top of page
Liquid Bubbles


We are currently working on research projects related to aquatic ecosystem modeling, physical limnology, aquatic metabolism, and data science.

Ensemble Modeling

Skærmbillede 2024-07-03 230237.png
Skærmbillede 2024-07-03 225500.png

Various one-dimensional, vertical, process-based aquatic ecosystem models exist that couple hydrodynamic with water quality models to replicate ecosystem dynamics. Although computationally similar, these models differ regarding their specific assumptions and mathematical formulations. In the last decades, several studies have pointed out that a systematic comparison of this plethora of aquatic ecosystem models is missing, as well as an evaluation of their accuracy, quantification of uncertainty and conceptual framework. Building on our past success of creating a common software framework for hydrodynamic ensemble modeling of lakes in R – the LakeEnsemblR package – we are working on an updated ensemble framework to run aquatic ecosystem models, assess their predictions on multiple ecosystem metrics and improve overall ensemble model confidence. This project is supported by an AUFF Recruiting Grant.

Modular Compositional Learning

Skærmbillede 2024-07-16 134823.png

The emerging scientific paradigm of “Knowledge-Guided Machine Learning” (KGML) focuses on creating hybrid models that combine process-based principles (scientific theory) with data-driven deep learning models. Modular compositional learning (MCL) is a novel design choice for the development of hybrid KGML models in which the overall model is decomposed into modular sub-components that can be process-based models and/or deep learning models. As part of the Eco-KGML team, I am working on developing and advancing the MCL methodology into a robust and transparent tool for scientific modeling. First results in Ladwig et al. (2024) highlighted how MCL can improve the performance of a 1D hydrodynamic model in replicating water temperature dynamics. MCL advances current hybrid KGML model designs, which generally focus on developing a single deep learning model that incorporates process knowledge. Here, hybrid MCL models would have a design similar to the modularized structure of process-based models, and individual sub-parts can either be process-based or deep learning models. Like a jigsaw puzzle, the MCL approach allows the modeler to balance between framework design (how much prior knowledge is inserted into the model formulation) and chances for the discovery of relationships through machine learning.

bottom of page