Lightweight and Interpretable Solar Flare Forecasting with Mamba: Feature and Temporal Attribution via Integrated Gradients

Lightweight and Interpretable Solar Flare Forecasting with Mamba: Feature and Temporal Attribution via Integrated Gradients The repository includes the datasets and experimental code and result used in our research, along with descriptions of these files. 1.Directory Introduction 1.1DATA The zip file named " DATA.zip" provides the datasets in the paper. 1. In this repository, the file named "10CV" contains ten groups of cross-validation datasets, corresponding to Section 2.1, where AR represents the NOAA number of the active region, time indicates the sample collection time, level denotes the flare level of the active region, and the remaining columns correspond to feature names. 2. In this repository, the "sharp_data_ten_feature.csv" corresponds to the comparative dataset in Section 2.2. Besides, "sharp_data_ten_feature.csv" also corresponds to the comparative dataset matched with the prediction results from LLMFalreNet. "ccmc_waitcompare.csv" corresponds to the comparative dataset matched with the prediction results from NASA/CCMC. "sc_waitcompare.csv" corresponds to the comparative dataset matched with the prediction results from SolarFlareNet. 1.2 Result The zip file named "Result.zip" provides the result of the manuscript. For details, see "https://huggingface.co/Qduck/MambaFlare/tree/main" to download the "Result.zip". 1. In the "Train_material" dictionary, the "model" folder contains all models.pt from the manuscript, named after the model names. 2. The files "Table1.xlsx" to "Table8.xlsx" in the "Table" folder contain the table results from the manuscript. 3. The files "figure1.eps" to "figure20.eps" in the "figure" folder contain the figure results from the manuscript. 2、Code Usage Instructions The zip file named "Code.zip" provides the code in the manuscript. 1. In the folder named "Compare_10CV" i. The file named "model", in each subfolder, named after the model, contains a "***.py" file. The "***.py" file holds the model's structure code. ii. The file named "run_main.py" contains the training code for all models. You need to use "run_main.sh" to run "run_main.py". For different models, you only need to modify the model_name parameter in "run_main.sh". For example, when training Mamba, change model_name to models_mamba, and so on. Enter ./run_main.sh in the terminal to run it. In addition, after saving the model, the code in "run_main.py" directly tests the model on the testing sets and calculates the corresponding metrics, i.e., it includes the testing functionality, which can produce the results in Table 5. "run_main.py" also includes the functionality for calculating model computational complexity; the results of Table 4 are output in the logs when each model is run. iii. The subfolder named "loss_draw" contains the CSV files of loss values saved during training and validation for each model. These files are named after the respective models, such as "models_***_train_loss.csv". Running "loss_curve_draw.py" will plot the training and validation loss curves for different models, producing Figure 3 and Figure 4. 2. In the folder "IG_interpretability", enter the following command in the terminal: cd /mnt/e/A_paper/VIM/models (place the data in the same-level directory) && python ig_analysis_10fold.py--model_base_dir./model_output/Mamba --data_dir ./data --n_samples -1 --steps 120 to run "ig_analysis_10fold.py", which will generate Figures 5 to 16. Then enter python r_value_time_step_global_ig.py in the terminal to run "r_value_time_step_global_ig.py", which will generate Figure 17. 3. In the folder "Compare_comparative dataset", running "2main.py" calculates the TSS values obtained from testing the Mamba model on the Comparative dataset under different thresholds. Running "2main_other.py" plots the TSS values obtained from testing the model on datasets matched with institutions such as LLMFlareNet, NASA/CCMC, and SolarFlareNet under different thresholds. Finally, running "draw_daily_mode_comparsion" plots the TSS values of the model under various thresholds as curves, as shown in Figures 18 to 20, and records the maximum TSS values of the model under their respective thresholds in the figures into Tables 6 to 8.