Skip to main content

Harnessing AI for Groundwater Predictions

Date
Harnessing AI for Groundwater Predictions

Groundwater is a vital resource, supplying nearly half of the world's drinking water and playing a crucial role in agriculture. However, predicting groundwater levels and flow accurately is challenging, often requiring complex and computationally demanding models. My PhD research explores a new approach, integrating advanced artificial intelligence (AI) techniques to enhance groundwater prediction models.

Unlike physically-based models, which typically demand a detailed understanding of local conditions and extensive calibration, data-driven models such as artificial neural networks can predict key variables using only relevant influencing factors. One of the most exciting aspects of using AI for groundwater prediction is the significant increase in speed and efficiency. Initially, AI models require a significant amount of time and resources to train. However, once trained, these models can make predictions incredibly quickly, often in just milliseconds. This rapid inference capability means that these models can be used in real-time applications, offering immediate insights and aiding quick decision-making processes.

From Theory to Practice: AI in Groundwater Prediction

My research journey began with creating a comprehensive dataset of synthetic data. This initial step was crucial for starting with complete data, allowing a solid foundation before tackling the more complex and data-sparse real-world scenarios. I began by applying computer vision techniques, which are highly effective in handling image-like data structures, to predict groundwater flow patterns.

As the research progressed, I then explored the use of neural operators, advanced AI models designed to learn physical relationships from scientific data. Unlike traditional methods that rely on complete datasets, neural operators can work with sparse data, learning relationships throughout the space-time domain. This capability is particularly valuable for real-world applications where data can be incomplete or unevenly distributed.

During the last months of my thesis, I dived into real-world data. This transition required a different approach due to the nature of the data. Real-world datasets often come with challenges such as missing values and inconsistent data quality. To address these, I adapted my models to handle the complexities of real-world scenarios, ensuring robust and reliable predictions. By integrating spatial and temporal data, these models can predict future groundwater conditions, helping in long-term planning and decision-making. This aspect of AI modeling is crucial for sustainable groundwater management.

Comparison of groundwater level distribution predictions made by an AI model. The left figure shows the ground truth distribution of groundwater levels with five pumping wells as generated by the traditional numerical model. The middle panel displays the AI model's prediction of the groundwater levels. The right panel presents the error map, which highlights the differences between the ground truth and the predicted values. The AI model effectively learns the impact of pumping wells on the underground groundwater distribution, as evidenced by the close alignment between the ground truth and predicted distributions.

The Impact of AI on Groundwater Management

The application of AI in groundwater modeling represents a significant advancement in the field of hydrology. These models are not only more accurate but also faster and more adaptable to different scenarios. By improving prediction accuracy and efficiency, AI models can help manage groundwater resources more effectively, making it a crucial tool for addressing global water challenges.

By Dr. Maria Luisa Taccari

Dr. Taccari recently graduated with her PhD from the University of Leeds, focusing on the development of deep learning surrogate models in groundwater flow simulation. During her PhD, Maria Luisa collaborated with the Dutch research institute Deltares and Brown University. Dr. Taccari now holds a position as a scientist at ECMWF, the European Centre for Medium-Range Weather Forecasts.

Paper references

  • M.L. Taccari, J. Nuttall, X. Chen, H. Wang, B. Minnema, and P. K. Jimack, “Attention U-Net as a surrogate model for groundwater prediction”, Advances in Water Resources, vol. 163, p. 104169, 2022.
  • M.L. Taccari, H. Wang, S. Goswami, M. De Florio, J. Nuttall, X. Chen, and P. K. Jimack, “Developing a cost-effective emulator for groundwater flow modeling using deep neural operators”, Journal of Hydrology, 2023.
  • M.L. Taccari, H. Wang, J. Nuttall, X. Chen, and P. K. Jimack, “Spatial-Temporal Graph Neural Networks for Groundwater Data”, In Review.
  • M.L. Taccari, O. Ovadia, H. Wang, A. Kahana, X. Chen, and P. K. Jimack, “Understanding the Efficacy of U-Net & Vision Transformer for Groundwater Numerical Modelling”, Synergy of Scientific and Machine Learning Modeling - ICML 2023 Workshop.