The research draws on data collected through the JUNEAU and JANUSS projects and organised within DATA, including field sensor measurements, weather observations and satellite data.
AI-based methods are investigated alongside the physical and mechanistic models traditionally used in environmental sciences. Comparing these approaches helps assess their respective performance, their ability to remain reliable as environmental conditions change, and the scientific insights they can provide. The resulting models and findings are progressively integrated into the DIGITAL TWINS project, which oversees the assembly of JUNON's digital twins.
Evaluating predictive approaches
Environmental forecasting relies on a wide range of methods, from physics-based and statistical models to machine learning techniques. Comparing their performance is a key part of the PREDICTION project.
A dedicated evaluation pipeline brings these approaches together and tests them against real-world environmental time series. The assessment considers several criteria, including prediction accuracy over short, medium and long time horizons, the ability to generalise to new situations, and model performance as data patterns change.
An initial version of the benchmarking framework is already being used to compare different models across five test years and provide a consistent basis for performance analysis. Ultimately, this work will contribute to methodological guidelines to help researchers select the most appropriate forecasting approaches according to the available data and their scientific objectives.
Identifying common patterns in environmental time series
Environmental time series can reveal similar patterns across different monitoring sites, even when those sites are geographically distant. For example, comparable groundwater dynamics may be observed in different parts of France.
To identify these similarities, clustering techniques have been applied to groundwater level time series, grouping datasets with shared characteristics. These groups can then be used to train models that work across multiple monitoring sites, rather than developing a separate model for each piezometer.
This approach can improve prediction reliability and generalisation while reducing the number of models that need to be maintained.
Another way to address differences between time series is through meta-learning. These techniques can automatically identify the most suitable models based on the characteristics of a given dataset, helping to develop tools that support researchers in selecting appropriate modelling strategies.
Combining physics-based models and Artificial Intelligence
PREDICTION also explores hybrid approaches that combine physical knowledge with machine learning techniques. This includes research into Physics-Informed Neural Networks (PINNs), which incorporate physical laws directly into the learning process.
These hybrid models offer a way to make better use of existing scientific knowledge while benefiting from the flexibility of AI-based methods. For example, they can help estimate variables that are difficult to observe directly or improve prediction reliability by ensuring that model outputs remain consistent with physical principles.
Integrating spatial information and multiple data sources
Environmental conditions vary significantly across regions, while the information needed to understand them comes from a wide range of sources. PREDICTION explores ways to bring these different types of data together within predictive models.
Satellite observations are particularly valuable for studying atmospheric pollution and air quality. Although Earth observation instruments provide extensive geographical coverage, their spatial resolution is not always sufficient to capture local variations.
Machine learning methods can help overcome this limitation by reconstructing finer spatial patterns from satellite observations while maintaining consistency with the underlying physical measurements.
Developing models that adapt to new data
Environmental systems are constantly changing in response to climate variability, human activities and changes in land use. Predictive models therefore need to remain effective as new conditions emerge.
PREDICTION investigates continual learning approaches, allowing models to update automatically as new observations become available. Rather than retraining models from scratch whenever new data are collected, these methods enable more efficient updates, reducing computational costs, storage requirements and the environmental footprint of model operation.
This research also challenges conventional AI evaluation practices, which often focus primarily on prediction accuracy against fixed benchmark datasets. In environmental applications, assessing a model also requires considering its resilience to changing data, its ability to generalise and the resources needed throughout its lifecycle.
Initial findings from groundwater forecasting studies suggest that more complex AI architectures do not necessarily deliver significant improvements over simpler methods.
These results support the development of more sustainable predictive approaches, better suited to the scientific, operational and environmental requirements of digital twins.
Explainability and evaluation of predictive models
Another area of research focuses on making AI-based predictions easier to understand and interpret.
This includes identifying the variables that have the greatest influence on predictions, estimating the level of confidence associated with model outputs, and exploring relationships between environmental variables.
Improving model explainability helps researchers better understand the processes being studied and makes predictive tools more useful for practical environmental management.
Towards predictive Digital Twins
The models developed within PREDICTION will contribute to JUNON's environmental digital twins. Their integration will help create tools capable of forecasting environmental changes over different time horizons, updating predictions as new data become available, and simulating how complex natural systems may respond to different scenarios and future projections.
In the longer term, this research will help establish methods for developing environmental digital twins that can anticipate changes in natural systems and support more sustainable resource management.
Scientific lead:
Chef de projet Grande Infrastructure de Recherche
BRGM, 3 avenue Claude Guillemin, BP36009
45100 Orléans
France