PREDICTION OF LARGE EVENTS ON A DYNAMICAL MODEL OF A FAULT

We present results for long-term and intermediate-term prediction algo rithms applied to a simple mechanical model of a fault. The long-term techniques we consider include the slip-predictable and time-predictab le methods and prediction based upon the distribution of repeat times between large events. Neither the slip-predictable nor time-predictabl e method works well on our model. In comparison, the time interval met hod is much more effective and is used here to establish a benchmark f or predictability. We consider intermediate-term prediction techniques which employ pattern recognition to identify seismic precursors. Thes e methods are found to be significantly more effective at predicting c oming large events than methods based on recurrence intervals. The per formances of four specific precursors are compared using a quality fun ction Q, which is similar to functions used in linear cost-benefit ana lysis. When the quality function equally weights (1) the benefit of a successful prediction, (2) the cost of maintaining alerts, and (3) the cost of false alarms, we find that Q is optimized in algorithms based on the most conventional precursors when alarms occupy 10-20% of the mean recurrence interval and approximately 90% of the events are succe ssfully predicted. The measure Q is further used to explore optimizati on questions such as variation in the space, time, and magnitude windo ws used in the pattern recognition algorithms. Finally, we study the i ntrinsic uncertainties associated with seismicity catalogs of restrict ed lengths. In particular, we test the hypothesis that many shorter ca talogs are as effective as one long catalog in determining algorithm p arameters, and we find that the hypothesis is valid for the model when the catalogs are of the order of the mean recurrence interval.