A scalable framework for goal-driven task-based simulations
- 1Max Planck Computing and Data Facility Giessenbachstrasse
- 2Fritz-Haber-Institut der Max-Planck-Gesellschaft Faradayweg 4-6
Abstract
Realistic surfaces of active materials that involve the motion of atoms or ions in their function typically change their (sub-)surface composition and geometry under operating conditions due to varying environmental stimuli, degradation or healing processes, but also just while exhibiting their intended function, making it difficult to build simulation models that remain valid as the system evolves. Such problems often demand large-scale computation together with efficient data processing, so scalable task execution and data-efficient synchronization become critical. We present a scalable computational framework that supports large collections of independent computation tasks with incremental result aggregation. We demonstrate the framework using adaptive kinetic Monte Carlo (akMC) as a showcase application, where transition-state searches are performed on-the-fly. The new framework employs a hierarchical orchestrator–worker architecture implemented with modern MPI constructs and operates efficiently from single workstations to multi-node high-performance computing systems. It integrates well-established simulation engines such as LAMMPS and supports multi-core CPU and GPU parallelism within each task. Performance is demonstrated using classical, machine-learning, and GPU-accelerated interatomic potentials. Benchmarks show strong scaling across multiple compute nodes with several hundreds of CPU-cores, efficient utilization of multi-core executors and GPU sharing. The results indicate that the framework provides an efficient solution for rare-event and mesoscopic simulation workflows.
Related
- A Shared-Task-State-Driven Multi-Agent Framework for Knowledge Graph Construction in Substation Operation and Maintenance
- A Lightweight Universal Machine-Learning Interatomic Potential via Knowledge Distillation for Scalable Atomistic Simulations
- Mamba Quantum Eigensolver (MQE): A Selective State Space Model for Scalable Ground State Simulations in Quantum Chemistry
- Growth driven phase transitions in Zinc Oxide nanoparticles through machine-learning assisted simulations
- Toward Chemical Accuracy in Biomolecular Simulations through Data-Driven Many-Body Potentials: III. A Library of Amino Acid Building Blocks
BibTeX
@misc{lopez2026scalable,
title = {{A scalable framework for goal-driven task-based simulations}},
author = {Carlos Lopez and King Chun Lai and Markus Rampp and Christoph Scheurer},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008845/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008845/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - A scalable framework for goal-driven task-based simulations
AU - Carlos Lopez
AU - King Chun Lai
AU - Markus Rampp
AU - Christoph Scheurer
PY - 2026
DA - 2026/09/15
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008845/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008845/v1
AB - Realistic surfaces of active materials that involve the motion of atoms or ions in their function typically change their (sub-)surface composition and geometry under operating conditions due to varying environmental stimuli, degradation or healing processes, but also just while exhibiting their intended function, making it difficult to build simulation models that remain valid as the system evolves. Such problems often demand large-scale computation together with efficient data processing, so scalable task execution and data-efficient synchronization become critical. We present a scalable computational framework that supports large collections of independent computation tasks with incremental result aggregation. We demonstrate the framework using adaptive kinetic Monte Carlo (akMC) as a showcase application, where transition-state searches are performed on-the-fly. The new framework employs a hierarchical orchestrator–worker architecture implemented with modern MPI constructs and operates efficiently from single workstations to multi-node high-performance computing systems. It integrates well-established simulation engines such as LAMMPS and supports multi-core CPU and GPU parallelism within each task. Performance is demonstrated using classical, machine-learning, and GPU-accelerated interatomic potentials. Benchmarks show strong scaling across multiple compute nodes with several hundreds of CPU-cores, efficient utilization of multi-core executors and GPU sharing. The results indicate that the framework provides an efficient solution for rare-event and mesoscopic simulation workflows.
ER -