FastAgent uses rapid physics screening to target detailed simulations

Computational fluid dynamics, or CFD, predicts how fluids move and transfer heat. Industrial studies become expensive when engineers must compare many designs while following processes that involve melting, drainage and structural loading. A new research preprint introduces FastAgent, a controller that decides which calculations need more detail and which can remain inexpensive. Its results are reported research findings, not established industrial performance.

The system combines FastFluent, a lightweight numerical engine, with selective calculations in Fluent and structural analysis in Abaqus. The central idea is not to replace detailed simulation everywhere. It is to match computational effort to the particular engineering question being asked.

Choosing detail by the question

FastFluent uses different numerical formulations for different stages, including flow calculations and a reduced-order thermal model that represents melting. It returns physical fields, target measurements, conservation checks and computational cost.

FastAgent uses that evidence to choose the physical model, spatial resolution, time step, simulation duration and candidates requiring detailed evaluation. When Fluent disagrees with the rapid calculation, the discrepancy informs later rankings and simulation choices.

The paper calls the target measurements “quantities of interest.” Examples include the time of a loading peak, the point when wax finishes melting, or the opening of a crack. A cheap calculation may estimate one accurately while failing on another. FastAgent checks physical and numerical requirements before considering cost.

What the benchmarks show

The authors evaluated six public CFD campaigns covering channel flows, separated flow, heat transport, cavity circulation and an unsteady cylinder wake. Against a Generic Agent workflow, they report reductions of 5.68–99.14% in normalised worst-target error and decision-relevant computational speedups of 1.53–15.36 times. That error measure tracks the largest target error after scaling by its prescribed tolerance.

These comparisons evaluate the combined controller and rapid engine. The timed Generic Agent workflow did not have FastFluent and used fixed Fluent setups, so the speedups do not isolate the benefit of the decision policy alone. Separate comparisons used expert-defined Fluent schedules and produced different speedups.

A matched-evidence replay more directly examined decisions: both agents received the same progressively revealed simulation evidence. FastAgent matched all 19 prescribed actions, versus 18 for the Generic Agent. The difference concerned recognising that a separated-flow result was still evolving and needed rescreening. Each system used one independent session per case, limiting evidence about repeatability.

Steam dewaxing exposes the trade-offs

The industrial application concerns removing wax from ceramic shells used in investment casting. Steam heats the assembly; expanding wax can load the shell before melting and drainage release that load.

For the first 4.1 seconds, the simulations identified uneven condensation heating as a larger contributor to the adopted shell-response proxy than direct steam impact. Condensation-aware models also avoided the much higher chamber temperature predicted without condensation. These startup conclusions depend on the paper’s simplified conversion from thermal and pressure fields to shell response; they are not direct measurements of fracture.

Over the full 420-second cycle, the workflow separated locating a global screening-risk window from confirming complete melting. A rapid numerical route estimated the risk-window time with 0.044% error while requiring 81.39% fewer cell-time updates than a fine-grid, small-time-step reference. Cell-time updates measure numerical work, not elapsed runtime.

That route’s melt-completion error was 12.1%, exceeding the 6% acceptance criterion. All five tested resolution routes missed that endpoint criterion, so Fluent was retained for confirmation. Local structural calculations placed blade-edge cracking much earlier than the roughly 100.7-second global screening-risk peak. The two timings describe different phenomena, not interchangeable failure predictions.

Screening many layouts, checking a few

Seventeen possible outlet positions generated 131,072 layouts. FastFluent screened them all; eleven received Fluent evaluation, and three selected alternatives underwent Abaqus structural comparison.

The selected Sparse-6 layout balanced drainage utility and local fracture safety. At the common simulated endpoint of 8.88 seconds, its maximum crack opening was about 0.121 micrometres, compared with 3.00 micrometres for the Original design: a 96.0% reduction. Connected crack extent fell from 21 elements to seven, a 66.7% reduction. These are simulation outputs, not measured reductions in factory defects.

The calculations located damage at the observed blade-edge region, providing a spatial consistency check. However, only three alternative layouts received structural evaluation, and confidentiality prevents public release of the industrial geometries, commercial-solver models and proprietary process data. The preprint therefore supports a targeted computational strategy: explore broadly with inexpensive physics, then use detailed analysis where simplified models cannot adequately answer the question.