用量子轨道信息统一建模分子，速度快且精度高。
OrbitAll: A Unified Quantum Mechanical Representation Deep Learning Framework for All Molecular Systems
- 融合量子轨道与图神经网络，物理可解释性强。
- 用1/10数据达化学精度，比DFT快1000~10000倍。
- 适合复杂、带电、溶剂环境分子的快速模拟。
我们提出OrbitAll，一种几何与物理信息融合的深度学习框架，通过底层量子力学方法提取的自旋极化轨道特征，结合SE(3)等变图神经网络，对任意电荷、自旋及环境效应下的分子系统进行编码。该框架在预测带电、开壳层和溶剂化分子时表现优异，能稳健外推至远超训练数据规模的分子体系。OrbitAll仅需竞品模型1/10的训练数据即可达到化学精度，相比密度泛函理论（DFT）提速约10³~10⁴倍。在化学多样性数据集上训练后，其在挑战性分子系统中表现稳健，即使仅使用35倍更少的分子数据和50倍更小的模型，仍优于基础机器学习势能模型UMA，尤其在高度带电体系中。学习溶剂效应后，其计算溶剂依赖反应路径的成本仅为显式溶剂模拟的1/100，效率显著提升。
原文摘要 · Abstract (English)
We introduce OrbitAll, a geometry- and physics-informed deep learning framework that encodes any molecular system with arbitrary charges, spins, and environmental effects using electronic structure information. It utilizes spin-polarized orbital features from the underlying quantum mechanical method and combines them with SE(3)-equivariant graph neural networks. OrbitAll demonstrates superior performance and generalization in predicting charged, open-shell, and solvated molecules, and robustly extrapolates to molecules significantly larger than the training data. OrbitAll achieves chemical accuracy using 10 times fewer training data than competing AI models, with approximately $10^3$ - $10^4$ speedup compared to density functional theory. Trained on a chemically diverse dataset, OrbitAll performs robustly on challenging molecular systems, and outperforms the foundational machine-learned interatomic potential, UMA, for highly charged species, despite using 35 times less molecular data and a 50-times-smaller model. After learning solvent effects, it accurately predicts solvent-dependent reaction pathways at about 100 times lower cost than explicit-solvation simulations using UMA.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。