This presentation will discuss the use of Adaptive Neural Networks (ANNs) in atomistic computer simulations, focusing on their implementation and efficiency on High-Performance Computing (HPC) massively parallel architectures. This work builds on the emerging effort in materials science to employ machine-learning methods, including ANNs, for reproducing material properties from physics-based first principles. When properly trained, ANNs are shown to successfully emulate the complex atomic energy landscape, while achieving orders of magnitude faster performance compared to quantum mechanical calculations. The relatively simple functional form of the ANNs, composed of series of matrix operations, allows them to scale very well on multicore machines, reaching the speed of the classical empirical atomic potentials. The use of ANNs opens the possibility of simulating multi-million atoms systems, achievable so far only with classical potentials, while preserving the accuracy of the quantum mechanics-based calculations.
