
Purpose: The overall heat-transfer coefficient (U) is an important descriptor of heat-exchanger performance in natural-gas dehydration. This study develops a compact artificial neural network (ANN) surrogate model for predicting U in a rich-monoethylene-glycol (MEG) heat exchanger represented by a steady-state Aspen HYSYS V14 model.
Materials and Methods: The dataset contains 1,500 structured computational operating points, comprising one converged HYSYS reference case and 1,499 points generated using a thermal calculation framework calibrated against the HYSYS reference effectiveness and effective heat capacities. Predictor auditing identified four varying operating variables: hot- and cold-stream inlet temperatures and the two stream mass flow rates. Exchanger geometry and fouling parameters were fixed. A 4-10-1 multilayer perceptron was developed using grouped operating-condition validation.
Results: On the fixed test partition, the ANN achieved R² = 0.9999990, RMSE = 1.39 × 10−3 W/(m².K), MAE = 1.15 × 10−3 W/(m².K), and MAPE = 0.0088%. Across ten repeated grouped partitions, the mean test RMSE was (1.90 ± 0.32) × 10−3 W/(m².K). Flow-edge hold-outs produced RMSE values of 2.41 × 10−2 to 2.47 × 10−2 W/(m².K). The reference U was 13.320 W/(m².K). Because the generator holds effectiveness, effective heat capacities, geometry, and fouling resistance constant, the target exhibits a flow-dependent relationship with 16 distinct U values.
Conclusions: The ANN accurately reproduces the HYSYS-calibrated computational relationship within the investigated operating domain. The proposed ANN provides a compact reduced-order surrogate for rapid repeated evaluation of the specified exchanger configuration. (J Multidiscip Eng 2026;1:125-149)
Citation data from Dimensions.
Vancouver (ICMJE)