Raju Rajesh K
National Research Council Canada, Clean Energy Innovation (CEI) Research Centre, Mississauga, Ontario L5K 1B4, Canada.
School of Chemistry, University of Birmingham, Birmingham B15 2TT, U.K.
J Phys Chem A. 2024 Oct 24;128(42):9122-9134. doi: 10.1021/acs.jpca.4c04416. Epub 2024 Oct 13.
The search for global minimum (GM) configurations in nanoclusters is complicated by intricate potential energy landscapes replete with numerous local minima. The complexity of these landscapes escalates with increasing cluster size and compositional diversity. Evolutionary algorithms, such as genetic algorithms, are hampered by slow convergence rates and a propensity for prematurely settling on suboptimal solutions. Likewise, the basin hopping technique faces difficulties in navigating these complex landscapes effectively, particularly at larger scales. These challenges highlight the need for more sophisticated methodologies to efficiently scan the potential energy surfaces of nanoclusters. In response, our research has developed a novel deep reinforcement learning (DRL) framework specifically designed to explore the potential energy surfaces (PES) of nanoclusters, aiming to identify the GM configurations along with other low-energy states. This study demonstrates the framework's effectiveness in managing various nanocluster types, including both mono- and multimetallic compositions, and its proficiency in navigating complex energy landscapes. The model is characterized by remarkable adaptability and sustained efficiency, even as cluster sizes and feature vector dimensions increase. The demonstrated adaptability of DRL in this context underscores its considerable potential in materials science, particularly for the efficient discovery and optimization of novel nanomaterials. To the best of our knowledge, this is the first DRL framework designed for the GM search in nanoclusters, representing a significant innovation in the field.
在纳米团簇中寻找全局最小（GM）构型会因错综复杂的势能景观而变得复杂，这些景观充满了众多局部最小值。随着团簇尺寸的增加和组成多样性的提高，这些景观的复杂性也会升级。进化算法，如遗传算法，受到收敛速度缓慢和倾向于过早陷入次优解的阻碍。同样，盆地跳跃技术在有效导航这些复杂景观方面也面临困难，尤其是在更大尺度上。这些挑战凸显了需要更复杂的方法来有效扫描纳米团簇的势能面。作为回应，我们的研究开发了一种新颖的深度强化学习（DRL）框架，专门设计用于探索纳米团簇的势能面（PES），旨在识别GM构型以及其他低能态。这项研究证明了该框架在处理各种纳米团簇类型（包括单金属和多金属组成）方面的有效性，以及其在复杂能量景观中导航的能力。即使团簇尺寸和特征向量维度增加，该模型仍具有显著的适应性和持续效率。在这种情况下，DRL所展示的适应性凸显了其在材料科学中的巨大潜力，特别是在新型纳米材料的高效发现和优化方面。据我们所知，这是第一个为纳米团簇中的GM搜索设计的DRL框架，代表了该领域的一项重大创新。
