TL;DR
FastAgent is a physics-aware campaign controller that bridges the gap between rapid screening and high-fidelity Computational Fluid Dynamics (CFD). By intelligently allocating computational resources between a lightweight engine (FastFluent) and heavy-duty solvers based on specific design goals, it accelerates complex industrial design tasks by up to 15x while significantly reducing errors compared to standard automation tools.
What Problem Does This Paper Address?
Industrial CFD is notoriously expensive. When engineers need to evaluate complex geometries, long transient processes, and coupled physics (like heat transfer and phase change) across a massive design space, the computational bill becomes prohibitive.
Traditional "agents" or automation scripts often lack a deep understanding of the underlying physics. They might automate a workflow, but they can't decide if a coarse, fast simulation is "good enough" for a specific question or if a high-fidelity run is strictly necessary. This paper tackles the "campaign-level" gap: how to explore thousands of design permutations efficiently without sacrificing the physical accuracy needed for final industrial certification.
Methods
The authors introduce a two-part framework:
- FastFluent: A purpose-built, lightweight numerical engine. It uses Lattice-Boltzmann (LBM) for flow calculations and reduced-order thermal models for rapid screening. It provides "numerical evidence"—not just results, but also diagnostics on conservation and computational cost.
- FastAgent: The decision-making "brain." It uses an evidence-driven policy to manage four coupled choices: the physical model, numerical resolution, simulation horizon, and which candidates to "escalate" to high-fidelity solvers like Ansys Fluent or structural solvers like Abaqus.
A critical feature is the Discrepancy Update. When a high-fidelity simulation is performed, FastAgent compares it to the FastFluent prediction. This discrepancy is used to update the agent's internal ranking and refine its policy for subsequent decisions in the campaign.
Datasets
The framework was validated across two main categories:
- Public CFD Benchmarks: Six cases including laminar channel flow, backward-facing steps (transient diagnosis), heated channels (thermal transfer), and lid-driven cavities. These also included reproductions of classic literature cases like the Ghia cavity and Schäfer–Turek cylinder.
- Industrial Case Study: High-pressure steam dewaxing in investment casting. This involved a combinatorial design space of 131,072 possible outlet layouts for a turbine blade shell.
Key Insights and Ideas
- QoI-Dependent Fidelity: The agent recognizes that different "Quantities of Interest" (QoIs) require different levels of effort. For example, in dewaxing, it uses the rapid engine to locate the "risk window" (when the shell is most likely to crack) but reserves the high-fidelity solver to confirm exactly when the wax has finished melting.
- Mechanism-Guided Design: During the startup phase of dewaxing, FastAgent identified that nonuniform condensation heating—rather than direct steam pressure—was the dominant mechanism driving early shell failure. This allowed the optimization to focus on thermal management.
- Combinatorial Reach: By using FastFluent to screen the entire 131k+ layout space, the agent could identify high-performing candidates that a human engineer or a purely high-fidelity workflow would likely miss due to time constraints.
Key Results
- Efficiency: FastAgent achieved speedups of 1.53x to 15.36x relative to a matched "Generic Agent" control across benchmark campaigns.
- Accuracy: It reduced the normalized worst-QoI error by 5.68% to 99.14%.
- Industrial Impact: In the steam dewaxing case study, the "rapid route" required 81.39% fewer cell-time updates than a standard fine-grid reference.
- Design Optimization: The agent identified a specific outlet layout ("Sparse-6") that reduced simulated crack opening by 96.0% and connected crack extent by 66.7% compared to the original industrial design.
Limitations and Future Work
- Geometric Complexity: While the framework handles complex industrial geometries, the authors note that commercial confidentiality prevents the public release of the specific turbine blade models used.
- Physics Scope: The current FastFluent engine is optimized for flow and thermal phase change. Extending this to more diverse multiphysics (like chemical reactions or complex combustion) would require expanding the lightweight engine's capabilities.
- LLM Integration: The agent currently utilizes a server-side Codex model for reasoning. The authors suggest that future iterations could benefit from more specialized, locally-hosted scientific LLMs to further refine the "rationale" behind simulation decisions.
Summary updated