AI Modeling & Applications

AI Applications in Land Surface Modeling

With the rapid development of artificial intelligence technology, machine learning, deep learning, and symbolic regression methods have shown great potential in land surface process modeling. Through collaboration with other AI teams, we are committed to combining advanced AI technologies with traditional physical modeling methods to improve the accuracy and efficiency of land surface models.

Core Research Directions

Machine Learning-Driven Parameterization Schemes

Developing physical process proxy models using deep learning techniques to significantly improve computational efficiency while maintaining physical mechanism rationality
Neural network-based meteorological element downscaling methods to generate high-resolution atmospheric forcing data
Intelligent parameter optimization systems that can automatically adjust model parameters for different regions and application objectives

Advancing symbolic regression for earth science

Roadmap for developing evapotranspiration parameterization schemes based on symbolic regression

Applications of Symbolic Regression in Earth Science

Automatically discovering physical laws and mathematical relationships in land surface processes
Symbolic regression-based evapotranspiration modeling to improve simulation accuracy of evapotranspiration processes
Developing interpretable AI models to help understand the physical mechanisms of complex land surface processes

Multi-source Data Fusion and Quality Control

Using machine learning methods to identify new empirical relationships and parameterization schemes from massive observational data
Deep learning-based multi-source data fusion techniques
Intelligent data quality control algorithms to automatically identify and process anomalous data

Evapotranspiration dataset developed based on multimodal AutoML. Compared to other existing datasets, our multimodal machine learning-generated data has the longest time scale and best accuracy among many other characteristics.

Intelligent Evaluation and Diagnostic Systems

Applications of explainable AI in land surface process diagnostics to help understand the physical causes of model errors
Reinforcement learning-based adaptive modeling frameworks that enable models to continuously learn and improve during operation

Application Areas

Climate Prediction: Improving the accuracy of numerical weather prediction and climate forecasting
Hydrological Simulation: Improving simulation of hydrological processes such as runoff and evapotranspiration
Agricultural Applications: Crop yield prediction and agricultural management decision support
Environmental Monitoring: Ecosystem change monitoring and assessment
Urban Climate: Urban heat island effect and urban climate simulation

Open Research Topics

The application of AI technology in land surface modeling provides rich research opportunities for graduate students and collaborators. Here are our current research focuses:

Physics-Constrained Deep Learning Models: Developing methods to embed physical laws into neural networks, ensuring AI models maintain both high accuracy and physical rationality. This topic will focus on: designing physics-constrained loss functions that incorporate physical laws such as energy and mass conservation as constraints; developing differentiable physical process modules that enable end-to-end training of physical processes with neural networks; constructing physics-guided attention mechanisms that allow models to focus on physically important features; studying the impact of physical constraints on model generalization and interpretability; developing automatic discovery methods for physical constraints to automatically identify physical laws from data.

Multimodal Data Fusion AI Framework: Integrating multiple data sources including remote sensing, ground observations, and reanalysis data to build a unified multimodal AI modeling framework. Research includes: developing multimodal data alignment and fusion algorithms to handle data with different spatiotemporal resolutions; constructing multi-task learning frameworks to simultaneously predict multiple land surface variables; studying cross-modal knowledge transfer methods to leverage knowledge from one modality to improve predictions in another; developing uncertainty quantification methods for multimodal data; building end-to-end multimodal data preprocessing and modeling pipelines.

Explainable AI in Land Surface Process Diagnostics: Developing explainable AI methods to help understand the physical mechanisms behind model predictions and sources of errors. Key directions include: constructing attention mechanism-based interpretable models to identify the most important input features for predictions; developing counterfactual analysis methods to understand model decision logic; studying model uncertainty quantification methods to provide prediction confidence intervals; building physical process diagnostic tools to compare AI predictions with physical theory; developing visualization tools for model interpretation to help scientists understand AI "black boxes".

AI-Driven Parameterization Scheme Optimization: Using AI technology to optimize parameterization schemes in land surface models, improving model accuracy and computational efficiency. Research focuses include: developing reinforcement learning-based parameter optimization methods for automatic model parameter adjustment; constructing AI methods for parameter sensitivity analysis to identify key parameters; studying automatic discovery methods for parameterization schemes to learn optimal parameterization forms from data; developing intelligent algorithms for parameter space exploration; building frameworks to assess the impact of parameter uncertainty on model outputs; studying the adaptability of parameterization schemes under different climate conditions.

We welcome graduate students and collaborators interested in any of these topics to contact us and jointly advance innovative applications of AI technology in land surface modeling!

Selected Related Publications (# corresponding author):

Li, Q.#, Zhang, C.,Wei, Z.#,, Jin, X., Shangguan, W., Yuan, H., Zhu, J., Li, L., Liu, P., Chen, X., et al. (2024). Advancing symbolic regression for earth science with a focus on evapotranspiration modeling. npj Climate and Atmospheric Science, 7(1), 321.
Xu, Q., Li, L., Wei, Z.#,, Lu, X., Wei, N., Lee, X., Dai, Y. (2025). A multimodal machine learning fused global 0.1° daily evapotranspiration dataset from 1950-2022. Agricultural and Forest Meteorology, 372, 110645.
Cai, Y., Xu, Q., Bai, F., Cao, X., Wei, Z.#,, Lu, X.#, Wei, N., Yuan, H., Zhang, S., Liu, S., et al. (2024). Reconciling global terrestrial evapotranspiration estimates from multi-product intercomparison and evaluation. Water Resources Research, 60(9), e2024WR037608.
More related research results are being published…