Disruptions in tokamaks are instabilities events which can damage the machi
ne components. The avoidance and mitigation of these events is desirable in
present machines as well as in Next Step devices (such as ITER). A neural
network has been developed to predict the occurrence of disruptions caused
by edge cooling mechanisms in ASDEX Upgrade. The network works reliably and
is able to predict the majority (85%) of the disruptions. The neural netwo
rk has been trained to predict the time interval up to the disruption and t
his makes it suitable to be used on-line either to avoid disruptions (by me
ans of auxiliary heating and reduction of gas puffing) or to mitigate the u
navoidable ones. For this last purpose, a solid pellet injector has been de
veloped and tested; the injected impurity pellets have been shown to reduce
the vertical forces and the conductive fluxes to the divertor. (C) 2001 El
sevier Science B.V, All rights reserved.