Dr. Arash Rabbani

Profile

I lead the Data Flow Lab at the University of Leeds, where we develop AI and machine vision solutions for characterizing micro-structured materials and biological tissues. Our research addresses fundamental challenges in computational modelling of fluid-solid interactions, a phenomenon critical to applications ranging from sustainable energy systems and water resources to pharmaceutical manufacturing and biomedical engineering.

The lab specializes in computational methods that integrate advanced imaging, deep learning, and physics-based modelling to extract meaningful structure-property relationships from complex three-dimensional microstructures. We can 3D print our materials and test our hypotheses! We work across scales, from nanometre-resolution characterization of separation membranes to millimetre-scale analysis of cardiac tissue, combining data-driven approaches with domain knowledge to accelerate materials discovery and diagnostic capabilities.

Our interdisciplinary research has contributed to projects funded by UKRI, the US Department of Energy, and NSERC Canada, with applications in naturally occurring materials, engineered materials, and bio-materials. The lab maintains active collaborations with industry partners including Johnson & Johnson Innovative Medicine and AstraZeneca, translating fundamental insights into practical engineering solutions for pharmaceutical and healthcare applications. We welcome researchers interested in computational materials science, bio-material image analysis, and data-driven fluid-solid interaction modelling.

Research interests

Below, you can find selected research case studies by our lab. Please feel free to get in touch via a.rabbani@leeds.ac.uk if any of following resonates with you.

DeePore: Rapid Characterization of Porous Materials Using Deep Learning

Traditional characterization of porous materials through physical simulations on pore network models can require substantial computational resources and time. DeePore addresses this challenge by using convolutional neural networks to rapidly estimate 30 different physical properties of porous materials directly from binarized micro-tomography images. The workflow generates 17,700 semi-real three-dimensional microstructures from naturally occurring porous textures and trains a feed-forward CNN to predict morphological, hydraulic, electrical, and mechanical characteristics in fractions of a second. With an average coefficient of determination of 0.885 across testing samples, the method achieves prediction speeds orders of magnitude faster than traditional simulation approaches. Published in Advances in Water Resources, this workflow has been widely adopted by the computational materials community and has become a benchmark dataset for developing new machine learning approaches in porous media research. (Link to the paper)

DeepAngle: Automated Contact Angle Measurement in Porous Media

Measuring contact angles in three-dimensional tomography images of porous materials presents significant challenges due to the need to measure angles within surfaces perpendicular to interface planes in discretized voxel space. DeepAngle employs machine learning to determine contact angles of different fluid phases directly from tomographic images, bypassing computationally intensive surface vectorization approaches. When tested on both synthetic and realistic images against direct measurement techniques, DeepAngle improved accuracy by 5 to 16% while reducing computational costs by a factor of 20. Published in Geoenergy Science and Engineering, this rapid method proves especially valuable for processing large tomography datasets and time-resolved images where conventional approaches become prohibitively expensive, with applications in carbon storage, environmental contaminant flows, and any multiphase flow process where interfacial wettability governs fluid behavior. (Link to the paper)

Hybrid Pore Network and Lattice-Boltzmann Modeling Accelerated by Machine Learning

Direct simulation of fluid flow using Lattice Boltzmann Methods provides high accuracy but demands substantial computational resources. This research developed a hybrid workflow coupling pore network modeling with Lattice Boltzmann simulations, where watershed segmentation extracts pore networks from three-dimensional rock images and LBM calculates throat permeabilities to replace simplified cylindrical formulations based on the Hagen-Poiseuille equation. To minimize computational costs, the workflow simulates steady-state incompressible flow through 9,333 different throat images using LBM and trains an artificial neural network to predict throat permeabilities from cross-sectional images Hybrid pore network and lattice-Boltzmann permeability modelling accelerated by machine learning. Published in Advances in Water Resources, this approach dramatically reduces computation time while maintaining prediction accuracy, enabling practical permeability calculations for large-scale porous media samples that would otherwise require prohibitive computational resources. (Link to the paper)

Superpixel-Based Pore Network Extraction

Traditional pore network extraction methods based on watershed segmentation or medial axis approaches can struggle with computational scalability when processing large tomography images. This work introduces superpixels, a classic image segmentation technique, as the foundation for pore network extraction from geological tomography images. The method analyzes three-dimensional microstructures to construct pore network models representative of hydraulic and electrical behavior, introducing an effective throat radius correction factor to compensate for over-segmentation effects. Despite differences in network morphology compared to watershed methods, superpixel networks predict macroscopic properties with equivalent accuracy while proving exponentially faster and linearly more memory intensive than watershed approaches Superpixels pore network extraction for geological tomography images. Published in Advances in Water Resources, this computational advantage makes superpixel extraction particularly valuable for analyzing large-scale tomography datasets where memory constraints would otherwise limit analysis capabilities. (Link to the paper)

Data-Science-Based Reconstruction of Three-Dimensional Membrane Pore Structure

Conventional two-dimensional scanning electron microscopy provides rapid qualitative evaluation of membrane pore structure but lacks information about three-dimensional spatial arrangement and connectivity crucial for understanding membrane performance. While experimental three-dimensional reconstruction via tomography exists, it remains complex, expensive, and not easily accessible. This collaborative research employs data science tools to reconstruct the full three-dimensional structure of a membrane from a single two-dimensional image extracted from tomographic datasets. When comparing reconstructed and experimental three-dimensional structures, important properties including mean pore radius, mean throat radius, coordination number, and tortuosity differed by less than 15% Data-science-based reconstruction of 3-D membrane pore structure using a single 2-D micrograph. Published in Journal of Membrane Science, this methodology enables rapid membrane characterization without expensive tomography equipment, accelerating the design and optimization of separation membranes for water treatment, gas separation, and other filtration applications. (Link to the paper)

Structure-Property Relationships in Fibrous Meniscal Tissue Through Image-Based Augmentation

Understanding mechanical function in fibrous biological tissues requires characterizing structure-property relationships across many samples, yet obtaining sufficient experimental tissue data remains prohibitively expensive. This work introduces an adaptive three-dimensional image synthesis technique that creates variational realizations of fibrous meniscal tissue microstructures with controlled deviations in parameters including porosity, pore size, and specific surface area. The unbiased reconstructed samples matched morphological and hydraulic properties of original tissues with relative errors generally below 10%. Analysis of 1,500 synthesized geometries revealed relationships between microstructural features, hydraulic permeability, and mechanical properties, yielding empirical correlations that predict longitudinal and transverse hydraulic permeability as functions of porosity with R² values of 0.98 and 0.97 respectively Structure–property relationships in fibrous meniscal tissue through image-based augmentation. Published in Philosophical Transactions of the Royal Society A, this augmentation framework provides a pathway for understanding injury mechanisms and designing biomimetic scaffolds for tissue engineering applications throughout the musculoskeletal system. (Link to the paper)

Qualifications

• Ph.D. University of Manchester

Student education

I am the Academic Integrity Lead of the School of Computing and presently teach two modules of Artificial Intelligence and Machine learning to UG and MSc students.

Current postgraduate researchers

• Ali Sadeghkhani

Projects

<li><a href="//phd.leeds.ac.uk/project/2059-generative-ai-for-3d-functional-material-discovery">Generative AI for 3D Functional Material Discovery</a></li>