Manifold Learning-based Surrogate Modeling for Uncertainty Quantification in Amorphous Solids
Dimitris Giovanis – Assistant Research Scientist, Johns Hopkins University
Optimization of material strength, failure tolerance, and formability requires a capacity to predict mechanical response on the continuum or meso-scale, given the uncertainties that exist in the material’s microstructure. However, quantifying parametric (real-world variability and randomness) and/or model-form (imperfect assumptions and idealizations in our models) uncertainties and propagating them across different length-scales presents a grand challenge for materials engineering. This seminar will focus on a novel surrogate modeling strategy to overcome the challenge of modeling plastic deformation in amorphous solids using the Shear Transformation Zone Theory of plasticity, in the presence of parametric uncertainty. This new approach combines surrogate modeling (Gaussian process) with manifold (Grassmannian) learning for the interpolation of reduced-order solutions, that can be used (through inversion of the dimensionality reduction) to predict the full solution (i.e. strain field of a material specimen under large shear strains) at a new point in the parameter space, without requiring computationally expensive model evaluation. Although the interpolated solutions are not constrained to satisfy physical principles, the resulting model can be viewed as a “physics-informed” surrogate in the sense that direct interpolation in the reduced-space of the physically/mathematically relevant bases is performed.
Dimitris Giovanis is an Assistant Research Scientist in the Department of Civil and Systems Engineering at Johns Hopkins University. He joined the University in 2016 as a Postdoctoral Fellow. He earned his five-year Diploma in Civil Engineering, his M.Sc. in Computational Mechanics from the Department of Chemical Engineering and his Ph.D. in Civil Engineering from the National Technical University of Athens in Greece. His primary research interests are data-driven uncertainty quantification (UQ) approaches for mathematically characterizing parametric and model-form uncertainties, that will inform decision making and eventually lead to the design of high performance physical and structural systems. Dr Giovanis is also a registered (licensed/chartered) professional Civil Engineer in Greece.