I am interested in the uncertainty quantification and optimization under uncertainty for solving complex physical problems, for instance fluid flow in porous media and full-wave form inversion. I am also interested in the use of machine learning as a choice of low-fidelity model to reduce computational cost and deal with big data. 

• MSc Energy Resources and Petroleum Engineering, King Abdullah University of Science and Technology, Saudi Arabia (Aug 2017- Dec 2018)
• BSc Petroleum Engineering, Institut Teknologi Bandung, Indonesia (Jul 2012- Oct 2016)

Project Title: Physics-based machine learning for real-time simulation update
Host Institutions: RWTH Aachen, ETH Zürich, Geophysica, Fraunhofer IEG
Supervisory Team: Florian Wellmann, Martin Saar, Renate Pechnig, Rolf Bracke
Start date: 1.4.2021

Enabling real-time update of geophysical simulations of coupled processes (thermal and hydraulic, hydraulic and mechanical) for realistic settings using a physics-based machine learning method and a combination with the High-Performance Computing finite element simulation environment MOOSE.

Consideration of various spatial parameterisations to capture heterogeneous parameter fields based on conventional spatial random field models, as well as geometric interface representations. 

A physics-based machine learning for geothermal reservoir simulations:

Thermo-Hydro-Mechano-Chemical (THMC) simulations are important for describing fluid-flow, reactive transport, and geo-mechanical processes in geothermal reservoirs. They are to assess efficiency in heat production, ensure safe operations, and evaluate the associated-environmental impacts, such as induced seismicity. These simulations are numerically expensive (especially for large-scale simulations) since they require relatively fine meshes to accurately capture the physical processes and guarantee numerical stability. This aspect makes them difficult and challenging to perform detailed parameter estimation studies (e.g., uncertainty quantification and global sensitivity anaysis). Therefore, surrogate models are needed to significantly reduce the computational costs.

In this project, we introduce the use of physics-based machine learning as the surrogate to perform large-scale TH(M)C simulations. As our physics-based machine learning technique, we use the non-intrusive reduced-basis (RB) method (see Figure (a)). It has two components: the construction of basis functions for preserving the physics (obtained through the Proper Orthogonal Decomposition) and the prediction of the associated weights via machine learning (e.g., Gaussian Process Regression and Neural Networks). At the current stage, we managed to compare the performance of non-intrusive RB towards data-driven machine learning using a hydro-mechanical benchmark case. We found that the data-driven machine learning produces small-scale perturbances of the prediction for a small number of training samples which is physically unreasonable, while the non-intrusive RB gives a high-accuracy for a comparable number of training samples.

(a)

(b)

Figure: The components of the physics-based machine learning (based on non-intrusive reduced-basis approach) (a), the comparison between the physics-based machine learning (POD-GPR and POD-NN) and the data-driven machine learning (GPR and NN) in term of accuracy toward Finite Element Method (FEM) result at small training set.