A Domain-Driven, Physics-Backed, Proximity-Informed AI Model for PVT Predictions—Part II: Differential Liberation Expansion and Viscosity Tests

Differential Liberation Expansion (DLE) and viscosity tests are core elements of the Pressure–Volume–Temperature (PVT) laboratory suite used to characterize reservoir oils under depletion and to support compositional modeling and reservoir simulation. Nevertheless, both DLE and viscosity testing remain expensive and time-consuming due to specialized equipment, strict operating procedures, and the need for experienced laboratory personnel. Building on our prior work that introduced the proximity-informed Local Interpolation Model (LIM) framework for Constant Composition Expansion (CCE), this study demonstrates how the same end-to-end, neighborhood-based workflow is applied to DLE and viscosity test data. A target fluid is embedded in a compositional–thermodynamic descriptor space and paired with a small set of thermodynamically similar fluids drawn from a PVT data archive. Within this locality, LIM is used to infer DLE behavior by combining local interpolation for key scalar quantities (e.g., saturation-point and endpoint PVT values) with shape-preserving reconstruction of pressure-dependent curves. For viscosity, the same approach reconstructs the oil viscosity curve μοp across the undersaturated and saturated regions. Evaluation on a proprietary database of DLE and viscosity tests shows strong agreement across diverse fluids for both DLE and oil viscosity trends. For example, across Tier 1–3 fluids, the mean curve mean absolute percentage error (MAPE) is 1.01% for Bo, 0.51% for ρo, and 1.32% for the liberated-gas Z-factor, while the conditioned baseline viscosity workflow yields a mean diphasic-branch MAPE of 7.75%. This supports reducing reliance on new DLE and viscosity measurements while maintaining engineering-grade fidelity in reservoir engineering and simulation workflows. This approach has been fully automated through software so it can be set up and directly utilized by the field operators on their own databases to significantly reduce their fluid sampling and laboratory analysis costs. Moreover, the proposed (artificial intelligence) AI model does not use others’ data, respecting data privacy and data ownership.