Register now After registration you will be able to apply for this opportunity online.
Modeling GPR data via physics-informed Neural Networks
This master thesis explores the use of Neural Networks (NNs) to model Ground Penetrating Radar (GPR) data. To this end, Physics-Informed Neural Networks (PINNs) can be used to incorporate electromagnetic wave propagation into the NN training in order to enhance the learning task. The inferred deep learning model can be used for fast simulation of GPR data for decision support tasks within asset management of the railway infrastructure.
One of the most common causes for deterioration of the railway track infrastructure is water intrusion leading to subsurface ballast moisture. This type of damage results in significant costs for mitigation and repair. Early detection of subsurface moisture can result in annual savings in the order of 35 millions per year and about 5~10 year extension of track lifetime.
Ground Penetrating Radar (GPR) is a non-destructive geophysical technology that uses radar pulses to image the subsurface and can potentially detect railway substructure damages and water infiltration. However, collecting GPR data and the associated ballast condition labels is a demanding task, which makes the development of railway health indicators far from easy.
To this end, previous works conducted at our chair explored the use of gprMax (www.gprmax.com), an open source software designed for simulating GPR data, in order to obtain simulated GPR data of railway tracks under controlled varying deterioration conditions. However, the software involves the numerical solution of complex equations, resulting in a slow inference compared to our research needs.
As such, this thesis will explore the use of deep learning models to build a reliable surrogate model for fast simulation of GPR data at inference time. The gprMax software will be used to build the offline dataset of GPR simulated data. Physics-Informed Neural Networks (PINNs) can be used to incorporate the underlying physics laws, i.e., electromagnetic wave propagation, to enhance the learning task. The embedding of physics into the learner has benefits for: better generalization/extrapolation, more efficient learning, and more reliable surrogate models. The neural network, trained on the offline simulated GPR dataset, can thus be used for fast emulation of new GPR data of varying railway conditions.
One of the most common causes for deterioration of the railway track infrastructure is water intrusion leading to subsurface ballast moisture. This type of damage results in significant costs for mitigation and repair. Early detection of subsurface moisture can result in annual savings in the order of 35 millions per year and about 5~10 year extension of track lifetime.
Ground Penetrating Radar (GPR) is a non-destructive geophysical technology that uses radar pulses to image the subsurface and can potentially detect railway substructure damages and water infiltration. However, collecting GPR data and the associated ballast condition labels is a demanding task, which makes the development of railway health indicators far from easy.
To this end, previous works conducted at our chair explored the use of gprMax (www.gprmax.com), an open source software designed for simulating GPR data, in order to obtain simulated GPR data of railway tracks under controlled varying deterioration conditions. However, the software involves the numerical solution of complex equations, resulting in a slow inference compared to our research needs.
As such, this thesis will explore the use of deep learning models to build a reliable surrogate model for fast simulation of GPR data at inference time. The gprMax software will be used to build the offline dataset of GPR simulated data. Physics-Informed Neural Networks (PINNs) can be used to incorporate the underlying physics laws, i.e., electromagnetic wave propagation, to enhance the learning task. The embedding of physics into the learner has benefits for: better generalization/extrapolation, more efficient learning, and more reliable surrogate models. The neural network, trained on the offline simulated GPR dataset, can thus be used for fast emulation of new GPR data of varying railway conditions.
The main goal of this thesis is to build a reliable deep learning surrogate model for fast simulation of GPR data of railway tracks. This will be achieved by completing the following tasks:
- Literature review on GPR technology, data, and simulation.
- Simulation of GPR data through the gprMax software to build the training dataset.
- Implementing the PINN architecture.
- Training and validation (both qualitatively and quantitatively) of the model output.
- Comparison against simple (non physics-informed) NNs, e.g. VAE.
The main goal of this thesis is to build a reliable deep learning surrogate model for fast simulation of GPR data of railway tracks. This will be achieved by completing the following tasks:
- Literature review on GPR technology, data, and simulation.
- Simulation of GPR data through the gprMax software to build the training dataset.
- Implementing the PINN architecture.
- Training and validation (both qualitatively and quantitatively) of the model output.
- Comparison against simple (non physics-informed) NNs, e.g. VAE.
Desired competencies: Confident with python, good knowledge of deep learning, and some prior experience with at least one deep learning framework (e.g., Tensorflow/Keras, Pytorch, etc...). Ideally, although not strictly necessary, good physics knowledge.
If interested, please send an email to:
- Giacomo Arcieri - giacomo.arcieri@ibk.baug.ethz.ch
- Dr. Marcus Haywood-Alexander - marcus.haywood-alexander@ibk.baug.ethz.ch
- Prof. Dr. Eleni Chatzi - chatzi@ibk.baug.ethz.ch
Desired competencies: Confident with python, good knowledge of deep learning, and some prior experience with at least one deep learning framework (e.g., Tensorflow/Keras, Pytorch, etc...). Ideally, although not strictly necessary, good physics knowledge.
MasterThesis_GPR-PINN.pdf
