MODELS HERE, MODELS THERE; MODELS, MODELS EVERYWHERE OR: HOW I LEARNED TO STOP WORRYING AND LOVE BEING WRONG

Scarcely anything we do as environmental and engineering geoscientists can be accomplished without abstraction to reduce processes, data, and space-time relationships to more tractable forms. We do that using models that, as the statistician George Box wrote, are in all cases wrong but in some cases useful. This presentation will describe the kinds of models we use, ways they can be wrong, and things we can do to make them useful.

Models are wrong in that they do not perfectly replicate reality. Useful models are complex enough to allow insights or make predictions that would not be obvious without them. Useful models are also simple enough to easily parameterize and implement, especially to support policy or planning decisions in which comparison of alternatives or understanding departures from baseline cases is important. Because useful models will always be somewhat wrong, we need to love being wrong in the right way. We also need to work with, not avoid, the uncertainty that comes with being wrong by understanding its bounds and behaviors. The most straightforward way to do that is through probabilistic approaches.

Models useful to environmental and engineering geologists can be broadly classified as either empirical models based solely on field, lab, or office observations, or rational models based on underlying physical, chemical, or biological principles. The same models can be simultaneously classified as deterministic if they are based on inviolable cause-effect relationships or probabilistic if they are based on uncertain cause-effect relationships. Digital elevation models, geologic maps, and engineering geologic ground models are typically empirical deterministic models. Machine learning and artificial intelligence models for tasks such as estimating landslide susceptibility or mapping surface deposits are typically empirical probabilistic models because they are based upon observations and convey some degree of uncertainty. A traditional factor-of-safety slope stability calculation is a rational deterministic model whereas one incorporating parameter uncertainty is a rational probabilistic model.

Selection of the best model for a particular purpose requires consideration of project objectives, fitness for purpose, data and computational resource availability, and project timelines.