ABSTRACT Software sensors, or soft sensors, can be a feasible option to monitor parameters that are difficult (or impossible) to measure with hardware sensors. At Henriksdal water resource recovery facility (WRRF), the operators have long experienced issues with a clogging sensor for the dry solid (DS) content in thickened primary sludge. A soft sensor was developed, and in the process, two methods were compared: long short-term memory (LSTM) network and linear regression. The first is a recurrent neural network that can capture non-linear dynamics, whereas the latter is a linear static model. The LSTM network was the best at predicting the DS content, with a mean squared error (MSE) of 0.341 with respect to laboratory data. The linear regression model performed worse than estimating a long-time average of daily manual samples but outperformed the online sensor. Replacing the existing sensor with the developed soft sensor can open up possibilities for more efficient control and operation of the thickener unit.