Neural Surrogate HMC: On Using Neural Likelihoods for Hamiltonian Monte Carlo in Simulation-Based Inference

Data and results used in the paper "Neural Surrogate HMC: On Using Neural Likelihoods for Hamiltonian Monte Carlo in Simulation-Based Inference" by Wolniewicz et al. The code used to work with this data can be found on GitHub: https://github.com/hawaii-ai/GalacticCosmicRays. ams_pamela_observations.zip: .dat files with the observations from PAMELA and AMS-02 used for generating posterior samples. nn_models.tar.xz: all the NN .keras files used to produce results (these are used to create Figure 3 in the paper). posterior_predictive_test.zip: .dat files with flux values used to perform the posterior predictive test in the paper. train_test_data.zip: .h5 files that contain the train and test data used to train and test the NN models (and perform the posterior predictive test). logprobs_d1_b0_init1_hmc1_h5.zip: log probability .h5 files from sampling for the two intervals shown in the paper in Figures 11 and 14 (AMS-02 interval #48, and PAMELA interval #107). predictions_d1_b0_init1_hmc1_h5.zip: prediction .h5 files from sampling for the two intervals shown in the paper in Figures 11 and 14 (AMS-02 interval #48, and PAMELA interval #107). posterior_predictive_samples_h5.zip: all posterior samples for various NN models and bootstrapped training datasets across a subset of 100 held-out simulation runs (these are used to create Figures 8 and 9 in the paper), stored in .h5 format. all_sample_files_h5.zip: all PAMELA and AMS-02 posterior samples for various NN models, HMC initializations, and bootstrapped training datasets stored in .h5 format. Sample files with folder name "d1_b0_*" are used to create Figures 10, 11, and 14 in the paper. Sample files with folder name "*_b0_*" are used to create Figures 4, 5, 6, 7, 12, and 13 in the paper. Naming convention of sample and model files: There are two kinds of dataset sampling: b1 for bootstrap-sampled training data, and b0 for non-bootstrapped. b0 is only used for training on the entire training set. There are 5 random HMC re-initializations: hmc1, hmc2, hmc3, hmc4, and hmc5 There are 5 bootstrap sampled datasets: d1, d2, d3, d4, and d5 There are 5 NN model re-initializations: init1, init2, init3, init4, and init5 Training data is of size: 0.0001, 0.001, 0.01, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0 This dataset contains all the train and test data, NN model files, posterior predictive test observation and posterior samples, AMS-02 and PAMELA observation and posterior samples (non-bootstrap sampled datasets), the logprobs and predictions needed to recreate Figures 10, 11, and 14, and all the AMS-02 and PAMELA posterior samples for changing HMC initial states, NLE initializations, and bootstrap-sampled training datasets, which are needed to recreate Figures 4, 5, 6, 7, 12, and 13 from the paper.