Fast Computer Model Calibration using Annealed and Transformed Variational Inference 

Abstract: Computer models play a crucial role in numerous scientific and engineering domains. To ensure the accuracy of simulations, it is essential to properly calibrate the input parameters of these models through statistical inference. While Bayesian inference is the standard approach for this task, employing Markov chain Monte Carlo methods often encounters computational hurdles due to the costly evaluation of likelihood functions and slow mixing rates. Although variational inference (VI) can be a fast alternative to traditional Bayesian approaches, VI has limited applicability due to boundary issues and local optima problems. To address these challenges, we propose flexible VI methods based on deep generative models that do not require parametric assumptions on the variational distribution. We embed a surjective transformation in our framework to avoid posterior truncation at the boundary. Additionally, we provide theoretical conditions that guarantee the success of the algorithm. Furthermore, our temperature annealing scheme and fine-tuning can prevent being trapped in local optima through a series of intermediate posteriors and weight adjustment. We apply our method to infectious disease models and a geophysical model, illustrating that the proposed method can provide fast and accurate inference compared to its competitors. 

Forecasting Subseasonal Temperature Extremes with Machine Learning: Skill, Interpretability, and Paths Toward Uncertainty Quantification

Abstract: Subseasonal temperature extremes pose major challenges for energy systems, agriculture, and public health, yet their prediction remains difficult due to weak atmospheric persistence and limited skill of physics-based models beyond 10 days. In this talk, I will present recent work evaluating the ability of machine learning methods to forecast weekly mean temperature and temperature extremes at 1–4 week lead times across five regions of the contiguous United States. We compare a tree-based ensemble method (random forests, RF) and a recurrent neural network approach (ensemble echo state networks, EESN) against persistence and enhanced climatological baselines.

Across all lead times and regions, RF achieves the highest accuracy, particularly for extremes, reducing RMSE by up to 50% relative to climatology. EESN provides modest improvements over persistence at 1-week lead times but performs similarly to climatology at longer lead times. To improve interpretability and model parsimony, we apply a grouped permutation feature importance (PFI) framework to iteratively rank and select correlated predictor groups from an initial pool of over 1,100 atmospheric and land-surface variables. This approach reduces the effective input dimensionality to fewer than 10–75 features per model with no loss of predictive skill, while revealing region- and lead-time–dependent drivers of subseasonal predictability.

While the primary focus of this work is on point prediction skill, I will conclude by discussing ongoing efforts to extend this framework to probabilistic forecasting and uncertainty quantification. In particular, I will outline challenges associated with uncertainty estimation for subseasonal extremes and highlight how conformal prediction methods, combined with machine-learning–based forecasts, may provide a principled pathway toward calibrated, distribution-free uncertainty estimates for environmental forecasting applications.

Modeling Dependence in Ecological Data: From Mosquitoes to Human Migration 

Abstract: Ecological data are inherently dependent due to spatial and temporal autocorrelation, species interactions, seasonality, and environmental gradients. In this seminar, I will present several statistical methods that account for various forms of dependence that arise in ecological data. Technical details will be limited; emphasis is on the overall ideas. First, we'll briefly look at a dynamic population model for the analysis of abundance data, which often experience a high degree of temporal autocorrelation. The model was fit to mosquito abundance data collected across North America. An additional source of dependence came in the form of temporal preferential sampling (TPS), where data collection was informed by the observed abundance. I accounted for temporal autocorrelation and TPS in a Bayesian hierarchical model (BHM), which was specified mechanistically such that the observed abundance was related to the Gompertz growth function. Due to the mechanistic specification, inference was made for abundance and phenological quantities of interest. Next, we'll explore a BHM used to infer the relationship between population genetic data and the environment. Such data arise in the field of landscape genomics, where there is interest in the functional connectivity and migration patterns of populations. Population genetic data are highly dependent, and I accounted for spatial and temporal autocorrelation via composite likelihoods. The BHM contains a dyadic regression, which is linked to an advection-diffusion differential equation to infer how Bronze Age humans migrated throughout Europe. Finally, I'll present an analysis of spatial compositional data, which often arise in community ecology, organismal composition studies, and ecosystem forecasting. The proposed model accounts for spatial autocorrelation and the compositional nature of the data. In particular, I transformed the compositional data to directional data and modeled it using a novel hyperspheric distribution. This is the first spatial hyperspheric approach to contain fixed and latent spatial random effects, while accounting for the compositional constraints inherent to the data. I applied the model to bioacoustic signal classifications obtained from a machine learning classifier.

