CHEMFET Reponse for Supervised Learning of Neural Network

Electrical response from chemical field-effect transistor (CHEMFET) sensors intended to be selective to a specific ion is influenced by interfering chemical ions present in the solution. To be able to detect the main chemical ion of interest, we include a neural network post-processing stage after a readout interface circuit. This work focuses on the training data collection of potassium sensors in the presence of ammonium ions intended for the supervised learning of the neural network module. Using function fitting approach, the network aims to find the potassium ion concentration. Training data is obtained from sample solutions prepared by keeping the main ion concentration constant while the activity of the interfering ions based on the fixed interference method. The training algorithm is back-propagation with generalized delta rule on a multilayer feed-forward network. Activation function based on the MOSFET drain current equation in the linear region is attempted in the hidden layer. We find that referencing voltage readings to sensor response in deionized water prior to measurement improves repeatability of measured training data.