Supporting the OpenMM Community-led Development of Next-Generation Condensed Matter Modelling Software

原子模拟是高性能计算研究的主要应用，并越来越多地支持化学和生命科学行业的创新研发流程。OpenMM是当前开源学术软件生态系统中增长最快的原子模拟引擎。OpenMM最初针对生物分子模拟受众，用户群呈指数级增长，并已渗透到各种相关领域，包括材料建模，量子化学，结构生物信息学，化学信息学，人工智能和机器学习。OpenMM的成功归功于一种设计，它在可扩展性（通过强大的用户界面）和GPU性能（通过自动生成的CUDA内核）之间实现了良好的平衡，用于分子动力学（MD）模拟。OpenMM可单独使用或通过插件连接到其他原子仿真引擎，为整个原子仿真生态系统提供GPU加速的MD仿真功能。我们对OpenMM用户社区进行了调查，以确定其最迫切的需求。OpenMM目前由单个核心开发人员维护，该开发人员无法再支持其快速增长的用户社区的培训和支持需求。我们将把OpenMM过渡到一个更可持续的社区驱动的发展模式。我们将开发培训资源来提高用户的技能，并让社区参与开发和维护OpenMM功能。机器学习（ML）潜力有可能彻底改变原子模拟方法的未来。我们的社区调查发现了对ANI神经网络和GAP高斯过程回归方法的浓厚兴趣。我们将在OpenMM中提供一个自包含的GPU优化的GAP实现，并与致力于OpenMM ANI实现的项目合作伙伴协调，为社区提供一个可以随时插入现有仿真引擎的ML潜力库。（ML潜力）和技术（增加的硬件异构性）驱动程序在未来十年内继续为用户群提供速度和易于修改之间的优化权衡。我们将在OpenMM中集成一个多级中间表示编译器（MLIR），以从用户指定的Python指令自动生成针对不同硬件的优化低级代码。通过使用户能够将自定义原子特征化技术指定为OpenMM操作，可以与Tensorflow或Pytorch操作进行精细交织，我们将OpenMM定位为首选的模拟引擎，以支持将下一代ML潜力部署到当前的GPU和新兴的AI硬件加速器上。我们的社区还需要支持，以促进独立开发的OpenMM软件解决方案与更广泛的原子仿真生态系统中的其他软件的组合使用。这项研究将开发一个标准化的接口，将OpenMM社区软件与CCPBioSim的可互操作Python框架BioSimSpace集成。我们将通过生产GAP ML管道来展示这项研究的所有工作包的集成，这两个用例针对软凝聚态建模中的巨大挑战（有机催化-最近获得2021年诺贝尔化学奖-和蛋白质-配体结合）。总之，这项研究将使OpenMM用户社区处于下一代混合机器学习的最前沿。软凝聚态建模的分子力学势。与人工智能和HPC社区的更深入集成将为原子模拟铺平道路，以利用新兴的百亿亿次机会。从单一开发人员过渡到社区驱动的开发治理模型将提高代码库的可持续性，并鼓励相关学术社区和行业更多地采用OpenMM。

Atomistic simulations are the main application of high-performance computing research, and increasingly underpin innovative R&D processes in the chemical and life sciences industry. OpenMM is the fastest growing atomistic simulation engine among the current ecosystem of open-source academic software. Originally targeting a biomolecular simulation audience, the OpenMM user base is growing exponentially and has permeated diverse related domains, including materials modelling, quantum chemistry, structural bioinformatics, chemoinformatics, artificial intelligence and machine learning. The success of OpenMM is down to a design that achieves an excellent tradeoff between extensibility (via a robust user interface) and performance on GPUs (via auto generated CUDA kernels) for molecular dynamics (MD) simulations. OpenMM is used standalone or via plugins to other atomistic simulation engines, providing access to GPU-accelerated MD simulation capabilities for the whole atomistic simulation ecosystem.We have surveyed the OpenMM user community to identify its most pressing needs. OpenMM is currently maintained by a single core developer who can no longer support the training and support needs of its rapidly growing user community. We will transition OpenMM to a more sustainable community-driven development model. We will develop training resources to upskill users, and engage the community to widen participation in developing and maintaining OpenMM functionality. Machine learning (ML) potentials have the potential to revolutionise the future of atomistic simulation methodologies. Our community survey has identified strong interest in ANI neural network and GAP Gaussian process regression methods. We will deliver a self-contained GPU-optimised GAP implementation in OpenMM and coordinate with project partners working on an OpenMM ANI implementation to offer the community a library of ML potentials that can be readily plugged into existing simulation engines.OpenMM must adapt to scientific (ML potentials) and technological (increased hardware heterogeneity) drivers to continue offering its user base an optimised tradeoff between speed and ease of modification over the coming decade. We will integrate in OpenMM a multiple level intermediate representation compiler (MLIR) to auto generate from user-specified Python instructions optimised low-level code targeting diverse hardware. By enabling users to specify custom atomic featurisation techniques as OpenMM operations, which can be finely interleaved with Tensorflow or Pytorch operations, we will position OpenMM as the simulation engine of choice to support deployment of next generation ML potentials onto current GPUs and emerging AI-hardware accelerators. Our community has also required support to facilitate the combined use of independently developed OpenMM software solutions with other software from the broader atomistic simulation ecosystem. This research will develop a standardised interface to integrate OpenMM community software with CCPBioSim's interoperable Python framework BioSimSpace. We will demonstrate integration of all the work packages of this research via production of GAP ML pipelines for two use cases that target grand challenges in soft-condensed matter modeling (organocatalysis - recently recognised by the 2021 Nobel Prize in Chemistry- and protein-ligand binding).Altogether this research will position the OpenMM user community at the forefront of next-generation hybrid machine learning/molecular mechanics potentials for soft-condensed matter modelling. Deeper integrations with AI and HPC communities will pave the way for atomistic simulations to harness emerging exascale opportunities. Transitioning from a single developer to a community-driven development governance model will improve sustainability of the codebase and encourage greater adoption of OpenMM in associated academic communities and industry.