Storms wash water, soil and carbon from farmland into streams. A model of the watershed can describe this, but it has to be tuned to each stream, and for carbon it does poorly even then. This study lets a neural network learn from an untuned model and from the measurements together. The result matches a tuned model for streamflow and is clearly better for carbon.
At a glance
- Theme: Streamflow and organic carbon in storms
- Where: Tuckahoe Creek, a farming watershed of 221 km² by the Chesapeake Bay, United States
- Data: Daily streamflow and organic carbon at the outlet, 2014 to 2019
- Approach: A watershed model used untuned, with a network that follows the river and remembers the days before
- Lead author: Dr. Venkatesh Budamala
- Status: Published, code open

The basics
In a farming region most of the water and carbon that leave the land do so in a few storms. The organic carbon travels in two forms. Carbon in particles is washed off with eroded soil and peaks with the flood. Dissolved carbon seeps through the soil and arrives later. Both shape the quality of the water downstream, here in the Chesapeake Bay.
The challenge
A model built on the physics of a watershed has dozens of settings, and it has to be tuned against local measurements before it can be trusted. That takes time and has to be done again for every watershed. A neural network needs no such model, but it needs long records to learn from, and records of carbon in streams are short.
The idea

The study uses the model as it comes. SWAT-C, the carbon version of a widely used watershed model, is run without tuning. For each of the 19 parts of the watershed it gives the water in the soil, the flow from groundwater, the soil washed off and the carbon in the soil, day by day. These become the inputs of a network with two parts. One follows the river from each part of the watershed to the next. The other looks back over the days before.
The network first predicts the flow of the stream. It then predicts the carbon in particles and the dissolved carbon, with the predicted flow as an extra input, because the flow is what carries them. It learned from the years 2014 to 2017 and was tested on 2019, a year it had not seen.
What was found

For streamflow the new approach reaches an efficiency of 0.82, the same as the model after tuning, where the untuned model alone reaches 0.23. For carbon it is well ahead: 0.68 against 0.04 for carbon in particles, and 0.74 against 0.43 for dissolved carbon. A simpler network combined with the model does worse on all three. The error in the volume of water over the year falls from 40 per cent for the untuned model to 7 per cent.
In the storms

The largest storm of 2019 came on 21 to 24 March, with about 40 mm of rain and a peak flow of 26.8 m³ per second. The untuned model overshot that peak by a wide margin, and the peaks of carbon with it. The new approach stayed close to what was measured. The timing of dissolved carbon changes too. In the model alone, dissolved carbon rises and falls with the flood. In the new approach it lags behind, as carbon that travels through the soil is known to do.
What the network needs the model for
The same networks were also trained without what the model supplies. They failed, most clearly for carbon, where their skill was close to zero or below. A few years of record are too few to learn a watershed from. The untuned model is far from the measurements, but it knows where the water goes, and that is what the network builds on.
What the result does not say
The method was tested in one watershed, flat and dominated by farming, and on daily values, with a single year held back for testing. How well it carries over to other climates, sizes of watershed and kinds of land is still to be shown. It also rests on the model: where the model leaves out the process that matters, the network has less to build on.
Why it matters
Tuning is the slow part of modelling a watershed. A method that reaches the same skill without it, and does better for carbon, can be carried more easily to the many watersheds where measurements are few. Knowing how much carbon storms carry off the land matters for the water quality of estuaries and for the carbon balance of farmland.
The paper
Budamala, V., Choi, E., Kim, M., Lee, S., and Kim, H. (2026). Hierarchical prediction of streamflow and organic carbon at event scale using a physics-guided AI hybrid framework. Journal of Hydrology, 679, 136248. The work was done with colleagues at Korea University. The carbon measurements were provided by the United States Department of Agriculture.












