Utilizing Physics Informed Neural Networks for Disrupt Signal Dynamics
Abstract
Physics informed neural networks (PINNs) have a long history of use in a variety of applications from engineering to the physical sciences such as biology or physics. PINNs allow for the incorporation of a priori understanding of a process’ structure into the neural network learning process. This allows for such models to incorporate the processes behind the data in a way that aligns with prior expectations for controlling dynamics. We attempt to leverage the physics informed neural network structure toward the evaluation of disruptions to classical dynamical models. We do so by combining the incorporation of elements of ordinary differential equations into our loss function with sigmoidal gating to control the balance of penalties for deviations from the data with the penalties for deviations from the expected signal structure. This allows for the identification of the signal structure as well as the bounds of the disruption. As a case study, we consider the SPDR S&P 500 ETF Trust stock value from 2019 to 2021. This stock expresses an approximately exponential growth curve which is interrupted by the 2020 COVID-19 induced market crash. We train a physics informed neural network on this data and identify the underlying exponential trend as well as the bounds of the disruption. This serves as a proof of concept for this application of physics informed neural networks. We aim for this new architecture to serve as an alternative to other PINN variations robust to discontinuous or noisy signals that combines signal identification with the identification of the disruption bounds and steepness of onset.
Start Time
15-4-2026 9:00 AM
End Time
15-4-2026 12:00 PM
Room Number
Culp Ballroom 316
Poster Number
48
Presentation Type
Poster
Presentation Subtype
Posters - Competitive
Presentation Category
Science, Technology, and Engineering
Student Type
Graduate and Professional Degree Students, Residents, Fellows
Faculty Mentor
Jeff Knisley
Utilizing Physics Informed Neural Networks for Disrupt Signal Dynamics
Culp Ballroom 316
Physics informed neural networks (PINNs) have a long history of use in a variety of applications from engineering to the physical sciences such as biology or physics. PINNs allow for the incorporation of a priori understanding of a process’ structure into the neural network learning process. This allows for such models to incorporate the processes behind the data in a way that aligns with prior expectations for controlling dynamics. We attempt to leverage the physics informed neural network structure toward the evaluation of disruptions to classical dynamical models. We do so by combining the incorporation of elements of ordinary differential equations into our loss function with sigmoidal gating to control the balance of penalties for deviations from the data with the penalties for deviations from the expected signal structure. This allows for the identification of the signal structure as well as the bounds of the disruption. As a case study, we consider the SPDR S&P 500 ETF Trust stock value from 2019 to 2021. This stock expresses an approximately exponential growth curve which is interrupted by the 2020 COVID-19 induced market crash. We train a physics informed neural network on this data and identify the underlying exponential trend as well as the bounds of the disruption. This serves as a proof of concept for this application of physics informed neural networks. We aim for this new architecture to serve as an alternative to other PINN variations robust to discontinuous or noisy signals that combines signal identification with the identification of the disruption bounds and steepness of onset.
