Machine learning, data assimilation, digital twins: how AI complements — without replacing — physical models of groundwater flow.
Physical models (MODFLOW, FEFLOW) remain the reference for simulating aquifers. AI does not substitute a black box for them: it accelerates their calibration and enriches data use.
What AI brings
- Anomaly detection in piezometric time series
- Gap-filling and denoising of measurement series
- Fast surrogates to explore scenarios
- Support for calibrating physical models
“AI does not remove the need to understand hydrogeology: it amplifies expertise, it does not replace it.”
A cautious adoption
Data quality and representativeness condition everything. Without a reliable dataset and physical validation, a high-performing statistical model can mislead with confidence.
Towards digital twins
The convergence of physical models, machine learning and real-time field data opens the way to aquifer digital twins: living replicas, continuously recalibrated, on which management scenarios can be tested before being applied.
This prospect does not exempt anyone from rigour: a digital twin is only as good as the data feeding it and the physical understanding underpinning it.
