Time Accounting Artificial Neural Networks for Biochemical Process Models

This paper is focused on developing more efficient computational schemes for modeling in biochemical processes. A theoretical framework for estimation of process kinetic rates based on different temporal (time accounting) Artificial Neural Network (ANN) architectures is introduced. Three ANNs that explicitly consider temporal aspects of modeling are exemplified: i) Recurrent Neural Network (RNN) with global feedback (from the network output to the network input); ii) Time Lagged Feedforward Neural Network (TLFN) and iii) Reservoir Computing Network (RCN). Crystallization growth rate estimation is the benchmark for testing the methodology. The proposed hybrid (dynamical ANN & analytical submodel) schemes are promising modeling framework when the process is strongly nonlinear and particularly when input-output data is the only information available.