Advancing Ecosystem Understanding through Knowledge-Guided Machine Learning

[Grigory Bronevetsky]

Modeling Talks

Advancing Ecosystem Understanding through Knowledge-Guided Machine Learning Licheng Liu, U Minnesota Tues, August 5, 2025 | 9am PT Meet | Youtube Stream Hi all, The presentation will be via Meet and all questions will be addressed there. If you cannot attend live, the event will be recorded and can be found afterward at https://sites.google.com/modelingtalks.org/entry/advancing-ecosystem-understanding-through-knowledge-guided-machine-learning More information on previous and future talks: https://sites.google.com/modelingtalks.org/entry/home Abstract:  Ecosystems are central to the global carbon and nitrogen cycles, yet their complexity and heterogeneity pose major challenges to modeling efforts. While process-based models offer theoretical rigor grounded in biophysical and biochemical principles, they often face limitations in computational scalability, data assimilation, and structural uncertainty. Conversely, purely data-driven models can leverage rich sensing data streams but tend to lack generalizability and scientific interpretability. Knowledge-Guided Machine Learning (KGML) bridges these paradigms by embedding domain knowledge into machine learning frameworks to produce models that are both accurate and mechanistically informed. In this talk, I will introduce a series of applications that demonstrate how KGML improves simulations of carbon and nitrogen fluxes across agricultural and natural ecosystems. Use cases include modeling carbon budgets from croplands, advancing digital twin frameworks for sustainable agriculture, and predicting methane dynamics in natural ecosystems. By blending scientific priors with data-driven flexibility, KGML facilitates ecosystem modeling that is explainable, scalable, and climate-action-relevant. Ultimately, this approach supports more robust decision-making in environmental management and provides a pathway for developing AI systems that are physically consistent, data-efficient, and applicable to complex Earth system processes. Bio: Dr. Licheng Liu is a research scientist at the University of Minnesota and the lead of the KGML division in the NSF AI-LEAF Institute, and the AI for Nature Methane working group. His research integrates process-based modeling, machine learning, in-situ sensing, and remote sensing to understand biogeochemical dynamics in agricultural and natural ecosystems, with a focus on greenhouse gas emissions and climate feedbacks. His work spans carbon-nitrogen-water cycle modeling, AI-enhanced crop and soil simulations, and the development of open-source KGML frameworks for ecosystem prediction. Starting January 2026, Dr. Liu will join the University of Wisconsin–Madison as an Assistant Professor in the Department of Biological Systems Engineering.

[Grigory Bronevetsky]

Video: https://youtube.com/live/61MK8csN_jE

Slides: https://drive.google.com/file/d/1ywXQ3KTf_FFjoQ7dgm77q9vCeUPsd2CJ/view?usp=sharing

Summary:

• Focus: understanding the behavior of the world's ecosystems.

  • Increasing global populations have put a strain on ecosystems and natural resources: land use, water use, energy use

  • To sustain ecosystems we need to understand them: carbon, water, nutrients

  • Strongly coupled/interdependent with

    • Global climate: rainfall, temperature, variability

    • Human management: fertilization, irrigation, land use

• Measuring ecosystems is challenging

  • Limited measurement data

  • Challenges: 

    • Ecosystem heterogeneity: soil, vegetation, management

    • Complex biogeochemical, physical processes

• Opportunity: AI for ecosystem understanding

  • Data: in-situ sensor networks, remote sensing, meteorological data, bopgeochemical, geospatial&survey, synthetic

  • Applications: greenhous gases, carbon sequestration,. Production, water, air quality, soils, biodiversity, natural hazards

• Modeling can infer unknown information from measurements

  • Process-based models incorporate known dynamics but are limited by process representation

  • Machine learning models (black box) adapt to data but have poor explainability and generalizability

  • Hybrid AIU models can leverage the best of both techniques (e.g. differentiable simulations, SciML)

  • Knowledge-Guided ML: Training based on domain-specific prior knowledge: invariants, useful examples

• Example: advancing agroeconomics in US corn belt

  • Challenge: simulating N2O emissions from use of fertilizer in corn farming

  • High spatial/temporal variability

  • KGML: 

    • Train ML model using runs oc ecosys model https://github.com/jinyun1tang/ECOSYS

    • KGML model is able to accurately predict N2O emissions

    • Supports data inversion across US midwest cornbelt: infer emissions hotspots from sparse regional measurements

    • Supports interpretability via causal diagrams, which can guide additional improvements in process-based model (using PCMCI: https://jakobrunge.github.io/tigramite)

• Example: understanding carbon budget in agriculture

  • Agriculture is both a source and a sink for carbon

  • Modeling emissions from agricultural activities

  • Train ML model based on traces of ecosys model

  • Used a knowledge-guided loss function: mass balance, threshold control, response control

  • Knowledge-guided extrapolation by assimilating remote-sensed data

  • Hybrid model

    • Outperforms both pure-ML and ecosys alone in accuracy

    • More efficient than ecosys

  • Can use emissions model to estimate carbon credit risk

• Example: Advancing natural ecosystem understanding with hybrid AI

  • Methane is a major contributor to climate change

  • Data on emissions is sparse and diverse

  • Pure AI models are good in-sample but don’t generalize

    • Most data are limited to high latitude in North Hemisphere

    • Poor accuracy in tropics, South hemisphere

  • Integrating scientific knowledge can improve accuracy

    • Knowledge-guided initialization TEM-MDM: https://www.eaps.purdue.edu/ebdl/resources/ecosystems-biochemical.html

    • Training ML model on its traces improves model accuracy

  • AI model can help plan new locations for placing measurement sites across the world to maximize predictive accuracy

  • AI for Natural Methane Working Group: 

    • https://esiil.org/working-groups/ai-natural-methane

    • https://github.com/CU-ESIIL/AI-for-Natural-Methane

• Ongoing work:

  • Water quantity & natural concentration

  • Precision agriculture

  • Global carbon cycle