Machine learning helps calculate how much water leaves the soil during heatwaves

A hybrid modeling technique separates plant transpiration from soil evaporation across diverse landscapes, offering clearer projections of ecosystem stress during severe heat waves.

Every day, vast volumes of water leave the ground and rise invisibly into the air. Part of this moisture evaporates straight from exposed soil and wet surfaces, driven by the direct warmth of sunlight.¹ Another large fraction travels a far more complex path: roots draw liquid up from deep underground, pull it through stems and branches, and release it as invisible vapor through tiny microscopic pores across green leaves.¹ Scientists group these two distinct pathways under the single term evapotranspiration, which serves as one of the primary engines connecting terrestrial ecosystems to the atmosphere.¹

Separating the two streams matters immensely when ecosystems enter periods of severe water stress. When a prolonged dry spell begins, moisture at the surface vanishes quickly as bare dirt bakes under the sun, shutting down direct evaporation within days. Deep-rooted vegetation, by contrast, can continue drawing water from deeper subterranean reserves, cooling local canopies and maintaining moisture flows until those deeper reserves run dry. Because soil drying and plant stress follow completely different physical schedules, predicting how a region will withstand severe heat waves or intensifying droughts requires knowing exactly how much water escapes through the dirt and how much moves through living leaves.¹

Why is separating evaporation from transpiration so difficult?

Evaporation and transpiration blend together instantly in the air, creating a single blended vapor signal that standard monitoring towers cannot easily disentangle. To calculate water movement, scientists trace an unbroken causal chain of physical drivers. Solar radiation delivers energy to warm the ground and leaves. Ambient air temperature and relative humidity set the atmospheric demand for vapor. Wind sweeps moisture away from the surface, while available soil moisture dictates whether roots and soil pores can supply liquid to meet that atmospheric thirst. Within this sequence, the most unpredictable gatekeepers are the leaf pores, known as stomata, which open and close dynamically in response to shifting sunlight, dry air, and dwindling groundwater.

Conventional physics-based land surface models use mathematical approximations to simulate how these plant pores open and close. The classical Penman-Monteith framework, a foundational mathematical system that balances surface energy and aerodynamic principles to estimate total evaporation rates, relies heavily on terms known as surface conductance values to gauge how freely moisture passes into the atmosphere.¹ Capturing surface conductance across mixed forest canopies, open grasslands, and varying soil textures has remained a long-standing challenge in Earth system science, because biological responses vary wildly across species and shifting weather patterns.¹

How does the new hybrid approach solve the partitioning problem?

The hybrid modeling system solves this bottleneck by embedding machine learning algorithms directly into the physical equations of the Penman-Monteith framework to calculate hard-to-measure surface conductance values while preserving strict physical boundaries.¹ Purely data-driven machine learning models often generate physically impossible predictions during weather extremes, such as creating phantom water fluxes out of dry ground, because algorithms simply chase statistical patterns in past data. Traditional physics models, conversely, maintain physical laws but frequently lack the biological nuance needed to track dynamic leaf behavior across shifting seasons.

A multi-institutional team of researchers developed the hybrid model to combine the strengths of both methodologies, creating a data-informed architecture that respects physical constraints like conservation of energy.¹ The scientists evaluated the hybrid model using observational records gathered from 47 National Ecological Observatory Network sites spanning a broad array of distinct vegetation types and environmental regimes.¹ When tested across these diverse landscapes, the hybrid model consistently outperformed traditional physics-based formulations as well as purely machine learning methods, maintaining high predictive accuracy when applied to entirely new test sites.¹

What can the model clarify during extreme environmental stress?

The hybrid framework maintained stable performance during extreme heat waves and severe droughts, revealing how key environmental variables dictate moisture partitioning across vulnerable ecosystems.¹ According to the research announcement released through Newswise, the hybrid approach highlights how radiation, soil moisture, and local vegetation structures govern the split between ground evaporation and vegetative water loss.¹ The model remained robust under intense climatic stress, demonstrating where traditional process representations break down and showing where physically constrained, data-informed calculations improve reliability.¹

Primary research funding for this work was provided by the Environmental System Science program within the Biological and Environmental Research program of the United States Department of Energy Office of Science.¹ Additional funding support was awarded by the Zegar Family Foundation and the National Science Foundation Division of Environmental Biology.¹ The resulting framework provides an operational pathway for integrating adaptive algorithms into larger continental and global simulations.¹

What are the boundaries of these findings?

These findings show how the model behaves across 47 terrestrial monitoring sites, but they do not represent a finished replacement for full-scale global Earth system models.¹ The hybrid model is a specialized computational framework evaluated against calibrated ground observations, meaning its current accuracy reflects conditions captured across the participating network locations. Its reliability depends on the availability of accurate input variables such as solar radiation and local soil moisture measurements.¹ Because the results come from a controlled multi-site evaluation rather than an operational global climate run, further testing is needed to verify whether the architecture maintains its stability when coupled directly into dynamic atmospheric forecasting engines over decades.

Even with these testing boundaries, the hybrid method establishes a scalable blueprint for resolving ecosystem water fluxes without abandoning fundamental physical laws. Climate modellers can use these data-informed insights to refine land surface simulations, helping regional authorities anticipate how local water reserves will behave as seasonal heat waves grow more intense.¹

What this rests on

19 sentences trace to 1 source.

All 1Documents 1Records 0Reporting 0People 0

1. 1 N Newswise Document · newswise.com Hybrid Modeling Improves Predictions of How Water Moves from Land to Atmosphere See the source