An end-to-end system that classifies objects from the NASA Exoplanet Archive into three classes: CONFIRMED, CANDIDATE and FALSE_POSITIVE. It was developed in response to a 2025 NASA exoplanet challenge. A Python script prepares the archive tables, a fully connected network written in C++ without machine-learning frameworks does the classification, and a local web interface lets a user try it on manual input or a CSV file.
On a class-balanced set of 4500 objects, evaluated on a held-out 20% split over five runs, the network reaches 71.1% ± 1.1% accuracy (chance level 33.3%). The weakest class is CANDIDATE. The full method, results and limitations are published as a technical note on Zenodo.
The Python script nasa_exoplanet_pipeline.py reads the archive CSV files, maps their column names to common features, normalizes them and writes a C++-ready training file. The C++ program trains the network and stores the weights in a text file; a separate C++ predictor loads them. A small Python web server calls the predictor and renders the results. The two illustrations below show the data path and the software stack. The learning curve and the confusion matrix in the gallery above come from the real evaluation runs: the curve is averaged over five runs and the matrix is summed over them.
Send a direct message about this project and it will be delivered to my email.