This thesis presents the development of a software framework, VampireMan, designed to automate the generation of diverse and reproducible data sets for training surrogate machine learning models in groundwater flow simulations with heat pumps. Groundwater flow and heat transport simulations are important tools for applications like geothermal energy systems, requiring extensive high-quality data sets for accurate predictive modeling. Surrogate machine learning models have emerged as efficient alternatives to computationally expensive numerical simulations, enabling rapid predictions of subsurface temperature fields. The success of these models relies on the availability of diverse and reliable data sets, encompassing variations in physical and operational parameters. However, manual and semi-automated data set creation approaches are limited in scalability and prone to errors.

VampireMan addresses this challenge by automating the entire data generation workflow: systematically varying simulation parameters, generating simulation input files, running simulations with PFLOTRAN, and visualizing outputs. The framework adheres to Research Software Engineering (RSE) and FAIR4RS (Findable, Accessible, Interoperable, and Reusable for Research Software) principles, ensuring reproducibility, scalability, and extensibility.

Key features include reproducible data set generation using different parameter variation modes (fixed, constant, and spatial), modular pipeline stages, and integration with PFLOTRAN. VampireMan's effectiveness is demonstrated through preconfigured examples that showcase parameter variations and simulation workflows. By enabling efficient and reproducibility data set generation, VampireMan can help advancing machine learning applications in environmental engineering, facilitating resource-efficient and real-time decision-making for subsurface energy systems.