CRSS Determination: Combining Analytical Framework and Surrogate Neural Networks

CRSS Determination: Combining Analytical Framework and Surrogate Neural Networks
by Orçun Koray Çelebi
(University of Illinois at Urbana-Champaign, Department of Mechanical Science and Engineering)
DATE: July 18, 2023 (Tuesday)
TIME: 14:00-15:00
The yield strength of a crystalline structural material is a fundamental mechanical property predominantly governed by the friction (critical) stress for a dislocation to glide. Existing approaches for critical stress determination are highly unsatisfactory because of empiricism associated with determination of dislocation “core-width” and nature of core-advance. This study proposes a predictive model addressing both shortcomings. The core-width is rigorously determined from an optimized balance between continuum strain-energy and atomistic misfit-energy of the dislocation’s core. The strain-energy is calculated using the fully-anisotropic Eshelby-Stroh formalism accommodating the inherent mixed characters of the partials constituting the extended dislocation. The misfit-energy is determined from critical fault-energies of the slip-plane input to a novel misfit-model capturing the lattice structure of the slip-plane and involving the discrete Wigner-Seitz cell area at each lattice site, advancing over an 80-year old misfit-energy model that has missed the role of both concepts. For the first time in literature, the nature of motion of the extended-dislocation’s core is rigorously derived from an optimized trajectory of its total-energy. It is shown that each partial’s core moves intermittently (“zig-zag” motion), and not together, allowing the stacking-fault width to fluctuate during advance of the extended-dislocation. The critical stress is shown to involve a trajectory-dependent combination of Schmid factors for each partial, also revealed for the first time. The proposed model is used to predict critical stress for multiple FCC and HCP materials including pure metals, solid-solution alloys, and High Entropy Alloys (HEAs), displaying excellent agreement with experiments. Further, hypothetical combinations of material properties are employed to train a machine learning-based Surrogate Neural Network (SNN), and the ones of real materials are utilized to validate the SNN model yielding a 94% accuracy for 1,033 materials. The generated dataset is used to unravel the sensitivity of each material parameter to the predicted CRSS establishing a general trend for the FCC materials guiding the field in achieving superior mechanical properties. The work opens future avenues for rapid reliable assessment of a multitude of compositions across varying lattice structures, addressing a major void in structure-property prediction for structural materials, also instrumental for ab-initio materials design.
Orcun received his B.Sc. degree in Mechanical Engineering from Bogazici University in 2018. He is currently a Ph.D. candidate in Mechanical Science and Engineering Department at University of Illinois at Urbana-Champaign. His research focuses on computational & theoretical modeling of plastic deformation mechanisms (slip and twinning) in metallic materials (pure metals, alloys, and high entropy alloys) employing Density Functional Theory (DFT), Molecular Dynamics (MD), Machine Learning (ML), and Monte-Carlo Simulations.