SFDT-DTVRA: Reproducibility Code for Closed-Loop Digital-Twin Validation of Federated

Abstract

This software package contains the code used for the simulation, analysis, and reproducibility of a study on closed-loop digital-twin validation of federated learning updates for methane control in underground mining. The package includes the plant–twin simulator, federated-learning routines, calibration scripts, intervention-budget analyses, attack calibration and backdoor diagnostics, candidate-update generation and scoring, the fresh-seed confirmatory analysis, twin-mismatch sensitivity experiments, exploratory Krum-based comparisons, FLTrust-style diagnostics, serious-harm analyses, and read-only manuscript audits. The evaluation separates the simulated plant used for operational outcome assessment from the imperfect digital twin used for candidate scoring. The repository also documents the distinction between direct plant evaluation and curve-interpolated budget matching used in different experimental stages. Raw methane measurements are not redistributed in this archive. The study uses the public dataset “Methane” by Sikora and Wróbel, available from Mendeley Data, Version 1, DOI: 10.17632/yd7vw4c5mk.1. Instructions for obtaining and preparing the data are provided in DATA.md. The archive includes reproducibility documentation, script-to-manuscript mapping, provenance notes, execution order, software requirements, and SHA-256 checksums. Historical scripts are retained where necessary to document the progression from early exploratory analyses to the final frozen analysis.