Session: Research Posters

Paper Number: 172100

Machine Learning-Augmented Finite Volume Modeling and Inverse Optimization of Velocity and Pressure Distribution of Bentonite Slurry in Slurry Shield Tunneling Applications 

A vast majority of fluids encountered in engineering and natural system deviates from Newtonian behavior due to their complex microstructure and internal interaction. These non-Newtonian fluids exhibit rheological responses that are highly dependent on applied deformation history encompassing shear thinning, shear thickening and yield stress. Bentonite slurries primarily composed of sodium bentonite, polymer and water illustrates such behavior. Its tunable rheology, structural stability and ease of formulation contribute to its widespread use in shield tunneling. Shield tunneling are utilized across the globe to construct highway, subway and other utilities. Bentonite slurry plays a critical role in maintaining excavation face pressure by forming an impermeable filter cake when infiltrating into soil.  Accurate modeling and real-time control of bentonite slurry flow behavior are essential for optimizing slurry shield tunneling operations, where flow characteristics directly impact excavation stability and system performance. This research presents an integrated computational framework that couples Finite Volume Method (FVM)-based simulations with supervised Machine Learning (ML) models for both forward prediction and inverse design of flow parameters in non-Newtonian bentonite slurry transport.

A total of 48 high-fidelity 2D laminar flow simulations were performed in ANSYS Fluent using a pipe geometry 30 cm in length, with four pipe diameters (0.25 cm, 0.5 cm, 0.75 cm, and 1 cm), four inlet velocities (0.05, 0.1, 0.2, and 0.3 m/s), and three carboxymethyl cellulose (CMC) concentrations (0, 0.1, and 0.3%) to capture the rheological variations governed by the Herschel–Bulkley model. The simulations were validated using analytical benchmarks for pressure drop and velocity profiles. This parametric variation yielded a rich dataset, which was used to train several supervised ML models, including Artificial Neural Networks (ANN), Decision Tree Regressors, and Ensemble models.

Model evaluation showed high predictive performance for both pressure drop and centerline velocity, with R² values exceeding 0.98, confirming the reliability of ML surrogates. Correlation heatmaps and feature importance analyses further revealed the relative impact of pipe diameter, inlet velocity, and CMC content on flow behavior. Notably, pipe diameter was the most influential factor in determining pressure gradients, while inlet velocity had a dominant effect on centerline velocity.

Beyond forward modeling, inverse analysis was performed by integrating the trained ML models with optimization algorithms to identify the required input parameters—such as CMC content, diameter, and inlet velocity—that would yield a desired pressure drop or velocity. This surrogate-based inverse modeling enables rapid, real-time estimation of operating conditions, bypassing the computationally expensive iterative CFD-based inverse simulations.

This ML-augmented approach provides a scalable and cost-effective framework for real-time rheological monitoring and process optimization in slurry shield tunneling. It opens new possibilities for predictive control, inverse rheology design, and adaptive automation in underground construction environments.

Presenting Author: Somnath Somadder FIU

Presenting Author Biography: Somnath Somadder is a PhD student in the MME department of FIU. He has expertise in using commercial software such as Abaqus, Ansys, COMSOL and open-source software such as openFoam, openPhase. He has five years’ experience of being a faculty member in Bangladesh. He is the author of several peer-reviewed journal articles, and he has 40 citations. His research interest includes Computational fluid dynamics, Finite element analysis, Machine learning, Multiscale modeling.

Authors:

Somnath Somadder FIU