Source-linked AI summary
GALAMOST: GPU-accelerated large-scale molecular simulation toolkit
You-Liang Zhu, Hong Liu, Zhan-Wei Li, Hu-Jun Qian, Giuseppe Milano, Zhong-Yuan Lu
TL;DR
Mesoscopic simulation of polymeric systems requires methods that address atomistic MD bottlenecks while exploiting GPU computation. GALAMOST packages GPU-optimized coarse-grained methods, force fields, anisotropic-particle models, and polymerization models, enabling larger polymeric systems and longer simulations, with reported capacity advantages over HOOMD-blue on one GPU.
Problem
Atomistic MD has bottlenecks at mesoscopic time and length scales, while few highly optimized GPU packages efficiently target polymeric systems.
Method
GALAMOST combines CUDA-optimized coarse-grained simulation methods with MD-SCF, structure-derived numerical potentials, anisotropic-particle models, and stochastic chain-growth polymerization.
Results
2.2 million LJ liquid particles or 3.0 million DPD liquid particles can be simulated on a GTX 580, versus 1.0 million or 1.5 million, respectively, for HOOMD-blue under the same conditions.
Takeaways & Limitations
GALAMOST facilitates studies of polymeric systems involving self-assembly, phase separation, anisotropic-particle superstructures, and polymerization.
Abstract
from arXiv · showhide
A new molecular simulation toolkit composed of some lately developed force fields and specified models is presented to study the self-assembly, phase transition, and other properties of polymeric systems at mesoscopic scale by utilizing the computational power of GPUs. In addition, the hierarchical self-assembly of soft anisotropic particles and the problems related to polymerization can be studied by corresponding models included in this toolkit.
INTRODUCTION
Atomistic molecular dynamics is limited for polymeric systems at mesoscopic time and length scales, motivating optimized parallel GPU implementations. GALAMOST addresses the scarcity of highly optimized GPU-based packages designed specifically for polymer simulations.
- Atomistic MD faces bottlenecks when addressing mesoscopic time and length scales in polymeric and biological systems.
- Coarse-graining and enhanced sampling accelerate simulations while aiming to maintain computational accuracy.
- 3.95 Teraflops is the theoretical single-precision throughput of NVIDIA’s Tesla K20X GPU processor.
- CUDA requires MD algorithms to be redesigned into massively parallel, data-parallel forms suited to GPU execution.
- GALAMOST is developed because few highly optimized GPU-based MD packages efficiently simulate polymeric systems.
A SKETCH OF GALAMOST
GALAMOST combines GPU-optimized coarse-grained simulation methods with tailor-made models for anisotropic-particle self-assembly and stochastic polymerization. Its force-field options include density-field interactions and structure-derived numerical potentials.
- Force fields: MD-SCF replaces intermolecular pair interactions with particle interactions through density fields to speed some collective processes.
- Force fields: Numerical potentials can be derived by iteratively matching coarse-grained structural distributions, such as RDFs, to atomistic simulations or experiments.
- Core simulation methods: GALAMOST implements coarse-grained MD, Brownian dynamics, and dissipative particle dynamics on GPUs.
- Characteristic models: The toolkit includes soft anisotropic-particle and stochastic chain-growth polymerization modules for specialized polymeric-system studies.
- Characteristic models: Different anisotropic potentials enable studies of disk-like, rod-like, Janus, and triblock Janus particle packing and self-assembly.
- Characteristic models: The polymerization model has applications including polymerization-induced phase separation and surface-initiated polymerization.
MOLECULAR DYNAMICS METHODS AND ALGORITHMS
GALAMOST organizes neighbor, cell, and bond data for data-parallel GPU calculations across several molecular-dynamics methods. Its algorithms support standard and density-field forces, tabulated potentials, and multiple integration schemes.
- GPU data structures: Neighbor-list and cell-list arrays store particle-neighborhood data in device memory for GPU-based force calculations.
- Short-range non-bonded forces: A GPU thread builds cell and neighbor lists for particles, searching neighboring particles through surrounding cells.
- MD-SCF forces: MD-SCF computes intermolecular forces from density fields while retaining regular intramolecular interaction forms.
- MD-SCF forces: GPU density-field calculations assign particle density fractions to cell vertices and use atomic operations during parallel updates.
- Bonded forces: Bond, angle, and torsion forces are computed per particle using bond-list and bond-number arrays that store molecular topology.
- Numerical forces: Tabulated potentials are fitted with cubic splines whose values and first two derivatives remain continuous across grid intervals.
