Research
My research develops computational methods for subsurface energy and environmental systems, with an emphasis on scientific machine learning, inverse modelling, uncertainty quantification, and optimization.
Many subsurface problems are defined by the same practical constraints: observations are sparse, geological structure is uncertain, high-fidelity multiphysics simulation is expensive, and engineering decisions must be made under incomplete information. I use machine learning where it can reduce this computational burden or extract information that is difficult to obtain directly, while retaining the governing physics as an explicit part of the modelling framework.
Current work focuses on geothermal reservoirs, geological CO2 storage, fractured-media flow, and transferable scientific models for subsurface systems.
Research directions
Scientific machine learning for subsurface multiphysics
Learning reduced and surrogate representations of pressure, temperature, saturation, and coupled reservoir dynamics across heterogeneous porous and fractured media. The objective is to make repeated simulation, uncertainty analysis, and history matching feasible at field scale without treating physical simulation as a black box.
geothermal · CO₂ storage · multiphysics simulation · surrogate modelling
Generative inverse modelling and uncertainty quantification
Developing generative approaches for reconstructing geological structure and subsurface states from sparse, indirect observations. A particular focus is fracture-network inversion, where non-uniqueness and limited observability require probabilistic rather than purely deterministic solutions.
diffusion models · inverse problems · fracture characterization · data assimilation
Optimization and decision-making under geological uncertainty
Combining surrogate models, evolutionary optimization, and active learning to reduce the number of expensive simulations required for reservoir design and operation. Applications include well placement, injection and production control, heat-extraction design, and multi-objective trade-offs among energy recovery, risk, and environmental performance.
multi-objective optimization · active learning · reservoir design · decision support
Transferable models and autonomous workflows for geoscience
My current research is moving from task-specific models toward reusable representations and scientific workflows that can transfer across reservoirs, physical conditions, and modelling tasks. This includes foundation-model concepts for subsurface systems and agentic workflows that connect observation, simulation, inference, and decision-making.
scientific foundation models · transfer learning · agentic workflows · scientific discovery
Current research
Subsurface multiphysics modelling and scientific foundation models
At UC Berkeley and Lawrence Berkeley National Laboratory, I work on machine-learning methods for geothermal-reservoir modelling and transferable subsurface representations, including research connected to The Geysers, Cape Station, and Utah FORGE.
Selected research
Physics-supervised generative inversion of fracture networks
A generative inverse-modelling framework for recovering fracture-network structure from indirect hydraulic observations while enforcing physical consistency in the inferred solutions.
Diffusion modelling of spatiotemporal flow in fractured media
Conditional generative modelling of transient pressure and temperature fields in stochastic fracture networks, with probabilistic ensembles for uncertainty characterization and subsequent inverse analysis.
Machine-learning-accelerated design of fractured geothermal systems
Surrogate-assisted multi-objective optimization for fractured geothermal reservoirs, designed to identify high-value operating strategies with substantially fewer high-fidelity simulations.
Multi-fidelity learning for geothermal system design
Knowledge-transfer methods that combine simulations of different fidelities to improve predictive accuracy and reduce the computational cost of geothermal-reservoir optimization.
Data-driven production optimization in subsurface reservoirs
A series of surrogate-assisted evolutionary methods for high-dimensional well-placement and control problems, forming the methodological basis for my later work on geothermal design and scientific machine learning.
Collaborative projects
- Subsurface Multi-Physics Modelling & Scientific Foundation Model
UC Berkeley & Lawrence Berkeley National Laboratory - ADVANCEA — Advancing Controlled Environment Agriculture Through Data-Driven Decision-Making and Workforce Development
U.S. Department of Agriculture · US$3.77M - Poshan Drainage Tunnel System as a Hillslope Critical Zone Observatory
Research Grants Council of Hong Kong · Collaborative Research Fund · HK$2.79M - Digital Twin-Empowered Landslide Emergency Risk Management
Research Grants Council of Hong Kong · Theme-based Research Scheme · HK$2.506M
See the Publications page for the full publication record and the CV for additional project and collaboration details.