Multiple fluidised bed reactors for water softening (crystallisation of calcium carbonate) have been in operation at drinking water treatment plants since the late 1980s. Research on the operation of these pellet reactors has been focussed on investigating crystallisation under constant fluidised bed composition. However, in practice the bed composition varies frequently, despite large effort of plant operators. An improvement in the control of the fluidised bed can be achieved by using model-based multivariable control. Due to the nonlinear behaviour of the reactor, caused by water temperature variation during the year, a nonlinear control approach is used. A particle filter, based on a first-principles model, estimates the state of the softening reactor and a nonlinear model-predictive controller determines the values of the manipulated variables. We show in a simulation experiment, that it is possible to keep the reactor at desired operational parameters (pellet size and bed height) under varying operational conditions. In this way, the cost of pellet softening can be reduced and irregularities prevented. (C) 2007 Elsevier Ltd. All rights reserved.