收录解读
- 分级：`颠覆性` - 正式标题：`BEACONS: Bounded-Error, Algebraically-Composable Neural Solvers for Partial Differential Equations` - 原文：`2026-02-16-P6_BEACONS-BEACONS_Bounded_Error_Algebraically_Composable_Neural_Solvers_for_Partial_Differ.pdf` - 抽取：`extracted.md`
## 重写摘要
这篇论文瞄准的是 AI for PDE 最致命的缺陷：神经网络求解器速度很快，但一旦进入训练分布之外，往往既不能保证稳定性，也不能保证误差边界，更谈不上用于高风险科学计算。BEACONS 的核心做法是把“神经求解器”与“形式化可验证性质”绑定在一起，构造带有严格收敛性、稳定性与守恒性质的 PDE 神经架构，使其在外推区间也具有机器可检查的正确性保证。
论文特别强调 bounded-error 和 algebraic composability。前者意味着模型输出不是只有经验分数，而是有严格误差控制；后者意味着不同模块可以像代数对象那样组合，而不会在组合过程中丢失验证性质。对长期困扰科学计算的“快但不可信”问题，这是一条很硬的路线。
## 为什么重要
如果这条路线站得住，AI for PDE 就不再只是数值替代器，而会进入“可认证科学软件”范畴。这对等离子体、流体、核工程和国防级仿真都非常关键，因为这些领域真正缺的不是再快一点，而是快且可证明地可靠。
## 局限
目前是 `2026-02-16` 的 arXiv 预印本。形式化保证往往伴随更强结构假设和更高建模复杂度，实际覆盖的 PDE 类型、边界条件和工程规模，仍要看后续扩展。
原始摘要与中文对照
中文对照翻译
摘要 神经网络架构在可靠地实质性泛化到其训练数据的凸包之外方面存在传统局限性，这给计算物理学带来了重大问题，因为在计算物理学中，人们通常希望在远远超出任何可以通过实验或分析验证的范围内求解偏微分方程（PDEs）。在本文中，我们展示了如何通过构建用于求解PDEs的经过形式验证的神经网络架构来规避这些局限性，这些架构具有严格的收敛性、稳定性和守恒性，因此即使在外推区域也能保证其正确性。通过使用特征线法先验地预测PDE解的分析性质（即使在距离训练域任意远的区域），我们展示了如何为浅层神经网络近似的最坏情况L∞误差构建严格的外推界限。接着，通过将PDE解分解为更简单函数的组合，我们展示了如何将这些浅层神经网络近似组合起来形成深度架构，这借鉴了组合式深度学习的思想，其中通过将不连续函数的近似与更平滑的函数组合，抑制了其近似中较大的L∞误差（从而推广了传统数值求解器中通量限制器的理论）。所产生的框架，称为BEACONS（有界误差、代数可组合神经求解器），既包含用于神经求解器本身的自动代码生成器，也包含一个定制的自动化定理证明系统，用于生成机器可验证的正确性证书。我们将BEACONS框架应用于各种线性和非线性PDEs，包括线性平流方程和无粘Burgers方程，以及完整的可压缩欧拉方程，涵盖1D和2D情况，并说明了与同等规模的传统神经网络架构相比，BEACONS架构如何能够以可靠且有界的方式将解外推到远远超出训练数据的范围。本文讨论了BEACONS方法相对于经典PINN（物理信息神经网络）方法的各种优势。
原始摘要
A BSTRACT The traditional limitations of neural network architectures in reliably generalizing substantively beyond the convex hulls of their training data present a significant problem for computational physics, in which one often wishes to solve partial differential equations (PDEs) in regimes far beyond anything which can be experimentally or analytically validated. In this paper, we demonstrate how it is possible to circumvent these limitations by constructing formally-verified neural network architectures for the solution of PDEs, with rigorous convergence, stability, and conservation properties, whose correctness can therefore be guaranteed even in extrapolatory regimes. By using the method of characteristics to predict the analytical properties of PDE solutions a priori (even in regions arbitrarily far from the training domain), we show how it is possible to construct rigorous extrapolatory bounds on the worst-case L∞ errors of shallow neural network approximations. Then, by decomposing PDE solutions into compositions of simpler functions, we show how it is possible to compose these shallow neural network approximations together to form deep architectures, building on ideas from compositional deep learning, in which the large L∞ errors in the approximations of discontinuous functions have been suppressed by composing them with smoother ones (thus generalizing the theory of flux limiters in traditional numerical solvers). The resulting framework, called BEACONS (Bounded-Error, Algebraically-COmposable Neural Solvers), comprises both an automatic code- generator for the neural solvers themselves, as well as a bespoke automated theorem-proving system for producing machine-checkable certificates of correctness. We apply the BEACONS framework to a variety of linear and non-linear PDEs, including the linear advection and inviscid Burgers’ equations, as well as the full compressible Euler equations, in both 1D and 2D, and illustrate how BEACONS architectures are able to extrapolate solutions far beyond the training data in a reliable and bounded way, as compared to conventional neural network architectures of equivalent size. Various advantages of the BEACONS approach over the classical PINN (Physics-Informed Neural Network) approach are discussed.