Atomistic molecular simulations for engineering applications: methods, tools and results Jadran Vrabec
Motivation Simulation methods vary in their level of detail The more detail, the more predictive power Quantum chemical methods scale with O(M 3 ) O(M 7 ) [M: basis functions] Force field methods are favorable with respect to scaling and thermodynamics
Force fields Thermodynamics Classical models for molecular interactions Parameters have a physical interpretation Contain all thermodynamic properties Static: thermal, caloric, entropic Dynamic: viscosity, diffusion, thermal conductivity, Surface properties, e.g. surface tension Iso-Butane Straightforwardly applicable to mixtures Excellent predictive power Technical accuracy Directly applicable for the study of fluids in geometries, e.g. wetting, adsorption, zeolites, in processes, e.g. condensation, flow,
Sampling force fields Molecular dynamics Monte Carlo
ms2: simulation tool for thermodynamic properties Deublein et al., Comp. Phys. Commun. 182 (2011) 2350 Molecular dynamics / Monte Carlo Arbitrary mixtures of rigid molecules Grand equilibrium method (for VLE) Several classical ensembles All static properties (thermal, caloric, entropic) Gradual insertion for entropic properties Transport properties (Green-Kubo) Consistent FORTRAN90 code Object oriented All loops vectorized MD and MC parallelized 3D visualization interface
Equation of state for CO 2 (Span and Wagner, 1996) T = 216 1100 K, p = 0 800 MPa F R T 0 Res,,, τ =T c / T δ = ρ / ρ c Ideal part: Residual part: 7 0 0 0 0 0 1 2 3 i i i 4, ln a a a ln a ln 1 exp n 7 34 Re s t i di ti di ci, ai ai exp i 1 i 8 39 ti di 2 2 ai exp i( i) i( i) i 35 42 bi 2 2 ai exp C i( 1) D i( 1) i 40
Derivatives of the Massieu-Planck potential F /(RT ) nm n nm / T m 1 NVT 10 01. 20. A total of nine independent properties sampled per NVT simulation with ms2 Any equilibrium thermodynamic property = combination of nm s
Cyclohexane Rigid 6CLJ united-atom model by Merker et al., Fluid Phase Equilib. 315 (2012) 77 Density / mol m -3
Cyclohexane potential energy Density / mol m -3 Present simulation data Span and Wagner, Int. J. Thermophys. 24 (2003) 41 Penoncello et. al., Int. J. Thermophys. 16 (1995) 519
Cyclohexane pressure Density / mol m -3 Present simulation data Span and Wagner, Int. J. Thermophys. 24 (2003) 41 Penoncello et. al., Int. J. Thermophys. 16 (1995) 519
Cyclohexane isochoric heat capacity Density / mol m -3 Present simulation data Span and Wagner, Int. J. Thermophys. 24 (2003) 41 Penoncello et. al., Int. J. Thermophys. 16 (1995) 519
Ongoing project with ms2 9 independent thermodynamic data types from one NVT simulation Generation of an extensive dataset in an automatized fashion Parallel execution of (~80 simulation runs) x (9 data points) Fit of a fundamental EOS in terms of F res to these data EOS may serve for the optimization of the force field EOS may be the starting point for technical EOS fitting Poster T21: Thol, Rutkai, Span, Vrabec: Molecular simulation of thermodynamic properties and an equation of state for the Lennard-Jones model fluid
Present status and outlook for ms2 An efficient molecular simulation tool for thermodynamic properties of homogeneous fluids New features Ewald summation for ionic systems Pair correlations functions MPI/OpenMP hybrid parallelization Next development steps Integer arithmetics Internal molecular degrees of freedom Execution on graphics processing units (GPUs)
Force fields for Hydrazine and two derivates 600 Hydrazine T / K 500 400 300 Hydrazine Dimethylhydrazine Monomethylhydrazine Monomethylhydrazine + 0 10 20 30 Simulation, this work r / mol/l Experimental data from the literature Simulation, Gutowski et al. 2009 Dimethylhydrazine
VLE and gas solubility of systems containing Hydrazine, Monomethylhydrazine and Dimethylhydrazine Elts et al., Fluid Phase Equilib. 322-323 (2012) 79 390 Water + Hydrazine @ 1 bar H i / MPa H i / MPa 20000 15000 10000 5000 0 1500 1000 500 0 Hydrazine Monomethylhydrazine Stickstoff Nitrogen N2 Argon Ar Stickstoff Nitrogen Argon T / K 380 Simulation, this work + Experimental data (literature) Peng-Robinson EOS 370 0.0 0.2 0.4 0.6 0.8 1.0 x,y H2O / mol/mol H i / MPa 300 200 100 0 Stickstoff Nitrogen Carbon Kohlenstoffmonoxid monoxide Argon Dimethylhydrazinee 260 280 300 T / K
