R. Vanbalen et S. Cloetingh, NEURAL-NETWORK ANALYSES OF STRESS-INDUCED OVERPRESSURES IN THE PANNONIAN BASIN, Geophysical journal international, 121(2), 1995, pp. 532-544
Artificial neural networks can learn relationships between sediment ch
aracteristics (burial depth, composition, coordinates and thickness of
overlying Quaternary deposits) and overpressures from well data, afte
r which they can interpolate and extrapolate to areas and depths not c
overed by wells. We analyse data from the south-eastern part of the Pa
nnonian Basin. We use a neural network for analysing fluid overpressur
es because of the complex interaction of the key variables, making it
difficult to derive the functional relationships required for a statis
tical analysis. The optimal topology of the network (number of hidden
layers and neurons) is found by minimizing the network's training and
testing errors. The optimal design of the network resembles the intera
ctions scheme of the key variables. The Pannonian Basin, originally fo
rmed in an extensional regime, has been in a compressive state of stre
ss since Late Pliocene, causing anomalous subsidence patterns. Numeric
al forward modelling of compaction-driven fluid overpressures shows th
at, due to an increase in the level of compressive interplate stress,
the fluid overpressures in the deep subbasins have increased substanti
ally since Late Pliocene, giving rise to a very high overpressure (up
to 45 MPa) at present. The neural network analyses provide an independ
ent estimate of the current amount of overpressuring in this basin, co
mplementing the numerical forward modelling results. The overpressure
profiles obtained by the two modelling approaches are in excellent agr
eement, showing the same magnitude of overpressures, a reversal of the
overpressure in the deepest parts of the subbasins and a general decr
ease of the overpressure from SW to NE.