A worked example of an iterative, scripted approach to stochastic model development and deployment in a highly contentious decision-support setting.

A scripted development and deployment approach was used for developing the next-generation groundwater flow and land-surface subsidence model of the region surrounding Houston, Texas, USA.  The area has historically experienced substantial land subsidence resulting from groundwater use. Python scripts leveraging the FloPy and PyEMU packages were written to build and run the MODFLOW 6 model, perform very-high-dimensional parameter estimation and uncertainty analysis using PEST++, and process results. Automating these processes allowed for fast and repeated iterations through all or part of the modeling workflow for purposes including: troubleshooting input errors, testing hypotheses about the hydrologic system characteristics, evaluating the influences of structural model assumptions, and experimenting with different and increasingly complex formulations of the prior parameter distribution and likelihood functions in Bayesian sense. Automated generation and storage of processed output allowed easy comparison between iterations of the modeling workflow, and Git version control software provided a self-documented model repository with full-featured “undo” for returning to previous states of the workflow and investigating outcomes. The modeling team convened regularly (monthly to twice-weekly) to review results of the latest iteration and decide the next course of action. Model performance was improved steadily and incrementally by focusing on one new feature or problem per workflow iteration until modeling goals were met. This workflow style fostered a sense of predictability and confidence in the project outcome, a welcome departure from the “typical” numerical modeling process of panic and despair.