TD Swinburne

Autonomous exploration at scale: TAMMBER
In massively parallel computatation, decisions must be made every few minutes to optimise resources. TAMMBER uses novel Bayesian methods to autonomously allocate computational effort only in the most relevant regions of configuration space, building kMC/Markov models with robust UQ.
TDS† and D Perez, NPJ Computational Materials (2020) TAMMBER
TDS† and D Perez, PR Materials (2018)

ML-assisted DFT for dislocation-defect interations
(Postdoc of Petr Grigorev, from ANR JCJC MeMoPas)
Ab initio methods such as DFT are essential to train and validate atomic models, but extended defects are at or beyond current computational limits. We design efficient methods to embed DFT accuracy only where it is needed, using custom-built machine learning interatomic potentials. This builds on earlier developments that enabled energy calculations.
P Grigorev et al., TDS†, arXiv (2022) LML-retrain
TDS† and JR Kermode, PRB (2017)

Vibrational free energies beyond harmonicity: PAFI
PAFI is a path-constrained sampling routine, which returns exact vibrational free energy differences for linear-scaling effort. This enables evaluation of activation free energies for very large systems of e.g. dislocation or twin migration, which can not be treated by cubic-scaling HTST approximations.
TDS† and M-C Marinica, PRL (2018) PAFI

Surrogate ML models for defect entropics
(PhD of Clovis Lapointe, with M-C Marinica)
A linear machine learning model typically used for energetics can predict the formation and migration entropy of defects. Generalised linear models allow a parameter-geometric interpretation of relations such as the Meyer-Neldel compensation law.
C Lapointe, TDS† et al. Submitted (2022)
C Lapointe, TDS† et al. PRMaterials (2020)

Analysis of Markov Models: PyGT
Much atomic data (from TAMMBER or other codes) is a connected network of metastable states, which can be mapped to a discrete state Markov model, or master equation. We have developed tools to automatically propagate atomic data, with UQ, to drift-diffusion equations or first passage distributions between distant regions.
TDS† and D Perez, arXiv (2022)
TDS† and DJ Wales, JCTC (2020) PyGT

Mean-field models for lattice vibrations: BLaSa
BLaSa is a lightening-fast mean field approach to evaluate anharmonic vibrational phase free energies for milliseconds of effort. For unary FCC we are within meV/atom of brute force thermodynamic integration, requiring hours of effort. Current developments are within the PhD of R Dsouza since 2020.
TDS† et al., J Neugebauer PRB RC (2020) BLaSa