Soft computing models for predicting polymer viscosity across wide thermal and ionic-strength conditions

Characterizing polymer properties under varying thermal and ionic-strength conditions is crucial across diverse fields, from materials science to environmental engineering. While experimental techniques are effective, they are often time-consuming and costly, underscoring the need for rapid, reliable computational alternatives. This study evaluates the performance of various soft computing techniques, ranging from traditional machine learning methods (MLP, SVR, RF, GBM, LGBM, and XGB) to advanced deep learning approaches (CNNs), for predicting polymer viscosity under diverse temperature and salinity conditions. The models are trained on a curated database of 4,366 experimental measurements spanning ten commercially used polymers and wide ranges of molecular weight, polymer concentration, temperature, salinity, and shear rate. Using hold-out validation, 10-fold cross-validation, and a strict unseen-polymer test, all models achieve high accuracy (test R² ≥ 0.98), with LGBM and CNN providing the best overall performance. Compared with traditional models (MLP and SVR), these methods reduce test-set error metrics by approximately 25–30%, and the CNN attains R² = 0.997 on the unseen-polymer dataset. Permutation-importance analysis identifies polymer concentration as the dominant predictor, followed by shear rate and molecular weight, in agreement with established rheological behavior. The results demonstrate that ensemble and deep learning models can serve as fast, reliable surrogates for polymer viscosity across high-temperature and high-salinity conditions, reducing experimental effort and supporting data-driven optimization of polymer selection and operating conditions.