CAMLET: A Combined Ab-initio Manifold Learning Toolbox for Nanostructure Simulations

摘要分子动力学模拟是研究材料科学、纳米科学和生物学问题的有力工具。它自然为需要数学、统计学、计算机科学、物理学、材料和生物学知识的跨学科研究提供了大量机会。本项目的重点是开发基于学习的计算和统计方法，用于势能景观建模，以加速从头算分子动力学模拟。所开发的这套工具将大大扩展时间和系统规模的限制，而不会影响从头计算模拟结果的精度和质量。查宏源、杜强、李润泽和Jorge Sofo将研究学习和计算方法1)从从头计算模拟的轨迹通过流形学习来表征模拟真正发生的低维流形的局部和全局结构；2)在与从头计算模拟自然出现并在整个模拟过程中经常访问的构形空间中的区域对应的降维空间中识别和提取合适的团簇；3)使用具有惩罚似然性的高斯克里金模型进行有效的能量和力内插。在这个学习和计算框架中，插值势能面将被评估，当它的精度足够好时，它将取代昂贵的从头计算。随着模拟的发展，内插的势能面将被重新测试，以检测在模拟正在探索配置空间的新区域的情况下最终是否需要重新训练。

AbstractMolecular Dynamics simulations are powerfultools to study problems of materials science, nanoscience, and biology. It naturally providesample opportunities for interdisciplinary researchthat requires knowledge in mathematics, statistics, computer science, physics, materials and biology.The focus of this project is on developing learning-based computational and statistical methods for potential energy landscape modeling to accelerate ab-initio molecular dynamics simulations. The set of tools developed will substantially expand the limits of time and system size without compromising the precision andquality of the ab-initio simulation results.Hongyuan Zha, Qiang Du, Runze Li and Jorge Sofo willinvestigate learning andcomputational methods 1) to characterize both the local and globalstructures of the low-dimensional manifold in which the simulationreally occurs through manifold learning from the trajectories of theab-initio simulation; 2) to identify and extract suitable clusters inthe reduced dimension spaces corresponding to regions in theconfiguration space that naturally emerge from the ab-initiosimulation and are visited frequently by the particles throughout thesimulation; 3) to conduct efficient energy and force interpolationusing Gaussian Kriging models with penalized likelihood. In thislearning and computationalframework, the interpolated potential energy surface will beevaluated and it will replace the costly ab-initio evaluation when itsprecision is good enough. As the simulation evolves, the interpolatedpotential energy surface will be retested to detect theeventual need of a retraining in case the simulation is exploringnew regions of the configuration space.