COLLOQUIUM ON COMPUTATIONAL SOCIAL SCIENCE/DATA SCIENCES RESEARCH – The JARVIS project: Accelerating discovery of materials and validation of modelsusing classical, quantum and machine-learning methods– Francesca Tavazza, PhD

April 19, 2019 @ 3:00 pm
Karen Underwood

Computational Social Science Research Colloquium /
Colloquium in Computational and Data Sciences

Francesca Tavaza, PhD
National Institute of Standards and Technology Materials Science and Engineering Division

The JARVIS project: Accelerating discovery of materials and validation of models
using classical, quantum and machine-learning methods

Friday, April 19, 3:00 p.m.
Center for Social Complexity Suite, 3rd Floor Research Hall

All are welcome to attend.


Identifying new materials for technological applications is the goal of the Material Genome Initiative (MGI). As a response, NIST started the JARVIS project, a combination of atomistic databases at the classical and quantum level, and machine learning models. JARVIS-DFT is a collection of physical properties computed using Density Functional theory (DFT) for about 30000 materials. For each material, we determined its heat of formation, conventional and improved DFT bandgaps, dielectric function, elastic, phonon, electronic and transport properties. Statistical analysis of such properties allows to identify novel trends as well as new materials with desirable
properties. JARVIS-FF is a database of classically computed properties, designed to facilitate the user in choosing the right classical force field (FF) fortheir investigation. It uses the LAMMPS code to compute the same property, for the same material, with as many force fields as available (more than 25000 classical force-field). We focused on quantities like relaxed structures, elastic properties, surface energies, vacancy formation energies and phonon vibrations. JARVIS-FF contains these calculations for more than 3000 materials, so that a direct comparison between FF is easily achieved. Lastly, using all the properties in JARVIS-DFT as a training set, and novel descriptors inspired by FF-fitting, we developed machine learning (ML) models for all the properties studied in JARVIS-DFT. This allows to make on the fly predictions, and, therefore, to use ML to pre-screen materials.

Short Bio:

  • Undergraduate degree in Physics in Milan, Italy, 1993 (Universita’ Statale di Milano, Milano, Italy)
  • Master in Material Science in Milan, Italy, 1996 (Universita’ Statale di Milano, Milano, Italy). Dissertation topic: Tight-binding modeling of Cobalt and Iron Silicides, including fitting of the tight-binding parameters.
  • PhD in Physics at The University of Georgia, GA, USA in 2003 (PhD. Advisor: Prof. Davis Landau). Dissertation topic: Classical Monte Carlo simulations of Si and Si-Ge compounds under various conditions.
  • PostDoc at NIST starting in  2003, focusing on Density Functional theory (DFT) modeling of mechanical properties in metals.
  • Brief hiatus working at the Army Research Laboratory in 2008 for a short time, otherwise at NIST ever since I got there as a postdoc.
  • Currently: running an atomistic modeling group (both classical and DFT modeling) focused on the investigation of specific, technological relevant materials (TaS2, TaSe2, Bi2MnSe4, for instance) as well as on compiling databases of material properties. My group extensively uses artificial intelligence (AI) tools to accelerate material discovery as well as to build novel force fields (physics-inspired, neuron network-based fitting of Si, Ge, SiGe, AlNi potentials).