Can River Chemistry Reveal Catchment Lithology?

Learning module
Python
Deep learning
Hydrogeochemistry
A guided, 40–45-minute PyTorch module by Tao Wen comparing linear and neural-network classifiers using river chemistry.

About the module

Developed by Tao Wen for the IAMG 2026 short course, this guided exercise uses river chemistry to distinguish carbonate-dominated catchments from those dominated by crystalline silicate rocks. Learners compare a linear classifier with a small neural network in PyTorch.

  • Audience: Geoscientists with basic Python skills; no prior PyTorch, deep-learning, or hydrogeochemistry experience required.
  • Estimated time: 40–45 minutes.
  • Tools: Google Colab on a free CPU runtime; no GPU or manual data upload required. Local Jupyter instructions are also available.
  • Developer: Tao Wen, Syracuse University.
  • Teaching data: A curated GLORICH subset with nine chemistry inputs for 2,986 stations.

Open the guided notebook in Colab

View the module repository · Read the instructor guide

Learning goals

  • Define predictors and target classes from a geoscience question.
  • Split data before fitting transformations, then train and compare models.
  • Evaluate class imbalance and interpret classification metrics.
  • Experiment with activation functions and discuss limits of geographic transfer.

A random station split tests performance in similar sampled settings. It does not establish transfer to unfamiliar regions or prove that lithology alone controls river chemistry.

Attribution and reuse

The repository citation credits Tao Wen and identifies version 1.0.0, released July 11, 2026. Cite the module and its source dataset when reusing the materials:

  • Wen, T. (2026). Can River Chemistry Reveal Catchment Lithology? (Version 1.0.0) [Software]. GitHub.
  • Hartmann, J., Lauerwald, R., and Moosdorf, N. (2019). GLORICH — Global river chemistry database [Dataset]. PANGAEA.

Original code, notebooks, and documentation are MIT-licensed unless otherwise stated. GLORICH-derived data have separate CC BY-NC-SA 4.0 terms and remain subject to the repository’s data-release conditions. Review the current license and data-use notice before redistribution or fee-supported teaching.