Research

Goal: reduce the cost of high-fidelity prediction, inference, and optimization through information fusion.

We build multi-fidelity methods for prediction, inference, and optimization in data- and compute-limited settings. A recurring goal in this work is to use partial, lower-cost, or imperfect information without reducing the problem to a brittle surrogate or a fixed model hierarchy. Instead, we develop statistical estimators and learning procedures that fuse information across models, data sources, discretizations, or operating conditions in a principled way.

This line of work includes foundational variance-reduction methods based on approximate control variates, Bayesian network formulations for multi-fidelity learning, transfer across dissimilar parameterizations, and more recent work on automated exploration-exploitation strategies for estimator construction. The broader aim is to make complex predictive pipelines more adaptive, data-efficient, and useful for decision-making, while preserving statistical structure and giving clear control over computational cost.

These methods are deployed in software and workflows used in broader scientific computing ecosystems, including work connected to Dakota, PyApprox, and NASA's MXMCPy.

Selected papers:

• A generalized approximate control variate framework for multifidelity uncertainty quantification (2020)
• MFNets: multifidelity data-driven networks for Bayesian learning and prediction (2020)
• Ensemble approximate control variate estimators: applications to multi-fidelity importance sampling (2022)
• Adaptive experimental design for multi-fidelity surrogate modeling of multi-disciplinary systems (2022)
• Multifidelity uncertainty quantification with models based on dissimilar parameters (2023)
• Covariance expressions for multi-fidelity sampling with multi-output, multi-statistic estimators (2024)
• BISTRO: a bi-fidelity stochastic gradient framework using trust-regions for optimization under uncertainty (2025)
• Optimally balancing exploration and exploitation to automate multi-fidelity statistical estimation (2026)
• Accelerating Bayesian inference via multifidelity transport map coupling (2026)

Goal: fuse models and data for inference and learning from imperfect information.

We develop Bayesian methods for updating beliefs and learning models from incomplete, noisy, and indirect data. Much of this work focuses on dynamical systems, where state evolves over time, observations arrive sequentially, or learning must be incremental. In these settings, the central challenge is not only to make predictions, but to manage parameter uncertainty, model error, measurement noise, and partial observability in a coherent way.

Our contributions in this area include Bayesian system identification objectives that explicitly account for parameter, model, and measurement uncertainty; structure-aware identification of Hamiltonian systems; learning for partially observed forced systems; sequential parameter-state estimation; and theory for understanding how approximation errors propagate through Bayesian sequential learning. This perspective also extends naturally to uncertainty-aware learning problems in sensing and perception, where predictive confidence matters for downstream decisions.

Across these problems, the aim is to build inference procedures that can update reliably over time, remain interpretable, and support downstream prediction or control rather than simply fit data in a static batch setting.

Selected papers:

• Bayesian system ID: optimal management of parameter, model, and measurement uncertainty (2021)
• Bayesian identification of Hamiltonian dynamics from symplectic data (2020)
• A new objective for identification of partially observed linear time-invariant dynamical systems from input-output data (2021)
• Bayesian inference for time delay systems with application to connected automated vehicles (2021)
• Likelihood-based generalization of Markov parameter estimation and multiple shooting objectives in system identification (2024)
• Predicting uncertainty in vision-based satellite pose estimation using deep evidential regression (2025)
• A global Lipschitz stability perspective for understanding approximate approaches in Bayesian sequential learning (2025)
• Sequential Bayesian parameter-state estimation in dynamical systems with noisy and incomplete observations via a variational framework (2026)

Goal: mitigate the curse of dimensionality with error-controlled scalable computation.

We develop tensor and low-rank methods that discover and exploit structure for scalable computation. The motivation is simple: many scientific computing and learning problems are high-dimensional enough that standard representations and solvers become infeasible. Our work studies how tensor-network structure can be used not only for compression, but also for learning, differentiation, uncertainty quantification, structure learning, and the solution of large-scale scientific models.

This thread includes foundational work on continuous tensor-train representations and functional tensor approximation, tensor-network structure learning, reverse-mode differentiation for arbitrary tensor-network contraction sequences, incremental tensor decompositions for streaming data, and more recent tensor-based solvers for scientific transport equations. A key theme is that these methods are not treated as purely black-box reduced models. We seek algorithms with clear numerical structure, controllable error, and the ability to support computation in settings where conventional methods fail.

Recent work has pushed these ideas into quantum dynamics, video and behavioral learning, scientific data compression, and high-dimensional PDEs such as thermal radiation transport. Across these settings, the emphasis remains on scalable scientific computation.

Selected papers:

• A continuous analogue of the tensor-train decomposition (2019)
• Gradient-based optimization for regression in the functional tensor-train format (2018)
• Reverse-mode differentiation in arbitrary tensor network format: with application to supervised learning (2022)
• An incremental tensor train decomposition algorithm (2024)
• Low-rank tensor-network encodings for video-to-action behavioral cloning (2024)
• Functional tensor-train Chebyshev method for multidimensional quantum dynamics simulations (2022)
• Thermal radiation transport with tensor trains (2025)
• Tensor networks for block adaptive thermal radiation transport (2026 preprint)

The research is methods-driven, but it is tested in serious scientific and engineering settings rather than only on stylized benchmarks. These applications are where we evaluate whether an algorithm is actually useful under real constraints on physics, data, uncertainty, and computation.

Current recurring domains include:

• Electric propulsion: Hall thruster model calibration, optimal experimental design, kinetic simulation, reduced-order modeling, and uncertainty quantification across facility conditions and model components.
• Fluid mechanics: multi-fidelity prediction and inference, including aerodynamic and turbulence-related settings where high-fidelity models are expensive and lower-fidelity approximations must be used carefully.
• Radiation and plasma transport: high-dimensional transport solvers, plasma simulation acceleration, and uncertainty-aware modeling for complex multiphysics systems.
• Aerospace sensing, estimation, and control: uncertainty-aware vision, connected and autonomous vehicle estimation, and dynamical-system learning where sequential updates and partial observability are central.

Representative papers:

• GPU-accelerated kinetic Hall thruster simulations in WarpX (2025)
• Hall thruster model improvement by multidisciplinary uncertainty quantification (2024)
• Dynamic mode decomposition for particle-in-cell simulations of a Hall thruster and plume (2024)
• Predicting uncertainty in vision-based satellite pose estimation using deep evidential regression (2025)
• Multigroup thermal radiation transport with tensor trains (2026)

Earlier work in stochastic optimal control, nonlinear filtering, experimental design, system identification, and multi-agent assignment helped establish several ideas that inform the current program, including structured high-dimensional computation, sequential inference, adaptive information acquisition, and decision-making under uncertainty.

• High-dimensional stochastic optimal control using continuous tensor decompositions (2018)
• Low-rank tensor integration for Gaussian filtering of continuous-time nonlinear systems (2017)
• Integrated variance experimental design for Gaussian process regression (2016)
• System identification via CUR-factored Hankel approximation (2018)
• Dynamic multi-agent assignment via discrete optimal transport (2022)

These threads continue to inform the larger program in scalable inference, uncertainty-aware computation, and decision-making.

For complete bibliographic information, please see the publications page.