How simple should simple models be?
Abstract
Process-based models have been extensively used to glean mechanistic understanding of natural processes. There is an ongoing debate about whether to use simple versus complex models, both in terms of spatial discretization (distributed versus lumped) and the level of process details (e.g., flow representation, reaction stoichiometry, thermodynamics, and kinetics). How simple should simple models be? Or how complex should complex models be? It is often a fine line to walk between simple and complex models. This talk will share a few examples and lessons learned in our group when using watershed reactive transport models and maneuvering along this fine line. Models are not reality and the best models are not the ones that can do everything. The level of model simplicity and complexity should be the most fit for answering proposed questions and for revealing key, bottleneck processes that lead to patterns and dynamics exhibited in data. Toward achieving this goal, we generally follow a simple but not simplistic, or fit-for-purpose approach. We start building a model at the conceptual level (with hypothesis) from simple representation, and gradually add complexity when necessary, until a satisfactory balance is achieved between minimizing complexity and maximizing fidelity to the data. Because additional details always introduce the need for more data, the level of complexity should also be consistent with available data: whether existing measurements (water, water chemistry) provide sufficient constraints to differentiate the influence of individual processes at the relevant spatial scales. We underscore the importance of using models as thinking tools and the importance of data-model confrontation. Failed models that do not reproduce data often offer key insights on processes that control system dynamics such that failure is necessary. When having a choice, we advocate the use of simple models, as smaller numbers of processes and parameters in simple models offer transparency of model structure, clear understanding of causal effects, and better accessibility for education, broader user groups, and research questions.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H32B..05L