Title: Think or Sink: unravelling anthropogenic causes of subsidence with a hybrid AI approach
Thibault G.G. Candela (1), Joana C. Esteves Martins (2), Peter A. Fokker (1), Aris Lourens (3), Willem Dabekaussen (3), Wilfred F.J. Visser (4), Erik A.F. Langius (5), Madelon S. Molhoek (6), Matthias S. Fath (6), Andrei Bocin-Dumitriu (3) & Kay Koster (3)
TNO Applied Geosciencess (1); TNO Advisory Group for Economic Affairs (2); TNO Geomodelling (3); TNO Geo Data and IT (4); TNO Monitoring and Control Services (5); TNO Data Science (6)
Event: Abstract GeoUtrecht2020
The Netherlands is subject to anthropogenic and natural subsidence with rates which are an order of magnitude higher than sea-level rise. Because one-third of the Netherlands lies below mean sea level, subsidence may threaten the country’s subsistence with major socio-economic consequences. Subsidence is a normal natural process but is overprinted and accelerated by anthropogenic activities which causes deep or shallow processes. Deep processes are caused by the extraction of hydrocarbons and salt, whereas shallow processes are primarily caused by lowering of phreatic groundwater levels. At present, the relative contribution of each process to total subsidence (i.e. natural plus anthropogenic) is unclear. Such information is important for stakeholders to support decision making on subsidence mitigation and it should be substantiated with independent scientific studies.
We present the outline of a hybrid Artificial Intelligence (AI) big data and model workflow to disentangle different subsidence forcing and we report on preliminary results for an area covering a gas field in the Friesland coastal plain and an area in the peat-rich central Rhine-Meuse delta plain. The proposed workflow is a hybrid approach between a knowledge-based physical model and machine-learning techniques. The big input data comprises a suite of static (structural model) and time-dependent subsurface data (phreatic groundwater level, reservoir pressure), and geodetic measurements. Geomechanical models provide the connection between the drivers (groundwater levels and reservoir pressures) and the surface movement.
In parallel, a Bayesian approach is developed solely coupling physics-based models for both deep and shallow processes and data assimilation techniques. Performances of both approaches, the hybrid AI and physics-based, are compared in the aim to achieve more accurate and more reliable spatiotemporal subsidence predictions for each subsidence forcing.