VLE and gas solubility of CO 2 -expanded liqiuds Cyclohexane (C 6 H 12 ) 6 LJ sites Carbon dioxide (CO 2 ) 3 LJ sites + point quadrupole Oxygen (O 2 ) 2 LJ sites + point quadrupole Cyclohexanol (C 6 H 10 O) 7 LJ sites + point dipole + CH - O Cyclohexanol (C 6 H 12 O) 7 LJ sites + 3 point charges Poster T37: Merker, Hsieh, Lin, Hasse, Vrabec: Fluid phase equilibria for the oxidation of cyclohexane in carbon dioxide expanded liquids from experiment, molecular simulation, Peng-Robinson EOS and COSMO-SAC + H
Speed of sound of Hexamethylsiloxane (MM) deviation [%] 20 15 10 5 0 365 K EOS: Colonna et al., 2006-5 0 5 10 15 pressure [MPa] Poster T12: Dubberke, Vrabec: Speed of sound of siloxanes as workings fluids in Organic Rankine Cycles
Acetone in N 2 and O 2 under extreme conditions 12-50 C -30 C 30 C 0 C 50 C 70 C Acetone Presentation Wed 10:40: Windmann, Köster, Vra.: Study on vapor-liquid equilibria of nitrogen + acetone and oxygen + acetone with a focus on the extended critical region p / MPa 8 4 0 0.00 0.02 0.04 0.06 x N2 / mol/mol 90 C 126.85 C
Massively parallel molecular dynamics code: ls1 Identical force field types as with ms2 Additionally, Tersoff potential for solids Large systems, long time scales Concurrency in space, not in time Flow Nucleation
Scaling tests on Cray XE6 (Hermit) Peak performance: 1.045 Pflops 10 15 operations / s Racks: 38 with 96 nodes each Nodes: 3552 Cores: 113.664 (2 sockets each with 16 cores / node) Processor: AMD Interlagos @ 2.3 GHz #12 (November 2011)
Rechenzeit / s Execution time / s Scaling of ls1 on Cray XE6 (Hermit) Computational effort ~ O(N 1 ) 4.096 cores 1000 100 10 10 5 10 6 10 7 10 8 Molecule number Molekülzahl 1.000 time steps
Octree load balancing strategy of ls1 Hierarchical subdivision of the simulation volume through recursive bisection Choice of planes for bisection that lead to an equal load in the subvolumes Designed for strongly inhomogeneous molecular systems Capable to deal with rapidly changing inhomogeneity
Scenarios for scaling tests of ls1 Bulk: homogeneous liquid (Ethylene oxide) Domain decomposition trivial Slab: liquid slab surrounded by vapor in equilibrium (Argon) Topology for domain decomposition simple Droplet: liquid droplet surrounded by vapor in equilibrium (Argon) Topology for domain decomposition more complex Bulk Slab Droplet
Scaling of ls1 on Cray XE6 (Hermit) Strong Scaling 2 22 = 4.194.304 molecules Execution time / s 1000 Bulk 100 10 Slab Film Tropfen Droplet 10 2 10 3 10 4 10 5 Cores 1.000 time steps
Scaling of ls1 on Cray XE6 (Hermit) Strong scaling 2 26 = 67.108.864 molecules Execution time / s 1000 100 10 Film Bulk Tropfen Droplet Slab Bulk 10 2 10 3 10 4 10 5 Cores 1.000 time steps
VLE of Nitrogen + Ethane Stoll et al., AIChE J. 49 (2003) 2187 15 ξ = 0.974 Simulation simulation Peng-Robinson equation of state Experiment experimental (lit.) data pressure [MPa] 10 5 290 K 200 K 0 0.0 0.2 0.4 0.6 0.8 1.0 mole fraction (N 2 )
Direct simulation of the LLE Mixture of 60 mol-% Nitrogen and 40 mol-% Ethane Identical molecular model as before 20,000 molecules Molecular dynamics with ls1 Canonical ensemble (NVT) Initial configuration with randomly dispersed components
LLE of N 2 + C 2 H 6 after 48 ns mole fraction 1,0 0,9 0,8 0,7 0,6 0,5 0,4 0,3 0,2 0,1 0,0 0 2 4 6 8 10 12 14 16 18 box length [nm] N2 C2H6 Temperature: 128 K Pressure: 11.0 MPa Average over 500,000 time steps
LLE temperature dependence of N 2 + C 2 H 6 135 130 temperature [K] 125 120 p = 1.8 to 4.0 MPa, depending on exp. data 115 + Simulation, this work Experimental data, literature 0.0 0.2 0.4 0.6 0.8 1.0 mole fraction of N 2 [mol/mol]
LLE pressure dependence of N 2 + C 2 H 6 Poster T25: Eckelsbach, Vrabec: Prediction of liquid-liquid equilibria of nitrogen + ethane with a molecular model that was adjusted to vapor-liquid equilibria pressure [MPa] 20 15 10 5 T = 127 to 129 K, depending on exp. data + Simulation, this work Experimental data, literature 0 0.0 0.2 0.4 0.6 0.8 1.0 mole fraction of N 2 [mol/mol]
Summary The computational effort of classical force field methods scales linearly with the molecule number Classical force fields contain the thermodynamic properties adequately Molecular dynamics simulations of inhomogeneous fluids may efficiently use Petaflop machines The spectrum of possible applications is very wide