PDEFlow: Autonomous Agentic PDE Pipelines for Neural Operator Learning and Solver-Free Inference

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Akshat Jani, Prathamesh Gadekar, Sakhinana Sagar Srinivas, Venkataramana Runkana
• Venue: arXiv: cs.LG
• Published: 2026-07-06
• Collected: 2026-07-07
• URL: http://arxiv.org/abs/2607.05134v1
• PDF: https://arxiv.org/pdf/2607.05134v1
• DOI: UNKNOWN

Plain-Language Brief

The paper introduces PDEFlow, a framework that converts user-provided ODE and PDE descriptions into neural operator pipelines. It includes a stateful input graph to validate problem specifications, a data-generation module for parameter sampling and solving equations, a registry-based interface for different neural operators, and an initial implementation with a multi-branch Bayesian DeepONet. Experiments on benchmark ODE and PDE tasks demonstrate its ability to create valid specifications, generate solver-backed datasets, train neural operators, and provide solver-free predictions.

What the Document Says

• The paper presents PDEFlow, a framework that turns user-provided ODE and PDE descriptions into neural operator pipelines.
• It includes a stateful input graph for validating problem specifications.
• A data-generation module samples parameters, solves equations with FEniCSx, and stores solutions as operator-ready tensors.
• The framework uses a registry-based interface for different neural operators with the current implementation being a multi-branch Bayesian DeepONet.
• Experiments on benchmark ODE and PDE tasks demonstrate its ability to create valid specifications, generate solver-backed datasets, train neural operators, and provide solver-free predictions.

Key Concepts Mentioned

• PDEFlow
• ODE
• PDE descriptions
• neural operator pipelines
• stateful input graph
• data generation
• FEniCSx
• Bayesian DeepONet

Claims Made by the Document

• PDEFlow can construct valid specifications.
• PDEFlow can generate solver-backed datasets.
• PDEFlow can train neural operators across steady and transient problem classes.
• PDEFlow can provide solver-free predictions from saved checkpoints.

Evidence Used

• The text describes the experimental results on benchmark ODE and PDE tasks.

Methods / Approach

• PDE descriptions, neural operator pipelines, stateful input graph, data generation, FEniCSx, Bayesian DeepONet

Limitations Stated by the Authors

• UNKNOWN: Information about potential limitations as reported by the authors is not provided in the text.

Limitations Noted by the Agent

Agent interpretation — not from the source.

• The framework is currently designed for repeatable scientific and engineering workflows, which might limit its applicability to other domains.
• The initial implementation uses a multi-branch Bayesian DeepONet, which may not perform optimally for all types of problems.

Open Questions

• How well does PDEFlow perform compared to traditional solver methods?
• Can PDEFlow be extended to handle complex, non-linear PDEs?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• The paper does not provide a comparison with traditional solver methods.
• The text suggests that PDEFlow can handle steady and transient problem classes but does not specify its capabilities for complex, non-linear PDEs.

Raw Abstract / Excerpt

We present PDEFlow, an autonomous agentic framework that turns user-level ODE and PDE descriptions into solver-backed neural-operator pipelines. The workflow links problem specification, data generation, operator training, and checkpoint-based inference. A stateful input graph converts multi-turn natural-language input and user edits into validated problem specifications. The data-generation module then samples parameters, solves the configured governing-equation with FEniCSx finite-element backend, and stores the solutions as operator-ready tensors. The training and inference stages use a registry-based interface, allowing different neural operators to be trained and deployed without changing the surrounding pipeline. In the current implementation, we instantiate this interface with a multi-branch Bayesian DeepONet. Experiments on benchmark ODE and PDE tasks show that PDEFlow can construct valid specifications, generate solver-backed datasets, train neural operators across steady and transient problem classes, and provide solver-free predictions from saved checkpoints. The framework is designed for repeatable scientific and engineering workflows where many related physics configurations must be specified, simulated, learned, and queried with minimal manual intervention.