发表机构
Johns Hopkins University; University of Tennessee, Knoxville; Northwestern University（约翰斯·霍普金斯大学; 田纳西大学诺克斯维尔分校; 西北大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
ALIGNN 2.0是一个纯PyTorch实现的统一线图神经网络框架，通过单一图结构整合标量、光谱、张量、逐原子和力场预测，在JARVIS-Leaderboard上26/30基准领先，并支持材料筛选、逆向设计、光谱模拟和显微镜成像。
AI 中文摘要
图神经网络是材料性质预测和机器学习原子间势的核心，然而它们对专用图库的依赖阻碍了可移植性和可复现性，并且性质和力场模型历来需要独立的图处理流程。我们提出了ALIGNN 2.0，这是原子线图神经网络的一个无依赖、纯PyTorch重新实现，其线图及其批处理从头构建，可在当前一代加速器上运行，并将标量、光谱、张量、逐原子和力场预测统一在单个图之后，据我们所知，现有框架没有提供这种组合。比较半径图和k近邻（kNN）图，更宽的kNN图在性质预测上更准确，而平滑变化的半径图是能量守恒分子动力学所必需的。在JARVIS-Leaderboard上，ALIGNN 2.0在30个单性质基准中的26个上领先于原始ALIGNN，在压电和介电最大值、剥离能、模量和超导转变温度Tc方面有大幅提升。与LAMMPS和OpenMM兼容的ALIGNN-FF在Matbench-Discovery和CHIPS-FF基准上以极小比例的参数量匹配领先的通用势，同时可扩展到十万原子规模的晶胞。我们进一步将ALIGNN 2.0用作条件晶体扩散模型中的去噪器，其中显式线图消息传递持续降低结构去噪损失。我们还展示了作为进行中工作，独立扩散的冗余键角状态是可学习的，但不能均匀地改善重建，也不能与线图相加性地结合。最后，从单个弛豫结构，同一框架重建的红外、拉曼、光学介电和中子谱与实验和密度泛函理论（DFT）一致，并驱动冻结声子电子显微镜图像模拟。
英文摘要
Graph neural networks are central to materials property prediction and machine-learning interatomic potentials, yet their reliance on specialized graph libraries hampers portability and reproducibility, and property and force-field models have historically required separate graph pipelines. We present ALIGNN 2.0, a dependency-free, pure-PyTorch reimplementation of the Atomistic Line Graph Neural Network, with the line graph and its batching built from scratch, running on current-generation accelerators and unifying scalar, spectral, tensorial, per-atom, and force-field prediction behind a single graph, a combination that to our knowledge no existing framework provides. Comparing radius and k-nearest-neighbor (kNN) graphs, the wider kNN graph is more accurate for properties while the smoothly varying radius graph is required for energy-conserving molecular dynamics. On the JARVIS-Leaderboard, ALIGNN 2.0 leads on 26 of 30 single-property benchmarks against the original ALIGNN, with large gains for piezoelectric and dielectric maxima, exfoliation energy, moduli, and superconducting Tc. The LAMMPS- and OpenMM-compatible ALIGNN-FF matches leading universal potentials on the Matbench-Discovery and CHIPS-FF benchmarks at a small fraction of their parameters while scaling to hundred-thousand-atom cells. We further use ALIGNN 2.0 as the denoiser in a conditional crystal-diffusion model, where explicit line-graph message passing consistently lowers structural denoising loss. We also show, as work in progress, that an independently diffused, redundant bond-angle state is learnable but does not uniformly improve reconstruction or combine additively with the line graph. Finally, from a single relaxed structure the same framework reconstructs infrared, Raman, optical-dielectric, and neutron spectra in agreement with experiment and DFT, and drives frozen-phonon electron-microscopy image simulation.
