Research

A very large part of the energy that you use today comes from the geology beneath your feet. This includes conventional oil, gas and coal, but also cleaner sources of energy such as geothermal energy, carbon capture and storage for conventional reservoirs, geological hydrogen, and even for aeolians, a characterization of the subsurface is necessary for engineering. The subsurface is also a vital reservoir of drinkable water, and of brines rich in critical minerals. Hence, a proper characterization of geology in the subsurface is essential to human activites.

Automatic core description is a big research theme in our lab, as previous studies have shown that humans are inconsistent in descriptive tasks. This is important, because the facies defined during core descriptions are used to interpret environmental changes, and also petrophysical classes for subsurface energy. Our work to date includes supervised classification and object detection, semi-supervised learning for core deformation, and rethinking core description as a regression task rather than a classification. We are now working towards foundation models for generalizable downstream applications.

Generative AI is more often associated with LLMs and vision-language models, less with subsurface geology. However, in a series of progressively more complex projects, we show that generative AI can help geologists interpret the subsurface more accurately than if they rely solely on log data. We first explore how formation microscanner (FMS) data can serve as a prior to a generative model (Pix2PixHD) and demonstrated that geologists were more accurate when describing Dunham texture on synthetic cores than on raw FMS data. More recently, we continued this work by integrating conventional logs and FMS data in a new architecture for the generator model of Pix2PixHD, achieving even more realistic results.

Seismic data is a key source of information for subsurface geology, but its processing and interpretation can be time and resource intensive, as well as subjective. We are working on novel mathematical loss functions for neural network to improve seismic processing such as missing trace reconstruction or inversions. These AI techniques are orders of magnitude faster than conventional inversion or reconstruction techniques. For automatic interpretation of seismic stratigraphy, we overcome the issue of lack of labelled training data by generating realistic synthetic data. We have experimented with several techniques to achieve this, including generating simple seismic termination and full geological cross sections with forward stratigraphic modelling. We then use forward modelling of a seismic impendance response of the synthetic geology to obtain our training data. We demonstrate on actual data that the trained models perform well.

Both statistical and physical simulations are important to create subsurface Earth models honoring sparse well and seismic ground truthing. Geostatistics and forward stratigraphic modelling (e.g., CRONOS, DionisosFlow, CarboCAT) are two well-known approaches to produce Earth models, and these can then be used in flow simulations. Increasingly, deep learning is important to create fast, calibrated Earth models. We have a long history of working with simulations for the subsurface.

Much of our research explores how different modelling philosophies can be combined to understand sedimentary systems and reservoirs. Geostatistical methods, for example, are fast and honour well data perfectly, but because they are not anchored in physics they can produce unrealistic reservoir geometries; we have developed geostatistical tools such as pluri-Gaussian simulations to capture cyclical sedimentary successions. Forward stratigraphic models offer the complementary trade-off: they may not reproduce well data exactly, but they always yield physically realistic geometries. This makes them well suited to questions about depositional controls — from Cretaceous carbonates in Oman, to high net-to-gross deposits in the Gulf of Mexico, to Jurassic organic shales in the Middle East. A third strand of our work turns to dynamic flow simulation, for instance predicting water breakthrough in oil and gas reservoirs, including NURBS-based techniques for modelling channel deposits in Saudi Arabia. Together these approaches illustrate the range of numerical methods we use, and how their contrasting strengths can be brought to bear on a given problem.

One of the drawback of physics-driven simulation is that they can be very slow to compute. This is especially problematic for geological simulations: because the inital conditions of the model are usually poorly constrained, an ensemble of simulations needs to be run, each costing a significant amont of computation. We use deep learning as a surrogate for these simulators: training convolutional and recurrent neural networks on libraries of forward model runs, then using the trained surrogates to explore parameter space orders of magnitude faster. This enables rapid sensitivity analysis, quick identification of most compatible simulation parameters, and orders of magnitude faster solving of complex simulations. We couple this with data assimilation techniques to improve the calibration of models, thus ensuring that our ensembles are composed of mostly calibrated runs.

Numerical weather prediction and climate modelling have long been dominated by physics-based models that are computationally expensive and imperfect in their parameterisations. AI offers a complementary approach: learn the statistical relationships embedded in decades of reanalysis and observational data to produce fast, accurate forecasts. Although our team is focusing mostly on geological simulation, we have estabished some collaboration with atmospheric scientists and have competed - and won - in weather forecast hackatons.

Tracking environmental changes requires a mix of in-situ measurments, and remote sensing with drones, remotely operated vehicles, and satellite imagery. We use artificial intelligence on this data to automated it's interpretation, and improve the way we can monitor our planet and other planets in the solar system.

Computer vision has a major role to play in the study of the oceans. For instance, we worked on automating the classification of deep-sea biota using advanced deep learning models trained on remotely operated vehicle (ROV) image datasets. We developed a dataset with nearly 4,000 images across 33 organism classes, achieving classification accuracies up to 65% using Inception-ResNet architecture. This work aims to improve biodiversity assessments and conservation efforts in remote marine environments.

We also have spent considerable efforts developping tools to automate the recovery and correction of marine scenes in satellite images, notably to monitor coral reefs. The issue is that most operator's corrections (e.g. NASA, ESA) rely on dark vegetation or other continental features to auto-correct the images. When one is trying to build satellite time-series from Islands or coral reefs, it is difficult because the quality of the scene advertised by the operator cannot be always trusted, and the corrections can be inconsistent. Machine learning can overcome these issues.

One of the most exciting things about remote sensed data is that it is not limited to Earth. In fact, we have a growing collection of data from other planets in the solar system, notably Mars and the moon. Mars offers some exciting challenges, with multimodal data comprising satellite imagery, high-resolution imagery, hyperspectral data, and in-situ rover data.

We worked extensively on the Compact Reconnaisance Imaging Spectrometer for Mars (CRISM) instrument, a satellite capturing data across Mars in a number of different wavelength. This hyperspectral data is powerful, allowing us to make inference about the surface composition of the planet. However, the data quality of the instrument was declining over time, and became unsable. We demonstrated that AI could be used to denoise the signal, and to confidently identify minerals with a calibrated confidence. This approach can be extended in future to other bodies in the solar system.

Computer vision is a versatile tool, and some of our work expends beyond the confined of Earth Sciences. For instance, we develop novel ways to train machine learning model using knowledge distillation, or look at vision-language models and leverage semi-supervised learning to uncover the meaning of ancient written characters. And an image being an image, we can use our toolkit to solve problems in medical imaging, and beyond.

The point is that research is interdisciplinary in this day and age, and we welcome collaboration with fundamental AI teams, or applied group. This is why our research group initiated and lead the "Scale-Invariant Spatial Intelligence Hub [SI2-HUB]: Cross-Scale, Multimodal and Physics Aware Computer Vision Across Disciplines" team at the Huang Collaboration and Innovation Complex at OSU.