Abstract
<jats:p>Data-driven property models support chemical process and molecular design, yet reliable prediction remains difficult beyond sampled chemical and operating-condition domains. Using pure-component viscosity, we diagnosed structural-and temperature-extrapolation errors and tested axis-specific, physics-guided derivative regularization. We analyzed 21,616 measurements for 2,501 compounds in a leakage-controlled comparison across linear and nonlinear model classes, treating descriptor selection as part of nested hyperparameter optimization without access to outer test data. When similar compounds were distributed across training and test sets, a standard neural network matched kernel ridge regression; however, its error increased more sharply than those of kernel and tree models as structural similarity decreased. Analysis of its descriptor-space prediction surface suggested that off-domain curvature contributed to this deterioration. Hessian regularization reduced median pathwise curvature by 43.2% and achieved the lowest compound-weighted mean absolute error among all evaluated models. For temperature extrapolation, a neural network with a soft penalty on deviations from inverse-temperature linearity provided the strongest performance across the 20 and 50 K tasks among the standard neural network, kernel ridge regression, and a network enforcing exact Arrhenius behavior. Notably, it yielded lower aggregate errors than compound-specific local Arrhenius fits, suggesting that shared molecular information enabled data-supported corrections to strict Arrhenius extrapolation. Combining the structural and temperature regularizers yielded the lowest mean integrated score. Together, these results show that axis-specific derivative regularization can translate neural-network flexibility into robust extrapolation, offering a transferable strategy for property prediction across chemical and operating-condition domains.</jats:p>