发表机构
BCMaterials, Basque Center for Materials, Applications, and Nanostructures, UPV/EHU; University of the Basque Country (UPV/EHU); Ikerbasque Basque Foundation for Science（巴斯克材料、应用与纳米结构中心（BCMaterials）; 巴斯克大学; 伊克拉巴斯克科学基金会）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究基准测试五种图神经网络机器学习原子间势在CdSe纳米团簇上的动力学稳定性，发现仅凭预测误差不足以选择势函数，需结合长时标动力学测试与不确定性引导的主动学习。
AI 中文摘要
机器学习原子间势（MLIPs）能够实现无机半导体纳米晶的纳秒级原子模拟，但在留出构型上的低误差并不一定保证分子动力学的稳定性。我们对五种图神经网络MLIPs——SchNet、PaiNN、NequIP、Allegro和MACE——在一个含149个原子的氯钝化硒化镉纳米团簇上的动力学稳定性进行了基准测试。这些模型在统一条件下使用1,000个构型进行训练，并针对使用密度泛函理论生成的2,000个留出构型进行评估，考虑了力精度、计算效率、不确定性以及300 K下1纳秒模拟中的结构稳定性。Allegro产生了最低的验证力平均绝对误差，范围在每埃12至14毫电子伏特之间，而SchNet产生的误差最大，范围在每埃84至110毫电子伏特之间。这一排名并未预测动力学鲁棒性：NequIP在没有额外训练数据的情况下保持了1纳秒的稳定，而MACE仅在基于集成的主动学习程序添加了34个不确定性选择的构型后才变得稳定。PaiNN和Allegro在分别添加100和95个构型后仍不稳定，SchNet在测试的增强预算内也未能实现稳定动力学。在基准硬件条件下，机器学习势每个分子动力学步骤需要6至66毫秒，而密度泛函理论大约需要20秒。这些结果表明，仅凭留出预测误差不足以选择用于有限、表面主导的纳米结构的原子间势。可靠的部署需要结合长时标动力学测试与不确定性引导的细化，而主动学习的有效性和数据效率仍高度依赖于架构。
英文摘要
Machine-learning interatomic potentials (MLIPs) enable nanosecond-scale atomistic simulations of inorganic semiconductor nanocrystals, but low errors on held-out configurations do not necessarily guarantee stable molecular dynamics. We benchmark five graph-neural-network MLIPs, SchNet, PaiNN, NequIP, Allegro and MACE, for dynamical stability in a chloride-passivated cadmium selenide nanocluster containing 149 atoms. The models were trained under harmonized conditions on 1,000 configurations and evaluated against 2,000 held-out configurations generated using density functional theory, considering force accuracy, computational efficiency, uncertainty and structural stability during 1 ns simulations at 300 K. Allegro produced the lowest validation force mean absolute errors, ranging from 12 to 14 meV per angstrom, whereas SchNet produced the largest, ranging from 84 to 110 meV per angstrom. This ranking did not predict dynamical robustness: NequIP remained stable for 1 ns without additional training data, whereas MACE became stable only after an ensemble-based active-learning procedure added 34 uncertainty-selected configurations. PaiNN and Allegro remained unstable after the addition of 100 and 95 configurations, respectively, and SchNet also failed to achieve stable dynamics within the tested augmentation budget. Under the benchmark hardware conditions, the machine-learning potentials required 6-66 ms per molecular-dynamics step, compared with approximately 20 s for density functional theory. These results show that held-out prediction errors alone are insufficient for selecting interatomic potentials for finite, surface-dominated nanostructures. Reliable deployment requires long-timescale dynamical testing combined with uncertainty-guided refinement, while the effectiveness and data efficiency of active learning remain strongly architecture dependent.
Comments15 pages, 5 Figures, 2 Tabls
