3D multi-physics uncertainty quantification using physics-based machine learning.

Area of Science:

• Geophysics and Earth Sciences
• Computational Science
• Machine Learning Applications

Background:

• Subsurface predictions rely on complex differential equations, leading to computationally intensive, high-dimensional problems.
• Estimating uncertainties in material parameters further exacerbates computational challenges for existing models.

Purpose of the Study:

• To develop a novel hybrid physics-based machine learning technique for creating efficient surrogate models.
• To enable reliable, scalable, and interpretable predictions for complex subsurface physical processes.
• To facilitate probabilistic analyses, including sensitivity studies and uncertainty quantification.

Main Methods:

• Introduction of the non-intrusive reduced basis method, combining physical process models with data-driven machine learning.
• Application to a thermo-hydro-mechanical reservoir simulation to demonstrate capabilities.
• Development of surrogate models that overcome limitations of purely physics-based or data-driven approaches.

Main Results:

• Achieved orders of magnitude computational gain compared to traditional methods.
• Maintained accuracy exceeding measurement errors for complex, non-linearly coupled physical problems.
• Enabled effective global sensitivity studies and uncertainty quantification.

Conclusions:

• The hybrid physics-based machine learning approach offers a powerful solution for computationally prohibitive geoscientific problems.
• The non-intrusive reduced basis method provides a scalable and interpretable alternative for subsurface modeling.
• The technique is broadly applicable to various geoscientific challenges beyond the illustrated reservoir application.