Hybrid Bayesian Process-Informed Neural Network Models for Spatio-Temporal Data

Music City Center 

Process (physics)-informed neural models have become ubiquitous across many areas of science in recent years due to the value of regularizing neural-networks based on an underlying partial differential equation physical constraint. This talk discusses a generalization in which mechanistic information about the process can be incorporated via a Bayesian hierarchical approach. In many scientific applications where there is substantial a priori process knowledge, incorporating this information can improve model performance and efficiency. The notion of including process knowledge in data-driven models for spatio-temporal data is not new (e.g., data assimilation, physical-statistical modeling, etc.), and considering hybrid statistical/neural approaches can provide more realistic modeling of complex processes while quantifying uncertainty. This talk presents a brief overview of neural and statistical approaches and presents a unifying hierarchical modeling structure that can accommodate flexible mechanistically informed neural or statistical models for spatio-temporal dynamic processes.

spatio-temporal

dynamics

large-language model

attention mechanism

stochastic antecedent