Equinor Fellowship Research

Research Overview

My research under the Equinor Fellowship focused on two main thrusts:

1. Accurate and well-calibrated production forecasting
2. Well placement optimization (including one-shot and sequential optimization approaches)

Publications

1. Low-Rank Quadratic Knapsack Formulation for Large-Scale Well Placement under Gaussian Process Uncertainty

Ready for submission to the Journal of Computational Geosciences

This work addresses the need for computationally tractable methods in well placement optimization by developing a mean-variance model formulated as a 0-1 quadratic knapsack problem (QKP). The approach simultaneously handles:

• Spatial correlations in EUR/NPV forecasts via full covariance matrices
• Large-scale candidate drilling locations
• Operator risk aversion in the objective function

2. A Comparative Analysis of Heuristic-Based and Machine Learning-Driven Planning Strategies

Ready for submission to the Journal of Geoenergy Science and Engineering

This comparative study systematically evaluates heuristic-based drilling strategies against reinforcement learning (RL) approaches in closed-loop reservoir management. Key focus areas include:

• One-attempt constraint implementation (no environment resets)
• Role of Bayesian updates under geological uncertainty

3. Robust Stochastic Production Forecasting with Kernelized Gaussian Processes

Ready for submission to the Journal of Geoenergy Science and Engineering

A novel production forecasting model specifically designed for unconventional resource plays, developed using Equinor datasets. This represents the most complete work on production forecasting (Thrust 1).

Additional Research

The research program has generated several additional developments:

• Implementation of XGBOOST-based decision tree methods (adopted by Equinor)
• Development of ReserVAE, a probabilistic modeling approach
• Ongoing work on closed-loop field development (CLFD) as a POMDP, in collaboration with Chandra