Transfer learning strategies for neural networks: A case study in amine gas treating units

This work presents a framework where strategies are applied within a workflow created to enhance the accuracy and transferability of the transfer learning process. We used a case study for predicting corrosion rates in gas-treating units, employing datasets from two different amines (A and B), where the amine A dataset is large compared to the amine B dataset. In the first neural network-based strategy, the inlet to hidden layer weights and biases are frozen after being trained with dataset A, while the ones from the hidden layer to the end response are freed. The freed weights and biases are then estimated via optimization. A set of X% of the amine B dataset values is added to the amine dataset A, and the neural network is then refitted, showing an increment in accuracy as more data from amine B is added. In the second neural network-based strategy, fine-tuning was used through a loss cross-entropy function in addition to' freezing' layers. Moreover, we compared our approach against Transfer-LASSO, an approach based on LASSO regression that looks at reducing model complexity while adding sparsity via regularization. Metrics of accuracy and transferability are introduced to evaluate class imbalance, computational time, and the effect of outliers. Our findings serve as a set of considerations when selecting transfer learning approaches in engineering problems. • Framework/workflow to enhance accuracy and transferability of the transfer learning process. • A comparison analysis between neural network-based transfer learning and Transfer-LASSO was performed. • Findings serve as a set of considerations when selecting transfer learning approaches in engineering.