Date & Time: 28th November 2025.
Title: Efficient Nonlinear Preconditioning for Reservoir Simulation History Matching Using Random Features Learning
Speaker: Antoine Lechevallier holds a French engineering degree in Geosciences with a major in Numerical Geosciences and a specialization in AI and Big Data. He completed a PhD in Applied Mathematics between Sorbonne Université and IFP Energies Nouvelles, focusing on Scientific Machine Learning for improving reservoir simulations. His postdoctoral work at NORCE continues this line of research, emphasizing the practical integration of ML into industrial workflows to robustly accelerate simulations and enhance real-world performance.
Read below the highlights from the webinar
Smarter Reservoir Simulation: Combining Physics and Machine Learning
Reservoir simulation is a critical tool in energy production, helping engineers predict how oil, gas, and water move underground. These simulations guide decisions on field development and resource management, but they are computationally demanding and often slowed down by nonlinear solver failures.
Our latest research based on Hybrid Newton, a machine learning–enhanced approach that accelerates reservoir simulations without replacing the underlying physics. This method acts as a “smart assistant” to traditional solvers, learning from previous runs and predicting better starting points for complex calculations. The result: fewer nonlinear solver failures at desired point in time, and improved efficiency in workflows such as history matching, where models are adjusted to match observed production data.
Key Innovations
- Data-Driven Preconditioning: Neural networks improve solver performance while preserving physical accuracy.
- Extreme Learning Machines: A lightweight training technique based on Random Features Learning that reduces model training time from minutes to seconds.
- Operational Integration: Now available in OPM Flow, an open-source reservoir simulator widely used in industry.
Why It Matters
- Efficiency: Reduced CPU time and fewer failures mean faster results.
- Scalability: Works seamlessly in history matching workflows requiring thousands of simulations.
- Future-Ready: Opens the door for broader applications in complex fields and other physical systems.
For Everyone: What Does This Mean?
Imagine trying to predict how water, oil, and gas move underground—it’s like solving a giant puzzle with millions of pieces. Traditionally, computers do this by running heavy calculations, which can sometimes get stuck and waste time. Our new method adds a “smart helper” powered by machine learning. It learns from past runs and gives the computer better hints, so the process is faster and less error-prone. This means quicker results, smarter decisions, and more efficient energy production, all without changing the science behind it.
Next Steps: This breakthrough is now integrated into OPM Flow and ready for operational testing. Future work will focus on scaling the method to complex fields and automating detection of nonlinear issues.