Topics & readings

A model is a description — an ordered set of descriptors built from a finite but extensible alphabet. So are data, and so are metrics. Descriptions are encodings that must be decoded, they reflect representational choices, and effective description is compression. Large language models are generative models of exactly such languages, and agents are built from them. That chain of reasoning sets up the rest of the semester.

The long-running debate over what theory buys us once data-driven models outperform our equations — framed here through information theory, hypothesis testing, and the difference between reducing uncertainty and gaining information.

Model development is a hierarchy of decisions — physical constraints, architecture hypotheses, process parameterizations, property-to-parameter mappings. Naming those decisions explicitly is what makes it possible to ask which of them could be “agentified”. The HWRS 428/528 lecture modules are repeated here from Topic 00 because the syllabus asks you to re-read them with that question in mind.

Calibration as a reasoning task rather than a search task. The new work asks whether an LLM can play the role of the expert hydrologist in multi-step, constraint-aware parameter specification — which is exactly what the classic MACS / manual-automatic hybrid papers were trying to encode by hand.

Can a machine write the equation? Symbolic regression and, more recently, text-generating AI have both been used to distill parameter transfer functions from physio-geographical properties — two very different routes to the same hypothesis-generation problem.

One level up from the equations: what structure should the model have at all? Structural adequacy, hydrologic consistency, knowledge graphs of process models, and distillation as a route to generating and testing architectural hypotheses.

Metrics are descriptions too — of what, exactly? This topic separates training metrics from evaluation/diagnostic metrics, poses the model–metric dual problem, and treats model development as a generative-adversarial process between the two.

A discussion session. The reading list was still open at the time this archive was assembled — the folders exist but are empty.

No readings have been posted for this topic yet.

Background library for the group project: the mass-conserving perceptron line of work, interpretable LSTMs, Bayesian structure estimation, and the early neural-network rainfall–runoff papers that started it.