Uhlherr A1, Clisby N1
1Swinburne University Of Technology, Hawthorn, Australia
Biography:
Alf Uhlherr obtained his BSc (Hons) and PhD at Monash University and postdoctoral fellowship at University of Cambridge. He worked for 23+ years at CSIRO, notably as a principal research scientist focusing on computational science and scientific computing. His main research area was development and application of molecular simulation methods for materials and fluids. As Senior Manager, Scientific Computing Services, he was responsible for establishing and delivering eResearch services across CSIRO, alongside governance roles with NCI, Pawsey, MASSIVE, NCMAS, AeRO and eResearch Australasia. Alf is currently an Adjunct Research Fellow at Swinburne University of Technology.
Abstract:
Molecular simulations form a major component of high-performance computing, data and visualisation workloads, as researchers address medical, engineering and environmental demands through painstaking development of new chemicals and materials.
The basic simulation techniques of energy minimisation or “molecular mechanics” (MM), Monte Carlo (MC) and molecular dynamics (MD) have existed since the 1950’s. They remain fundamentally limited to atomic (nano) size and time scales, which can be expanded somewhat by complementary resource-intensive “multi-scale” coupling approaches, such as quantum chemical descriptions, coarse graining, mesoscopic fluid models and machine learning (ML) methods.
The current status of these techniques can be likened to a jigsaw puzzle, with excess pieces but incomplete assembly.
A crucial piece is the successful incorporation of MM within a rigorous, generalised, microscopically reversible, minimum-mapping MC framework, dubbed “MinMap” (Uhlherr & Theodorou, J. Chem. Phys. 2006). MinMap allows uniquely efficient, controlled, collective local transitions in atomic positions, identities and connectivity. This enables complex and broad-range compositions and long effective timescales (years?) to be sampled within a single MC molecular simulation.
Current ML capabilities can now also enable automated iterative refinement of MC algorithm parameters. This will help in exploiting the versatility of MinMap, most likely coupled with MD using the Hamiltonian Monte Carlo (HMC) formalism. Such combinations also present new optimisation challenges, since MC, MM, MD and ML perform best across differing hardware architectures.
The presenters will outline their recent progress on identifying and optimising the full capabilities of the MinMap approach, within this wider current context.
