Computational Geosciences 2026 - Hierarchical variable-fidelity surrogate optimization
Description
Optimization histories and processed results supporting the manuscript "Robust well control optimization under geological uncertainty using variable-fidelity surrogates" (Computational Geosciences, Springer). Data cover five optimization scenarios (A–E) run on the UNISIM-I-M benchmark reservoir: two deterministic scenarios (A, B) and two robust scenarios (C, D) comparing a three-level variable-fidelity multilevel strategy (VF-ML) against a high-fidelity single-level baseline (HF-SL), plus two realization-reduction variants (STAT, ADAPT) built from Scenario C.
Files
Steps to reproduce
Well control optimization using a three-level variable-fidelity surrogate framework. At each fidelity level, a cubic RBF surrogate is fitted to Latin Hypercube samples within a local trust region and used to drive a trust-region optimizer. Grid coarsening is performed via wavelet-based multiresolution upscaling with stationary wavelet transform refinement decisions. The three levels (VF-L1, VF-L2, VF-L3) are executed hierarchically: the coarsest level performs broad exploration; finer levels refine the solution in progressively smaller trust regions. For robust optimization, the objective is the arithmetic mean of NPV over a subset of five representative geological realizations selected from a 48-realization ensemble (UNISIM-I-M benchmark). An adaptive variant re-ranks the ensemble at fidelity transitions using inexpensive VF-L3 evaluations to maintain subset representativeness as the optimal control evolves.