Past Event: PhD Dissertation Defense

Scalable, Risk-Aware, and Uncertainty-Driven Forward and Inverse Problems in Reservoir Modeling

Ryan Farell, PhD Candidate, Oden Institute

Abstract

This talk addresses three core computations that connect field data to drilling decisions. The first is Bayesian history matching, where subsurface geological properties are inferred by calibrating a multiphase flow simulator against sparse historical well rates and pressures. To accelerate this inversion, we train a stochastic‐Hamiltonian surrogate whose learned state‐dependent mass and damping operators reproduce the simulator’s gradients and preserve symplectic structure, cutting the number of forward runs needed while maintaining posterior accuracy. The second is probabilistic forecasting, where a neural-ODE mixture of Gaussian process kernels that produces probabilistic forecasts that honor the physics while under geological uncertainty. The third is risk-averse well placement, cast as a mean–variance 0-1 quadratic knapsack problem, in which a streamed SketchyCoreSVD low-rank factorization cuts each solver node’s cost from O(NxN) to O(Nr). This enables optimization of tens of thousands of wells with negligible loss in certainty-equivalent value. Together, these components form a scalable, uncertainty-aware reservoir development workflow that accelerates decision-making and maximizes economic returns.