Elastic Functional Changepoint Detection of Climate Impacts from Localized Sources 

Abstract: Functional data analysis (FDA) is an important research area, due to its broad applications across many disciplines where functional data is prevalent. An essential component in solving these problems is the registration of points across functional objects. Without proper registration the results are often inferior and difficult to interpret. The current practice in the FDA literature is to treat registration as a pre-processing step, using off-the-shelf alignment procedures, and follow it up with statistical analysis of the resulting data. In contrast, an elastic framework is a more comprehensive approach, where one solves for the registration and statistical inferences in a simultaneous fashion. Our goal is to use a metric with appropriate invariance properties, to form objective functions for alignment and to develop statistical models involving functional data. While these elastic metrics are complicated in general, we have developed a family of square-root transformations that map these metrics into simpler Euclidean metrics, thus enabling more standard statistical procedures. In this talk, we present a changepoint method for functional data which properly accounts for these types of invariances and demonstrate on simulated data and climate data from the eruption of Mt. Pinatubo in the Philippines in June 1991.

Planetary Snow Cover Changes

Abstract: This talk develops a statistical model and methods to quantify trends in binary snow cover observations (snow presence/absence, not depths).  The methods are used to analyze the Earth’s snow cover record, extracted from weekly satellite pictures, during the period 1972-2024. A two state Markov chain with periodic dynamics is developed for the weekly data arising from each land cell studied. Spatial aspects are handled by allowing cell parameters to spatially vary according to a Besag-York-Mollie model. A Bayesian estimation procedure is constructed to estimate model parameters, trends, and their uncertainty margins. Snow cover is seen to be declining in almost twice as many cells as it is advancing.  While Arctic and European snow cover is found to be rapidly receding, other locations, such as Eastern Canada and Japan, are experiencing advancing snow cover.   An interesting seasonal pattern also emerges in the results.

From Data to Action: AI and Hydroinformatics Applications in Environmental Science

Abstract: This talk uncovers how AI and Hydroinformatics are revolutionizing environmental science through cutting-edge information technologies. We highlight the transformative role of AI and web technologies in augmenting data analysis and predictive capabilities. These technologies enable dynamic simulations and real-time decision-making, essential for tackling environmental challenges. We also explore the digital twins, Large Language Models (LLMs) and Metaverse's potential—leveraging Augmented Reality (AR) and Virtual Reality (VR) to create engaging, immersive experiences for environmental data visualization and communication. Current projects demonstrate how these tools are applied to address challenges in hydrology and disasters. Focusing on future directions and vision, we foresee an integrated approach where cutting-edge information technologies transition data into impactful solutions with unparalleled efficiency. This vision emphasizes the development of a collaborative ecosystem that leverages AI, Hydroinformatics, digital twins, and immersive technologies to empower stakeholders across sectors.

Uncertainty Quantification, Agent-Based Models, and Synthetic Populations

Abstract: Agent-based models (ABMs) use rules at the individual (agent) level to simulate a social, ecologic, or social-technical system, producing structured behavior when viewed at an aggregated level. For example, dynamic network simulation models commonly evolve a very large collection of agents interacting over a network that evolves with time. Such models are often used simulate animal populations, epidemics, or transportation, typically producing random trajectories, even when model parameters and initial conditions are identical. This will be a largely conceptual talk showing some approaches for using such models for parameter estimation and quantifying uncertainty in resulting predictions. If time allows, the talk will also deal some with accounting for uncertainty in synthetic populations which are required for ABMs. 

Why a Multi-Representational Approach is Necessary to Facilitate Prediction, Understanding & Scientific Discovery

Abstract: This talk examines the question “How can we use Models to facilitate Scientific Discovery?”, as opposed to the manner in which models are normally used, which is to generate “Predictions” or to aid (we hope) in the development of “Understanding”. As a starting point, we recognize that the key to the analysis of any data, or development of any model, is the selection of an appropriate “representational system”. The problem, however, is that the representation we chose then completely determine the “questions we can ask”, the nature of “analyses and inferences we can perform”, and the “answers we can obtain”. Consequently, any model we develop using that representational approach may be highly suitable for learning certain things about a system, but completely unsuitable for learning other (possibly equally important) things.

To explore how different representational frameworks can affect what we learn from data, we present a case study in which three complementary models approaches (one physics/process-based and two based in machine-learning methods) are simultaneously used to develop an improved understanding of how catchment-scale hydrological processes vary across the diverse hydro-geo-climatology of Chile.  Based (in part) on these results, we argue that if our goal truly is “Scientific Discovery”, then we would be well served to adopt a “general multi-representational framework” when analyzing data, so that the models we build have the possibility of converging towards accurate and comprehensive representations of the underlying data generating process.  Our view is that such an approach is key to the use of Models in support of improved Prediction, Understanding and Discovery in the Earth & Environmental Sciences.