- Integration algorithms: Velocity-Verlet, GW-VV, and half-step leapfrog algorithms integrate the included CGMD, BD, DPD, and soft-anisotropic-particle models.
CHARACTERISTIC COARSE-GRAINED MODELS
GALAMOST combines coarse-grained models for soft anisotropic particles and stochastic chain-growth polymerization, with GPU-optimized algorithms for mesoscopic polymer simulations.
- Soft anisotropic particles: Adding two rotational degrees of freedom lets the one-site anisotropic particle model represent disk-like, rod-like, diblock, and triblock Janus particles.The model has been used to examine their packing and self-assembly.
- Soft anisotropic particles: The disk-like and rod-like models use soft anisotropic potentials whose factors encode particle orientations and angular interactions.The repulsion strength, particle shape, and angular width are controlled by model parameters.
- Soft anisotropic particles: Janus particle models extend the anisotropic potential with an attractive tail and describe attractive patch size through the Janus balance β.The angular factor depends on the orientations of the interacting particles and their relative vector.
- GPU implementation: The GPU implementation includes optimized translational and rotational dynamics for four anisotropic particle types, with forces and torques determined by the selected potential.Rotational motion uses the particle moment of inertia, orientation derivatives, and torque components.
- Chain-growth polymerization: The polymerization model represents free-radical linear chain growth, coupling reaction probability to the real reaction rate and updating bonds and active ends after reactions.Device-memory storage of bond connections and atomicMax() handle concurrent reactions involving the same monomer.
PERFORMANCE AND VALIDATION
GALAMOST is benchmarked against CPU and GPU simulation codes for performance, numerical precision, and model correctness. The results show greater device-memory capacity, faster DPD simulations, and reproducibility across several coarse-grained models.
- GPU performance: The neighbor-list buffer distance has an optimal system-dependent value because larger buffers reduce rebuild frequency but increase the number of particles searched for non-bonded interactions.Neighbor-list construction is generally the most time-consuming component for short-range non-bonded interactions.
- GPU performance: About 30% faster DPD liquid simulations were obtained with GALAMOST than with HOOMD-blue under the same simulation conditions.GALAMOST uses the Saru pseudo-random number generator for stochastic DPD forces.
- GPU performance: 2.2 million LJ liquid particles or 3.0 million DPD liquid particles can be simulated on a GTX 580 with 1.5 gigabytes of global device memory.Under the same conditions, HOOMD-blue supports at most 1.0 million LJ or 1.5 million DPD particles.
- Validation: GALAMOST reproduced reference results for phospholipid density profiles, polystyrene bond and angle distributions, and anisotropic-particle energies, momenta, temperatures, and self-assembled structures.These comparisons used OCCAM, IBIsCO, or established CPU implementations as references.
- Validation: GALAMOST reproduced the polymer-brush structure generated at specified initiator density and polymerization rate for surface-initiated polymerization.The GPU algorithm was tuned to handle simultaneous reactions of multiple active ends with one monomer.
- Numerical precision: For rcut = 3.4 and dt ≤0.004, RMS deviations of total energy remained below 4 × 10^-5, while larger time steps increased deviations.For rcut = 6.0 and dt ≤0.004, GPU SP and CPU SP showed several-times-larger deviations than GPU DP and CPU DP.
- Numerical precision: For rcut ≤3.0 and dt ≥0.002, numerical precision has comparatively small influence on coarse-grained MD trajectories, although GALAMOST also supports GPU double precision.For dt > 0.004, the integration time step dominates the dependence of relative energy RMS deviations on numerical treatment.
CONCLUSIONS
GALAMOST integrates specialized force fields and characteristic models into a single-GPU molecular dynamics toolkit for polymeric systems. Its optimized GPU implementations support larger systems, longer simulations, anisotropic-particle self-assembly, and polymerization studies.
- GALAMOST incorporates MD-SCF and coarse-grained numerical potentials derived by fitting structural distributions to atomistic simulations or experiments.The numerical potentials include non-bonded, bond, angle, and torsion interactions.
- GALAMOST includes soft anisotropic-particle and chain-growth polymerization models for studying self-assembled superstructures and polymerization-related problems.New anisotropic potentials can describe novel anisotropic particles.
- GALAMOST's optimized GPU algorithms are benchmarked against HOOMD-blue and corresponding CPU codes, supporting larger polymeric systems over longer time.The package is designed to facilitate studies of various polymeric systems.