Atomistic fracture modelling by inference-boosted first-principles techniques

TY - CONF

T1 - Atomistic fracture modelling by inference-boosted first-principles techniques

AU - Glielmo, A.

AU - Zeni, C.

AU - Caccin, M.

AU - De Vita, Alessandro

PY - 2017/1/1

Y1 - 2017/1/1

N2 - Using extended atomic configuration databases and inference techniques can enable QM-accurate molecular dynamics simulations of fracture processes that require model system sizes and simulation times well beyond the reach of standard QM-based techniques. This paper reviews examples of this, notably including the thermally activated fracture and stress corrosion of example brittle materials. It also illustrates a practical implementation of the scheme where QM-accurate information is either database-retrieved or generated on the fly only where/when “chemically novel” configurations are encountered, and then used to predict atomic forces via Bayesian inference. This notably implies recipes to predict QM-accurate forces via a Gaussian process, using suitable "covariant" kernel functions, and a framework to discuss (and some care in controlling) the incurred prediction errors. The approach naturally leads to novel QM-zone partition strategies well suited for high-end parallel simulation of fracture propagation.

AB - Using extended atomic configuration databases and inference techniques can enable QM-accurate molecular dynamics simulations of fracture processes that require model system sizes and simulation times well beyond the reach of standard QM-based techniques. This paper reviews examples of this, notably including the thermally activated fracture and stress corrosion of example brittle materials. It also illustrates a practical implementation of the scheme where QM-accurate information is either database-retrieved or generated on the fly only where/when “chemically novel” configurations are encountered, and then used to predict atomic forces via Bayesian inference. This notably implies recipes to predict QM-accurate forces via a Gaussian process, using suitable "covariant" kernel functions, and a framework to discuss (and some care in controlling) the incurred prediction errors. The approach naturally leads to novel QM-zone partition strategies well suited for high-end parallel simulation of fracture propagation.

UR - http://www.scopus.com/inward/record.url?scp=85066052642&partnerID=8YFLogxK

M3 - Paper

AN - SCOPUS:85066052642

SP - 350

EP - 351

T2 - 14th International Conference on Fracture, ICF 2017

Y2 - 18 June 2017 through 20 June 2017

ER -