRORS['JPLUS'],
    dist_col=None,
    correction_pairs=dict(zip(mg.FILTERS['JPLUS'], mg.CORRECTIONS['JPLUS'])),
    models={'Teff': best_pipeline},
    mc_reps=100,
    batch_partitions=10,
)
predictor.predict_parameters((catalog, 'results/teff_predictions.csv', ['ID'], 'w', True))

---

Bolometric Correction

MINAS includes pre-trained models for bolometric correction (BC) based on Jordi et al. (2010), trained on Gaia-observed stars using Teff, log g, and [Fe/H].

Model Performance

Model	R²	MAD	Std Deviation
XGBoost	0.9983	0.0062 mag	0.0430 mag
Random Forest	0.9970	0.0067 mag	0.0573 mag

Figure: Performance of XGBoost (left) and Random Forest (right) for bolometric correction prediction.

Usage

import minas as mg

df = mg.bolometric.apply_bc(
    data='catalog.csv',
    teff_col='Teff',
    logg_col='logg',
    feh_col='[M/H]',
    model_type='XGB',        # 'XGB' or 'RF'
    sigma_multiplier=3.0,    # uncertainty = multiplier x STD
    output_file='catalog_bc.csv',
)

print(df[['Teff', 'BC_pred', 'err_BC_pred']].head())

Reference

Jordi, C. et al. (2010). Gaia broad band photometry. A&A 523, A48. DOI: 10.1051/0004-6361/200913234

---

Supported Surveys and Filters

MINAS provides built-in filter definitions for the following photometric surveys. All filter lists are accessible via mg.FILTERS, mg.ERRORS, and mg.CORRECTIONS.

Survey	Filters	mg.FILTERS key
J-PLUS	uJAVA, J0378, J0395, J0410, J0430, gSDSS, J0515, rSDSS, J0660, iSDSS, J0861, zSDSS	'JPLUS'
S-PLUS	uJAVA, J0378, J0395, J0410, J0430, gSDSS, J0515, rSDSS, J0660, iSDSS, J0861, zSDSS	'SPLUS'
J-PAS	uJAVA + 56 narrow bands (J0378-J1007) + iSDSS	'JPAS'
WISE	W1, W2, J, H, K	'WISE'
GALEX	NUVmag	'GALEX'
Gaia	G, BP, RP	'GAIA'

import minas as mg

print(mg.FILTERS['JPLUS'])      # magnitude column names
print(mg.ERRORS['JPLUS'])       # photometric error column names
print(mg.CORRECTIONS['JPLUS'])  # extinction correction column names

---

Model Comparison — RF vs XGBoost

Feature	Random Forest	XGBoost
Pipeline steps	Imputer → SelectKBest → RF	SelectKBest → XGB
Missing value handling	Built-in (median imputation)	Must be handled externally
Training speed	Moderate	Fast
Typical accuracy	Good	Excellent
Model key	'RF-REG' / 'RF-CLA'	'XGB-REG' / 'XGB-CLA'
Saved format	.sav (joblib)	.json

import minas as mg

# Default models
rf_model  = mg.models.create_model('RF-REG')
xgb_model = mg.models.create_model('XGB-REG')

# With tuned hyperparameters
hp = (0.8, 0.05, 6, 500, 0.8, 0.1)  # colsample, lr, depth, n_est, subsample, gamma
xgb_tuned = mg.models.create_model('XGB-REG', hp_combination=hp)

---

Package Structure

minas/
├── preprocess/     magnitude correction, color creation, work DataFrame assembly
├── models/         ML pipeline factory (RF, XGB) and Monte Carlo predictor
├── tuning/         hyperparameter search with RandomizedSearchCV
├── evaluation/     metrics (MAD, R2), plots, feature importance
└── bolometric/     bolometric correction with pre-trained models

Key Functions

Function	Description
mg.preprocess.assemble_work_df()	Build feature DataFrame from magnitudes
mg.preprocess.correct_magnitudes()	Apply extinction corrections
mg.preprocess.calculate_abs_mag()	Convert apparent to absolute magnitudes
mg.models.create_model()	Create RF or XGBoost pipeline
mg.models.Predictor	Monte Carlo predictor with uncertainty estimation
mg.hyperparameter_search()	RandomizedSearchCV for RF or XGB
mg.evaluation.get_important_features()	Impurity-based feature importance (RF)
mg.evaluation.get_permutation_importance_rf()	Permutation importance (RF)
mg.evaluation.get_permutation_importance_xgb()	Permutation importance (XGB)
mg.evaluation.calculate_mad()	MAD per bin
mg.evaluation.plot_test_graphs()	Scatter + KDE error plot
mg.evaluation.plot_comparison_graph()	Bar chart comparison across models
mg.bolometric.apply_bc()	Apply pre-trained bolometric correction model

---

Examples

The examples/ folder contains complete Jupyter notebooks covering the full workflow:

Folder	Contents
data/	Catalog creation and preprocessing
tuning/	Hyperparameter search and feature importance
training/	Model training, evaluation, and visualization
apply/	Model application with Monte Carlo error propagation

---

Citation

If you use MINAS in your research, please cite:

@software{minas,
  author  = {Meidem, Icaro},
  title   = {{MINAS}: Machine-learning INtegrated Analysis with photometric Astronomical Surveys},
  year    = {2025},
  url     = {https://github.com/icaromeidem/minas},
}

Bolometric correction reference:

• Jordi, C. et al. (2010), A&A 523, A48 — doi:10.1051/0004-6361/200913234

---

License

MIT © Icaro Meidem