AutoNumerics is a multi-agent framework that autonomously designs, implements, debugs, and verifies PDE numerical solvers from natural language. Advances AI in automating scientific computing and solver development.
AutoNumerics is a multi-agent framework that autonomously designs, implements, debugs, and verifies numerical solvers for partial differential equations (PDEs) directly from natural language descriptions, aiming to make scientific computing more accessible by reducing reliance on expert-level mathematical knowledge . Unlike neural network-based solvers such as physics-informed neural networks (PINNs) or operator-learning frameworks, AutoNumerics generates transparent, classical numerical schemes—such as finite difference or spectral methods—grounded in first principles of numerical analysis rather than producing black-box models .
The system employs a coarse-to-fine execution strategy to decouple logic debugging from stability validation: initial solver implementations are tested on low-resolution grids to efficiently identify and correct syntax errors or logical flaws before scaling to high-resolution simulations . To assess correctness in the absence of analytical solutions, AutoNumerics incorporates a residual-based self-verification mechanism that evaluates the PDE residual norms, enabling autonomous quality assurance . Additionally, a reasoning module proactively detects and filters ill-designed or numerically unstable solver configurations, such as those violating Courant–Friedrichs–Lewy (CFL) conditions, thereby improving robustness .
Experiments on 24 canonical and real-world PDE problems show that AutoNumerics achieves competitive or superior accuracy compared to existing neural and large language model (LLM)-based baselines, including CodePDE, U-Net, FNO, and PINNs . For instance, on the geometric mean of normalized root mean square error (nRMSE), AutoNumerics achieves $$9.00 \times 10^{-9}$$, significantly outperforming CodePDE ($$5.08 \times 10^{-3}$$) and other baselines across problems like Advection, Burgers, Reaction-Diffusion (React-Diff), Compressible Navier-Stokes (CNS), and Darcy flow . The framework also demonstrates intelligent scheme selection aligned with PDE structural properties—for example, choosing spectral methods for periodic domains and finite differences for Dirichlet boundary conditions .
Despite these advances, AutoNumerics currently shows limited accuracy on high-dimensional (≥5D) and high-order PDEs, and its evaluation is restricted to regular domains . The implementation is currently coupled to a single LLM (GPT-4.1), and while it generates interpretable code, it does not provide formal convergence or stability guarantees . Nevertheless, AutoNumerics represents a significant step toward fully automated, transparent, and reliable scientific computing pipelines powered by AI .
AutoNumerics introduces a novel multi-agent framework designed to fully automate the end-to-end development of numerical solvers for Partial Differential Equations (PDEs) directly from natural language descriptions. Unlike standard code-generation tools that often produce isolated snippets, this system employs a collaborative team of specialized AI agents to manage the entire scientific computing lifecycle. The pipeline autonomously navigates the complex stages of solver design, method selection (such as Finite Difference or Finite Element methods), code implementation, debugging, and rigorous verification. By decomposing the task into distinct roles—such as a mathematician to derive discretization schemes and a developer to write optimized code—the framework ensures that the final output is not just syntactically correct, but physically consistent and numerically stable.
A key contribution of this work is its "PDE-agnostic" architecture, which allows the system to adapt to a wide variety of physical domains without relying on hard-coded templates for specific equation types. The integration of a robust verification loop, which compares numerical results against analytical solutions or established benchmarks, addresses the critical issue of reliability in AI-generated scientific code. This advancement is significant because it lowers the barrier to entry for high-fidelity simulation, democratizing access to complex computational tools and accelerating research in physics and engineering. By demonstrating that AI agents can successfully orchestrate multi-step reasoning tasks involving specialized domain knowledge, AutoNumerics represents a substantial leap toward autonomous scientific discovery.
AutoNumerics introduces a groundbreaking multi-agent framework designed to autonomously generate, implement, debug, and verify numerical solvers for partial differential equations (PDEs) from natural language descriptions. The system operates in a PDE-agnostic manner, meaning it can handle a wide range of PDEs without requiring pre-defined solver templates. By leveraging large language models (LLMs) and agent-based collaboration, AutoNumerics automates the entire pipeline—from problem formulation to code generation, testing, and verification—thereby significantly reducing the manual effort traditionally required in scientific computing.
The key contributions of AutoNumerics include its ability to: 1. Automate Solver Development: The framework decomposes the solver design process into modular tasks (e.g., discretization, linear algebra, error analysis) and assigns them to specialized agents, enabling parallel and efficient problem-solving. 2. Handle Uncertainty and Corrections: The system incorporates feedback loops where agents can revise their outputs based on runtime errors or verification failures, improving robustness. 3. PDE-Agnostic Flexibility: Unlike traditional methods that rely on hand-coded solvers for specific PDEs, AutoNumerics dynamically adapts to new problems, making it scalable for research and industrial applications.
This work is particularly significant because it democratizes access to high-quality numerical solvers by reducing the barrier to entry for non-experts while also accelerating the development cycle for researchers. By integrating AI-driven automation into scientific computing, AutoNumerics could transform how PDEs are solved across fields like physics, engineering, and computational biology, potentially leading to faster discovery and more reliable simulations.