Environmental Science and Engineering Seminar

Parameter estimation and uncertainty quantification (UQ) are critical for managing complex natural and engineered systems, yet traditional methods are limited by high-dimensional models involving millions of parameters and computationally expensive simulations. This talk highlights how scientific machine learning bridges physics-based forward and inverse models with data-driven approaches to overcome these challenges. Reduced-order models (ROMs), accelerated by physics-informed neural networks and neural operators, enable efficient inversion and UQ by approximating governing equations while ensuring generalization. Deep generative models, such as variational autoencoder, generative adversarial networks, and diffusion models, further streamline stochastic inverse modeling by encoding high-dimensional, non-Gaussian priors into low-dimensional manifolds, while multi-fidelity forward modeling frameworks can integrate high-cost simulations with cheaper ROMs to balance accuracy and computational cost. Applications in river and coastal dynamics and subsurface system characterization demonstrate the accuracy and scalability of these methods in integrating multi-modal, heterogeneous remote sensing and geophysics data sets, quantifying uncertainties, and supporting decision-making.