GTC 2016 San Jose, Californiae, 7 April, 2016 Xiaoxia Li Group of HPC & Cheminformatics Institute of Process Engineering Chinese Academy of Sciences, Beijing
Outline Reaction mechanisms of coal pyrolysis? 1 2 GPU-enabled ReaxFF MD (GMD-Reax) 3 Pyrolysis of coal, biomass, polymer 4 Concluding remarks and perspective 2
Background China is the largest producer & consumer of coal China has much more coal, less oil Reaction mechanism ? Mechanism still hardly accessible Experimentally, hard to detect and ReaxFF MD replicate the free radical initiation at high temperature in lab (Reactive molecular dynamics) Computationally with QM, extremely high computing cost, limited model scale: ~100 3 atoms
Overview of ReaxFF ReaxFF MD: reactive force field + molecular dynamics Publications on ReaxFF MD Subject searching hits from Web of Science by van Duin (Penn state), Goddard (Caltech) et al. for b ond breaking and forming with parameters based on experiments and QM (quantum mechanics approach) Faster than DFT (widely used QM) for models > 1000 atoms No priori knowledge of reaction pathways required A comprehensive knowledge on multiple reaction ReaxFF MD is promising pathways of coal pyrolysis is not available ! for coal pyrolysis simulation 4
Can large coal model simulated efficiently with ReaxFF? HPC Programs of ReaxFF - supercomputer/cluster F-ReaxFF, Univ. South. California, 2007 ( parallel ) PuReMD, Purdue Univ., 2011 ( single node performance ) In LAMMPS, Sandia National Lab. (open source) FORTRAN code (precise, based on van Duin’s original code) C code (2011, faster , based on PuReMD) In commercial software ADF (to enhance visualization, ~2011) GULP, Materials Studio 6.0 (2012) Desktop workstation Is it practical to simulate large coal model (~10,000 atoms) is more preferable on desktop workstation? 5
ReaxFF MD on Desktop workstation? Computational challenges – complexity of coal structure and pyrolysis ~10,000 atoms, s tate-of-the-art coal model scale ~1,000 atoms, practical scale for LAMMPS (Sandia National Lab) and ADF (Europe, a major player of QM software) on single computational node ReaxFF vs LJ potential 10 - 50 folds LAMMPS Benchmarks slower than 2012: classical MD http://lammps.sandia.gov/ bench.html#potentials ) C code FORTRAN code 6
Overview of ReaxFF MD ReaxFF MD MD Dynamic atom charge equilibration Fixed atom charge Time-step 0.1 fs Time-step 1 fs Bond order dependency 7
Computational cost of ReaxFF MD vs MD ReaxFF MD vs MD Similar computing loops, but Time-step: 0.1 fs (ReaxFF MD) vs 1 fs (MD) Atom charge: optimizing at each time-step (ReaxFF MD) vs fixed (MD) Additional computing introduced in potential & its corrections Taper + Morse for van der Waals in ReaxFF 8
ReaxFF MD on Desktop workstation? GPU Thanks for the GPU & CUDA Rapid development GPU computing since 2007 MD codes (major players and novel codes such as HOOMD) Stone, J.E., et al., GPU-accelerated molecular modeling coming of age. Journal of Molecular Graphics and Modelling, 2010. 29(2): p. 116-125. GPU infrastructure in IPE (in my office building) Mole-8.5 .5 (GPU enabled) d) 1 Pet eta, Double Top 500 Supe perco comput puter er 19 19 th th , 2010 33 th 33 th , 2011 37 37 th th , , 2012 55 55 th th , 2013 13 Potential seen from GMD we created in 2009 - 2010 (a GPU enabled code for MD) Polyethylene crystalization 9
GMD and its applications in polymer crystallization study GMD: a GPU enabled code for classical MD Our first attempt using GPU Performance is comparable with early version of GROMACS GPU Application in polymer chain crystallization (Polyethylene as model) PE models: 360,000 united atoms & 400,000 united atoms Simiao Wang, et al. Two mechanisms of polymer chain crystallization within nanoglobule. Polymer. 2013;54(15):4030-4036 10 folds larger model scale than 10 Students in GPU HPC companies (NVIDIA, Sugon) and more that simulated in CPU cluster
GMD-Reax: ReaxFF MD on GPU GPU works for MD the first GPU code for ReaxFF MD (C2050) Its implementation – tough job Constrained coding closely linked with GPU hardware faster memory limited, global memory access latency, and more 11
