发表机构
Center for the Transformation of Chemistry; Max Planck Institute of Colloids and Interfaces（化学转化中心; 马克斯·普朗克胶体与界面研究所）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
该研究提出PINN-Phase物理信息神经网络时间积分器，用于多晶微结构的曲率驱动多相场演化，在多类基准测试中实现了低不一致率的长范围预测与初始条件迁移，提升了模拟效率。
AI 中文摘要
当需要对多晶微结构的大量相关案例进行长时间演化时，其相场模拟会变得成本高昂。我们提出PINN-Phase，一种物理信息神经网络时间积分器，该积分器从初始条件推进完整的多相场，并在每一步强制相位界限和单位和；在报告的显式多相场基准测试中，初始条件后的参考状态仅用于评估。无需针对具体案例进行调整，一个经过训练的含25个晶粒的模型在评估前固定的8个 unseen 微结构中的7个以及两种应力案例中均满足所有预定义标准，10个案例的最终晶粒标签不一致率为0.94%-3.71%。每12000步的 rollout 在单个GPU上耗时约5.2分钟，达到训练期间所代表的时间范围的近三倍。一个预先注册的含64个晶粒的模型达到6.09%的不一致率，保留了全部21个参考存活晶粒外加1个额外晶粒；评估后的延续通过将训练范围加倍并增加25个 epoch，达到3.97%的不一致率和准确的存活晶粒集。在三维中，一个经过训练的含16个晶粒、分辨率为96³的模型在6个 unseen 微结构中恢复了准确的最终活动集和全部三个灭绝恒等式，其中5个满足完整的预定义资格。这些结果证明了在固定基准家族内的结构可容许长范围预测和预期初始条件迁移。
英文摘要
Phase-field simulation of polycrystalline microstructures becomes costly when many related cases must be evolved over long times. We introduce PINN-Phase, a physics-informed neural time integrator that advances the full multiphase field from its initial condition and enforces phase bounds and unit sum at every step. On the reported explicit multiphase-field path, it learns from the governing residual on its own predicted states, without post-initial-condition reference states in the loss or checkpoint selection. A fixed 25-grain model satisfies the complete predefined criteria in seven of eight prospectively fixed unseen microstructures and in both deliberately atypical stress cases, with 0.94-3.71% terminal label disagreement. At native 128^3 resolution, an eight-grain model trained on four initial microstructures retains 98.79% label agreement and exact survivor identities on a training case after 24,000 autonomous steps, 5.86 times its represented training horizon. On six prospectively fixed unseen 128^3 initial conditions, label disagreement at the last saved state within that horizon is 0.72-2.11%; after autonomous evolution to 24,000 steps, or 5.86 times the represented training horizon, five of six retain the exact terminal survivor set. At 64 grains, terminal disagreement decreases from 6.09% in the primary evaluation to 3.97% after continued training, with exact recovery of the 21-grain survivor set. These results demonstrate structurally admissible long-horizon neural integration and direct initial-condition transfer within fixed benchmark families.
Comments53 pages, 17 figures, 3 tables; Supplementary Material available as an ancillary file (52 pages, 13 figures, 22 tables)
