Can ML/AI make subsurface simulations faster without compromising physics?
Predicting how fluids and chemical species move through the Earth’s subsurface is essential for safe energy storage, carbon sequestration, groundwater management, and environmental protection. However, modelling complex geological systems over many possible scenarios can require substantial computational time and resources.
Scientific machine learning combines artificial intelligence, physical laws, and numerical modelling to create faster predictive tools while retaining the physical constraints that make simulations scientifically meaningful. These projects explore two complementary approaches to accelerating subsurface modelling and rigorously evaluate their accuracy, reliability, generalisability, and computational efficiency.
PROJECT 01. Physics-Informed Neural Networks
Combine data-driven learning with the laws of physics.
Physics-Informed Neural Networks (PINNs) embed governing equations, boundary conditions, and conservation laws directly into neural-network training. Rather than relying only on data, they encourage models to produce predictions that are consistent with known physical processes.
This project investigates whether PINNs can accurately represent coupled subsurface flow and transport processes, particularly where observations are sparse or uncertain. It also compares different approaches for enforcing physical constraints, including penalty-based loss functions and hard constraints, and examines whether unknown material properties can be inferred from limited measurements.
Research directions:
- How accurately can PINNs capture subsurface physics?
- Do hard constraints enforce physics more effectively than penalty-based methods?
- Can PINNs recover unknown geological or transport properties from sparse data?
- How robust are PINN predictions under uncertainty in data and model parameters?
Explore:
Physics-informed learning, partial differential equations, constraints, inverse modelling, and uncertainty quantification.
PROJECT 02. Neural Operators
Learn the patterns. Predict new scenarios.
Neural operators learn mappings between model inputs and solution fields, allowing rapid predictions across different initial conditions, boundary conditions, source terms, and geological settings. Once trained, they can act as efficient surrogate models for conventional numerical simulators.
This project assesses how well neural operators can reproduce subsurface flow and transport simulations while generalising to scenarios not included in their training data. The work will benchmark predictive accuracy, physical consistency, computational cost, and reliability against established numerical modelling approaches.
Research directions:
- How accurately can neural operators predict subsurface flow and transport?
- How well do they generalise to unseen geological settings and operating conditions?
- How much computational time can they save compared with conventional numerical simulations?
- When do fast surrogate predictions remain reliable enough for scientific or engineering decision-making?
Explore:
Surrogate modelling, operator learning, generalisation, benchmarking, and high-performance scientific computing.