GMD-Reax: ReaxFF MD on GPU Our approach Most of computations on GPU Faster SFU for some bond order based corrections (early version) T thread for charge evaluation/time-step – bottle neck Finely tuned data access for computation, and more 12
GMD-Reax: performances GMD-Reax on one C2050 achieved up to 16 times speedup against the LAMMPS’ codes on 8 CPUs (~fastest on CPU, Sandia National Lab & Purdue Univ) Single precision Zheng, M.; Li, X.; Guo, L., Algorithms of GPU-enabled reactive force field (ReaxFF) 13 molecular dynamics. Journal of Molecular Graphics and Modelling 2013, 41, (April), 1-11
GMD-Reax: performances GMD-Reax on one C2050 achieved up to 8 times speedup against the LAMMPS’ codes on 8 CPUs (~fastest on CPU, Sandia National Lab & Purdue Univ) Double precision 14
GMD-Reax: performance & impact GMD-Reax PuReMD-GPUs Notes (Ours, DP) (Purdue Univ.) Bulk water systems Coal models are Amorphous coal pyrolysis (6540 – 50 097 atoms) more complex than Systems Benchmarked systems Amorphous silica bulk water or silica (4976 – 27 283 atoms) (6000 – 48 000 atoms) systems, of which all Hardware of GPU Tesla C2050 Tesla C2075 energy terms must Speedups against be computed in 4.5 – 14.0 7.1 – 16.6 (water) PuReMD in LAMMPS potential evaluation (complex coal models) 5.8 – 11.4 (silica) (1 CPU core) of ReaxFF MD Tesla C2075 has Speedups against 1.5 – 4.0 2.0 – 2.9 (water) PuReMD in LAMMPS more global memory (complex coal models) 1.5 – 2.1 (silica) (8 CPU cores) than Tesla C2050 PuReMD-GPUs : Journal of Computational Physics Ours : Journal of Molecular Graphics and Modelling 2014, 272(Sept), 343-359 2013, 41, (April), 1-11 Top 5, NVIDIA GPU Award, 248 th ACS meeting, 2014 The only two GPU codes available have comparable performance, ours even better Ours published ~ 1.5 year earlier 15
ReaxFF MD of coal pyrolysis Challenges – complexity of coal structure and pyrolysis Coal model construction? Computing scale discrepancy? Lack of reaction analysis ability for revealing mechanism LAMMPS, ADF analysis tool (?) number of molecules (formula based) ~ time Manual analysis is a must? Manual analysis is not practical for revealing the n-dodecane (C 6 H 14 ) pyrolysis: complex reaction mechanism of coal pyrolysis 1279 species, 5056 reactions 16
VARxMD: the first reaction analysis tool for ReaxFF MD What we need to do? Reaction analysis - discovering the bonding and species changes 3D chemical structure processing Automatic perception of atomic connectivity, bonding type, species, reaction 17 Jian Liu, Xiaoxia Li et al., Journal of Molecular Graphics and Modelling 2014, 53(9):13-22
VARxMD: the first reaction analysis tool for ReaxFF MD What we have – detailed reaction list All reactions Product evolution & underlying reactions 2D & 3D Reaction details Allowing for “direct” observation of chemistry events computationally 18
VARxMD: the first reaction analysis tool for ReaxFF MD What we have – a view of all reaction sites Allowing for “direct” observation of chemistry events computationally 19
VARxMD: the first reaction analysis tool for ReaxFF MD What we have – a 3D view of a reaction with reaction sites highlighted Reaction site – bond breaking or forming highlighted 20
New methodology for large scale ReaxFF MD GPU high performance computing We created the first GPU-enabled codes Xiaoxia Li et al., Molecular Simulation, 2015, 41(1-3), 13-27 Cheminformatics approach We created the first reaction analysis tool 21
New methodology applications Typical time for one condition is one week Large scale ReaxFF MD simulations (GMD-Reax) Coal pyrolysis (~10,000 atoms) Liulin coal model: C14782H12702N140O690S37, 28,351 atoms, second largest ever simulated Pyrolysis of polymer (HDPE) (150x8, 7216 atoms) Pyrolysis of biomass 15,920 atoms for lignin 7572 atoms (C2160H3612O1800) Pyrolysis and oxidation hydrocarbon fuel 10,828 atoms for bio-oil Tingting Zhang, Xiaoxia Li, et al. Energy and Fuels 2016, just accepted Mo Zheng, Ze Wang, Xiaoxia Li, et al. Fuel, 2016. 177: p. 130-141 Xiaolong Liu, Xiaoxia Li, et al. Polymer Degradation and Stability 2014, 104(June), 62-70 22 Mo Zheng, Xiaoxia Li, et al. Energy and Fuels 2014, 28(1), 522-534
Recommend
More recommend