Parameter Nonidentifiability

Data-Driven Model Order Reduction, Parameter Estimation, and Prediction

  • Discovered reduced sets of efective parameters with which to describe dynamic models.
  • Confirmed equivalence of data-driven parameters to analytical expressions proposed by experts.
  • Used reduced representations to predict unseen system behaviors and parameter values.
  • Implemented conformal autoencoders to disentangle effective from redundant parameters.

More Details

Machine Learning for Materials Science

Parameterizing Structure-Process-Property Relations in Data

  • Constructed machine learning models of zeolite catalyst properties from small data set.
  • Used models to propose new synthesis conditions that improved catalytic performance.
  • Co-developed an active learning scheme for exploration of high-entropy alloy lattice space.
  • Including dimensionality reduction and comparing performance to composition-based approach.

Data-Driven Optimization

Manifold Learning for Accelerated Solution of High-Dimensional Problems

  • Analyzed trajectories of simulated annealing as time series from Langevin dynamics.
  • Combined diffusion maps with SDE parameter inference to inform larger, coarse-grained steps.
  • Achieved convergence with fewer objective evaluations than standard simulated annealing.
  • Extending approach to Bayesian optimization on reduced space of diffusion coordinates.