Keynote
FIRST-PRINCIPLES MOLECULAR DYNAMICS IN THE BIG DATA ERA

The current Big Data revolution has a potential to enhance virtually any materials modelling activity. A main expectation is that it will boost our ability to identify new key microscopic mechanisms and processes at the core of materials behaviour by enabling higher information efficiency through massive computational savings. Namely, modellers might eventually be able to compute just information on their target physical systems which is still genuinely missing, and infer all the rest from established or rapidly growing open-access “big” databases using appropriate mathematical tools such as neural networks or kernel-based regression techniques.

This talk will focus on how using extended databases and advanced inference techniques can enable quantum mechanical(QM)-accurate molecular dynamics simulation in problems requiring model system sizes and simulation times well beyond the reach of standard QM techniques. The target processes may e.g., include the thermally activated fracture of brittle material samples[1], or their stress corrosion when exposed to environmental chemistry[2]. In particular, I will present a data-boosted molecular dynamics technique where QM-accurate information is either located in (potentially massive) databases, or generated on the fly if unavailable, and used to predict atomic forces via Bayesian inference [3]. The approach is accurate and efficient as force validation can be accurately monitored, while new QM calculations are carried out only where/when “chemically novel” configurations are encountered along the simulated system’s trajectory.

[1] J.R.Kermode, et al., PRL 115, 135501 (2015).

[2] A. Gleizer et al., PRL 112, 115501 (2014).

[3] Z. Li et al., PRL 114, 096405 (2015).