Thesis Supplement: A Deep Learning Framework for Modeling Input Uncertainty in Neutron Reflectometry Predictions

This record contains the reproducibility materials for a Master’s thesis on uncertainty-aware amortized inversion in neutron reflectometry. It includes (i) all plots referenced or used in the thesis and (ii) the raw evaluation outputs used to compute reported metrics and tables. Included files thesis_plots.zip: all evaluation and diagnostic plots (primarily PDF; some PNG/PGF), including constraint-based MAPE histograms, per-parameter MAPE breakdown plots, and selected uncertainty diagnostics (coverage/calibration) where applicable. Plots are organized by method and evaluation regime. batch_jsons.zip: raw BatchInferencePipeline outputs for all evaluated methods and regimes. Each run folder contains batch_results.json (per-experiment predictions, ground truth, priors/constraints, and derived errors), batch_summary_1layer.json (pre-aggregated statistics), and protocol exclusions (outlier_experiments.json, and failed_experiments.json where present). Evaluation regimes encodedMethods are evaluated across a grid of regimes: prior widths of 5% / 30% / 99% of the global constraint span (“constraint-based priors”), all curves vs PROMINENT (feature-based filtering), no SLD fixing / backing SLD fixed / all SLD fixed. Run folder names encode these settings and timestamps to allow direct mapping between plots and raw JSON outputs. How to useTo reproduce thesis figures, select the matching method and regime folder inside thesis_plots.zip. To reproduce reported aggregate numbers (mean/median MAPE and per-parameter breakdowns), use the corresponding run folder in batch_jsons.zip, relying on batch_summary_1layer.json or recomputing from batch_results.json while excluding entries listed in outlier_experiments.json and failed_experiments.json (if present).