A Domain-Driven, Physics-Backed, Proximity-Informed AI Model for PVT Predictions—Part I: Constant Composition Ex-Pansion

Constant-composition expansion (CCE) experiments provide critical relative-volume and density information describing the thermodynamic behavior of reservoir oils and gases under varying pressure. These properties are vital inputs for hydrocarbon reser-voir engineering, as they impact how oil and gas move through the reservoir during production. However, the need for specialized personnel, high-end equipment and measures taken to ensure safety in handling high pressure fluids often render the CCE experiments expensive and slow. This work introduces a Local Interpolation Method (LIM), a proximity-informed, end-to-end CCE fluid properties prediction AI model that leverages domain expertise and existing PVT data archives to generate surrogate CCE behavior for new fluids, thereby eliminating or reducing the need for completing laboratory CCE tests. Each new fluid is embedded in a compositional–thermodynamic descriptor space, and its response is inferred from a small neighborhood of thermody-namically similar fluids. Within this locality, the LIM combines hybrid local interpola-tion for key scalar properties (such as saturation-point quantities and expansion end-points) with shape-preserving reconstruction of monophasic and diphasic rela-tive-volume curves, enforcing continuity at saturation and consistency between rela-tive volume, density and compressibility. The workflow operates purely at inference time and does not require case-specific retraining. Application to a synthetic database of CCE tests shows that LIM reproduces key CCE features with very good agreement to laboratory data across a range of fluid types, indicating that proximity-based AI mod-elling can substantially reduce reliance on new CCE experiments while maintaining engineering-grade fidelity for compositional simulation workflows. The proposed ap-proach has been fully automated through software so it can be set up and directly uti-lized by the field operators on their own databases to significantly reduce their fluid sampling and laboratory analysis costs. The proposed model does not use others’ data while respecting the data privacy and data ownership.