Link to Report: Coming Soon
Background :
The problem of bridge scour has typically been studied using empirical equations (like HEC-18 CSU equation). CFD has become a powerful tool for modeling the problem, offering detailed insights into the complex interactions between flow structures and sediment transport. Unlike empirical formulas, CFD enables high-resolution simulation of turbulent flow, vortex formation, and bed shear stress around bridge piers, which are critical drivers of scour. Abdelalim et al. Developed a shear-stress-based CFD model to estimate scour depth around complex bridge supports using ANSYS, while Yu et al. applied a refined wall shear stress model within a Reynolds-Averaged Navier-Stokes (RANS) framework to simulate scour evolution. These advancements highlight CFD’s potential to improve scour prediction accuracy and support risk-informed infrastructure design. However, CFD simulations can be costly and utilizing them in a probabilistic framework that required thousands of runs over the parameter space is computationally prohibitive. ML models can address this shortcoming by expediting simulations.
Machine learning application for predicting bridge scour has been explored very recently. The common ML architectures include artificial neural networks (ANNs), Support Vector Machines (SVMs), Random Forest, Long-term Short-Term Memory (LSTM), and Convolutional Neural Networks (CNNs). Recent studies also include physics-inspired neural networks (PINNs) which incorporate empirical equations for scour in a deep learning framework. However, since these models are trained purely by field data, their application to unseen extreme conditions will be erroneous. CFD simulations of scour for extreme flood conditions can augment the training data by filling this gap.
Robots and autonomous systems have been used to address the problem of scour detection. Particularly close to our approach is the work in, where a low-cost AUV, BlueROV2 (base cost $4,900) was equipped with a side-scanner for scour monitoring and a forward-looking sonar for obstacle avoidance. We build on this concept and recent work by Co-PI Bobadilla on developing algorithms, digital twins, and systems for autonomous systems to monitor partially submerged infrastructure along with strategies for efficient obstacle avoidance and detecting and tracking features on water bodies.
Objectives :
The project goal to develop and test a concept for real-time scour prediction and monitoring will be achieved by pursuing the following research objectives:
RO-1: Assess the performance of a high-fidelity CFD model in scour prediction
RO-2: Integrate a low-cost robotic underwater platform using camera, sonar, and forward-looking sonar for efficient detection and characterization of bridge scour damages, through the development of planning, control, and perception algorithms.
RO-3: Develop and validate an ML model as a fast surrogate for the CFD model.
RO-4: Develop a probabilistic, generalizable, predictive framework for bridge scour using the validated ML model
Scope :
Task 1 – Characterization of common bridge structures and environmental conditions
Task 1.1: Survey the bridge structure designs in the state of Florida and in the U.S. to identify common structural geometric features in bridge piers.
Task 1.2: Identify the range of soil characteristics in the coastal, estuarine, and riverine environments where vulnerable bridges are located.
Task 1.3: In collaboration with FDOT, Miami-Dade County, and Boward County, identify bridges that have been vulnerable to scour. These bridges will be candidates for field data collection.
Task 2 – Development and validation of a high-fidelity CFD model
Develop a high-fidelity CFD model (likely OpenFOAM as it is open source), with the proper flow conditions (collected in Task 1.3), structural design (Task 1.1) and soil characteristics, (Task 1.2), and simulate scour due to regular flow conditions. OpenFOAM has previously been used for scour problem (Figure 1 showing simulations) and PI Tahvildari is working with the program to simulate fluid-structure interaction [17]. The model will use flow boundary conditions measured by deployment of an acoustic doppler current profiled (ADCP) that PI owns, and it will be validated with in-situ measurements of scour patch depth and distribution (Task 3), at a location identified in Task 1.3. The model will then be applied over a wide parameter range including flow velocity and turbulence intensity, water depth and channel slope, pier geometry and alignment relative to the flow, sediment type and size distribution, and bed roughness. The outputs of each simulation will include maximum scour depth, shear stress distribution, and morphological change around the pier. These results will form a labeled dataset for ML training (Task 4).
Figure 1. Sample of OpenFOAM application for bridge scour.
Task 3 – Test operation algorithms of underwater robotic systems and field deployment
The field data collection process will be done after extensive testing of the algorithms in simulation and controlled conditions, in Co-PI Bobadilla’s lab. We have allocated recharge resources for two field deployments including boat rental, captain, and scuba diver collaborators.
Task 3.1: Sensor Integration with underwater robots- The team has several Blue Robotics Remotely Operated Vehicles (BlueROV2) (Figure 2) with imaging sonars and cameras for obstacle avoidance and perception. We will acquire a side-scanner to integrate to the robot for scour damage detection.
Task 3.2: Motion planning and control for BlueROV2- We will propose AI-driven algorithms based on our previous work to create strategies that will cover the required infrastructure. Vision and side scanner information will be used to help state estimation and control.
Task 3.3: Perception algorithms for scour detection- We will investigate two problems, one how to detect scour holes using side-scanner information and how to detect change from a based side-scanner image condition.
Figure 2. Underwater robot BlueROV2
Figure 3. Motion capture tank at FIU and its digital twin will be used to test the BlueROV2 algorithms prior to field deployment. A sample video of the test in the lab: https://www.youtube.com/watch?v=1PQ4jdE06vU
Task 3.4: Field deployment- After testing the three preceding tasks in realistic simulations and controlled environments in our motion capture tank facility and in its digital twin (Figure 3), we will do two test deployments to test the effectiveness of the proposed methodology.
Task 4 – Development of an ML surrogate for the CFD model
To enhance the predictive capability and scalability of bridge scour assessment, this project will develop a ML model trained on synthetic data generated using the high-fidelity CFD simulations (Task 2). This ML model will be a rapid surrogate model to estimate scour depth and distribution across a range of hydrodynamic, soil, and pier geometric conditions. Key input variables will be extracted and normalized to ensure consistent scaling across the dataset. Nondimensional numbers including Froude number, Reynolds number, Pier diameter-to-flow depth ratio, and Shields parameter will be included as input features to improve model generalizability and reduce sensitivity to site-specific conditions. Data augmentation techniques may be applied using the CFD model to expand the training set and simulate rare or extreme conditions. Several ML model architectures will be evaluated, including Random Forests and Gradient Boosting Machines for interpretable, rule-based predictions, Convolutional Neural Networks (CNNs) for capturing nonlinear relationships. The model will be trained using supervised learning, with cross-validation and hyperparameter tuning to optimize performance. Evaluation metrics will include Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R² score. PI Tahvildari has recently developed several ML models for flood prediction and detection using CNN models, and will leverage the experience in this task.
Task 5 – Develop a probabilistic framework for scour using the trained ML model
For any set of given inputs, the ML model developed in Task 4 produces point estimates for scour depth. This deterministic approach lacks information on the inherent uncertainty in the scour that stems from uncertainties in input conditions. To develop a probability distribution function (PDF) of scour depth, we will use Monte Carlo sampling technique to randomly sample from input parameters (listed in Task 2) from their PDFs. After numerous runs (more than 1000), a histogram or kernel density estimate (KDE) of the scour depth will be derived. This output is essentially the PDF of the scour depth which can be used to calculate mean scour depth, standard deviation, intervals with any confidence level (e.g., 90% CI), and the probability of exceedance of a critical scour depth. This probabilistic insight can be the basis of a risk-based design, asset management, and decision-making.
Research Team :
Principal Investigator: Dr. Navid Tahvildari
Co-Principal Investigator: Dr. Leonardo Bobadilla