Data and Algorithms for "Prediction of Intensity Variations Associated with Emerging Active Regions using Helioseismic Power Maps and Machine Learning" Manuscript at ApJ

Description This repository contains the code and sample data associated with the paper accepted in The Astrophysical Journal (ApJ):"Prediction of Intensity Variations Associated with Emerging Active Regions using Helioseismic Power Maps and Machine Learning" by Spiridon Kasapis et al. 2025 In this study, we present a machine learning framework based on Recurrent Neural Networks (RNNs) with Long Short-Term Memory (LSTM) architecture to forecast the early emergence of solar active regions (ARs) before they are visible in HMI continuum intensity maps and begin forming sunspots. Our goal is to provide interpretable predictions for the timing and location of AR emergence using subsurface helioseismic signals and magnetic field evolution. The model was trained on a curated dataset of 40 emergence events and tested on 5 held-out emerging ARs. Input features include: Time series of mean acoustic power maps (derived from SDO/HMI Dopplergrams across four frequency bands), Mean unsigned magnetic flux (from HMI magnetograms), Corresponding continuum intensity (from HMI continuum intensity maps). The model predicts when and where a decrease in continuum intensity will occur within a 12-hour forecast window, serving as a proxy for the beginning of the AR emergence. Our results show that acoustic power maps contain precursor signatures of active region emergence. The LSTM model is able to anticipate these intensity dips several hours in advance—at times when the surface magnetic flux is still only 4–9.6% of its eventual maximum—thus demonstrating the potential of machine learning and helioseismology in the early detection of ARs. Contents 📁 AREmergCode.zip Contains all Python source code and model output figures necessary to train and evaluate the ML model: train_w_stats.py — Main training script to build and train the LSTM model. functions.py — Utility functions used in both training and evaluation pipelines. evalXXXX.py — Evaluation scripts for the 5 test active regions (AR11698, AR11726, AR13165, AR13179, AR13183). These ARs were not included during training. *.pdf — Visual results showing prediction performance for each evaluation AR. t12_r4_i110_n3_h64_e1000_l0.01.pth — Trained model weights. run_validations.pbs — Batch script for automated evaluation of all test ARs (for use on HPC environments). all_training_results.txt — Training loss and evaluation metrics over all epochs. 📁 demo_data.zip Includes a reduced dataset of 50 active regions used in training and validation. Each subfolder corresponds to an individual active region and contains: mean_int<AR>_flat.npz: Continuum intensity time series data mean_mag<AR>_flat.npz: Unsigned magnetic flux time series mean_pmdop<AR>_flat.npz: Acoustic power maps derived from Dopplergrams across four frequency bands Each .npz file contains a 9×9 spatial tiling of time series covering the AR’s emergence, with 240 time steps per tile. The data were extracted and preprocessed from the SDO/HMI instrument as part of the SOLARED dataset. How to Use Unzip both folders. Use train_w_stats.py to train the LSTM model from scratch, or load t12_r4_i110_n3_h64_e1000_l0.01.pth to evaluate directly. Run any of the evalXXXX.py scripts to evaluate the model on held-out ARs. Customize input/output paths in the scripts as needed based on your local folder structure.