CDAC Webinar- “First-Principles Approaches and Machine Learning Applications for Thermodyamics and Kinetics of Solids” with Prof. Sara Kadkhodaei (UIC CME)
CDAC Webinar
April 16, 2025
1:00 PM - 2:00 PM
Calendar
Download iCal File"First-Principles Approaches and Machine Learning Applications for Thermodynamics and Kinetics of Solids
Sara Kadkhodaei
Department of Civil, Materials, and Environmental Engineering
University of Illinois Chicago
Abstract:
First-principles methods and artificial intelligence have demonstrated significant success across a wide range of tasks related to materials understanding and design. In this talk, I will present several examples of research that leverage either first-principles modeling or machine learning to uncover material phenomena and accelerate their discovery. My goal is to highlight opportunities for potential collaboration where these computational approaches could provide meaningful impact.
For machine learning applications, I first present our pioneering work using visual image representations and deep convolutional neural networks (CNNs) to predict synthesizability and formation energy across diverse crystal structures and compositions. I highlight the model’s predictions, its integration into XtalOpt—a widely used evolutionary crystal search software—and its role in accelerating materials design. I will also discuss the impact of data bias on predictions and introduce our recent model that leverages large language models to address data scarcity, with applications in graphene synthesis.
Second, I will present first-principles modeling of thermodynamics and diffusion kinetics in high-temperature solid phases that exhibit low-temperature mechanical (or phonon) instabilities. For thermodynamic modeling, I introduce a coarse-grained statistical model that uses first-principles data to construct Gibbs energy functions for these unstable phases. For diffusion kinetics, we developed a new saddle-point search method based on a Gaussian process regression surrogate model of a temperature-dependent energy surface (TDES). This approach uses stochastic atomic configurations to efficiently sample the TDES while converging toward the saddle point via the dimer algorithm. Identifying saddle points with this method enables the application of transition state theory to diffusion in these systems—offering a practical alternative to current approaches, which rely on direct molecular dynamics simulations of individual diffusive hops.
Finally, I present our newly developed first-principles CALPHAD model for dilute defects, which advances defect thermodynamics by removing the need to determine the equilibrium chemical potential or Fermi level for atom and electron exchange with the environment. Instead, the model introduces universal parameters—directly derived from first-principles data such as DFT-calculated excess energies and band gaps—that are incorporated into the Gibbs energy formulation. This direct linkage between Gibbs energy parameters and first-principles data eliminates the need for fitting to experimental or simulation results, significantly accelerating the prediction of charge carriers and defects in solids.
Additional Background:
Upcoming and Previous CDAC Talks: https://cdac.phys.uic.edu/cdac-webinar-series/
Recorded CDAC Talks: https://www.youtube.com/channel/UCah_ErVY7qVCRief1WaF7hA?view_as=subscriber
Date posted
Feb 17, 2025
Date updated
Apr 14, 2025