Bio

I am a Distinguished Member of Technical Staff at Sandia National Laboratories in Livermore, California (currently remote-working from Queens, New York).

My research evolves around uncertainty quantification (UQ), statistical learning and predictability analysis of physical and computational models. I have developed and applied methods for model reduction, UQ and data assimilation, targeting fundamental challenges such as structural errors, intrinsic stochasticity, high-dimensionality, limited data, discontinuities, and rare events, with a range of applications including climate modeling, chemical kinetics, turbulent combustion, fusion science, hardware architecture simulation.

Interests

  • Uncertainty quantification
  • Machine learning
  • Statistical modeling
  • Bayesian inference

Education

  • Ph.D. in Applied and Interdisciplinary Math, 2007

    University of Michigan, Ann Arbor

  • B.S. in Applied Math and Applied Physics, 2002

    Moscow Institute of Physics and Technology

Current Projects

(and my roles in them)

  • E3SM: A state-of-the-science Earth system model development and simulation project to investigate energy-relevant science using code optimized for DOE’s advanced computers.
    • Uncertainty quantification (UQ) lead of E3SM land modeling (ELM),
    • Development and deployment of UQ algorithms and tools for ELM.

  • FASTMath: Scalable mathematical algorithms and software tools for reliable simulation of complex physical phenomena and collaborates with application scientists to ensure the usefulness and applicability of FASTMath technologies.
    • Development of UQTk,
    • Advanced method for model structural error estimation,
    • Outreach and collaboration with physical scientists in support of UQ tools.

  • ECC: Exascale Catalytic Chemistry: predictive models for gas/solid heterogeneous catalytic systems.
    • Development of MPNN,
    • Advanced integration methods for partition function computation,
    • Uncertainty quantification of kinetic Monte Carlo simulations.

  • UQPANN: Visualizing and Quantifying Uncertainty of Physics-aware Neural Networks.
    • Partnership between SciDAC institutes FASTMath and RAPIDS,
    • Develop scientific visualization techniques to understand uncertainties in neural network (NN) predictions, and gain insights into the impact of physics constraints on the shape of NN loss landscape.

  • QBO: Improve the Quasi-biennial oscillation through surrogate-accelerated parameter optimization and vertical grid modification.
    • Surrogate-enabled calibration and uncertainty quantification of QBO.

  • ThermChem: Develop an integrated computational model linking the materials physics scale with the component-level scale to simulate the thermomechanical response of the full first wall/blanket structures during fusion reactor operation.
    • Uncertainty quantification and propagation in atomistic modeling.

  • FusMatML: Machine Learning Atomistic Modeling for Fusion Materials.
    • Model error estimation and active learning of machine learning interatomic potentials.
    • Development of uncertainty quantification tools within FitSNAP.

  • NNRDS: Neural Networks as Random Dynamical Systems: advanced regularization methods for improving residual NN training and generalization via Neural ODE analogy.
    • Stiffness score penalization of ResNets,
    • Regularization via weight parameterization,
    • Augmenting NN predictions with uncertainty estimation.

Software

  • UQTk: The UQ Toolkit (UQTk) is a collection of C++/Python libraries and tools for the quantification of uncertainty in numerical model predictions.

  • QUiNN: Quantification of Uncertainties in Neural Network (QUiNN) is a python library centered around various probabilistic wrappers over PyTorch modules in order to provide uncertainty estimation in neural network predictions.

  • MPNN: Minima-Preserving Neural Network (MPNN) is a small PyTorch-based library for neural network surrogate construction that preserves first- and second-order information at given minima.

Publications

Quickly discover relevant content by filtering publications.
Polynomial Chaos Surrogate Construction for Random Fields with Parametric Uncertainty. SIAM/ASA Journal on Uncertainty Quantification, 2024.

Active Learning for SNAP Interatomic Potentials via Bayesian Predictive Uncertainty. Computational Materials Science, 2024.

DOI

Ground Heat Flux Reconstruction Using Bayesian Uncertainty Quantification Machinery and Surrogate Modeling. Earth and Space Science, 2024.

DOI

Importance Sampling within Configuration Space Integration for Adsorbate Thermophysical Properties: A Case Study for CH3/Ni(111). Phys. Chem. Chem. Phys., 2024.

DOI

Measuring Stiffness in Residual Neural Networks. RAMSES: Reduced order models; Approximation theory; Machine learning; Surrogates, Emulators and Simulators, 2024.

DOI

Selected Presentations

Spatio-Temporal Surrogate Construction and Calibration of E3SM Land Model

Talk

Reduced-Dimensional Neural Network Surrogate Construction for the E3SM Land Model

Talk

Uncertainty Quantification and Calibration of E3SM Land Model

Poster

Quantifying Uncertainties in Weight-Parameterized Residual Neural Networks

Talk

Contact