FastAgent: A Physics-Aware FastFluent-to-Fluent Workflow for Rapid CFD Screening and Industrial Steam Dewaxing Design
Authors: Chen YanKai, Yu QiuHeng, Yuan Man, Zou Ting, Wu Lin, He Jian, Zhang QiongYuan, Jing Tao, Xu QingYan
Organizations: School of Materials Science and Engineering, Tsinghua University, Beijing, China · State Key Laboratory of Clean and Efficient Turbomachinery Power Equipment, Deyang 618000, China · Dongfang Electric Corporation Dongfang Turbine Co., Ltd., Deyang 618000, China
Industrial computational fluid dynamics (CFD) becomes costly when complex geometries, long transients and coupled physics must be evaluated across large design spaces. We introduce FastAgent, a physics-aware campaign controller built around FastFluent, a lightweight numerical engine for rapid physical screening. FastFluent supplies stage-specific fields and diagnostics; FastAgent allocates physical models, numerical resolution and simulation horizons by the active quantities of interest (QoIs), while returned Fluent discrepancies update subsequent decisions. Across six public CFD benchmark campaigns, the framework reduces normalised worst-QoI error by 5.68-99.14% and accelerates decision-relevant computation by factors of 1.53-15.36 relative to a matched Generic Agent control. Applied to high-pressure steam dewaxing in investment casting, FastAgent identifies nonuniform condensation heating as the dominant startup mechanism and separates rapid risk-window localisation from high-fidelity melt-completion confirmation. The rapid route requires 81.39% fewer cell-time updates than a fine-grid, small-time-step reference. Structural analysis localises fracture at the observed blade edge. Screening 131,072 outlet layouts and selectively evaluating candidates with Fluent and Abaqus identifies Sparse-6, which reduces simulated crack opening by 96.0% and connected crack extent by 66.7% relative to the Original design. These results demonstrate how rapid numerical evidence and adaptive fidelity control connect large CFD search spaces to mechanism-guided industrial design.
Figures & tables
Case
Physical test
Agent/FastFluent decision
Key result
Supported conclusion
Laminar channel
Refinement and mass balance
Diagnose and repair the inlet before high-fidelity confirmation
RMS 0.008656 ; speedup >2.595×
Numerical correction
Backward-facing step
Separated-flow reattachment
Screen three tiers and return a non-stationary result
2.655× acceleration; snapshot deviation 0.168%
Rapid screening with diagnostic return
Heated channel
Blocked thermal transport
Correct porous mapping through front and dose monitors
Errors 0.1956%/0.5847% ; speedup 1.710×
Stage-specific QoI transfer
Lid-driven cavity
Full-field confirmation
Transfer the velocity field and correct numerical settings
Ghia RMS 0.003452 ; speedup 1.283×
High-fidelity full-field confirmation
Ghia cavity reproduction
Re=100 – 1000 profiles
Screen nine candidates and confirm three
Mean RMS 0.0095285 ; speedup 2.512×
Discrepancy-driven correction
Schäfer–Turek cylinder
Unsteady wake at Re=100
Stop after consecutive multi-QoI criteria are satisfied
Strouhal error 0.672% ; speedup 2.037×
Adaptive confirmation
Table 1: FastAgent decisions and principal evidence across six CFD campaigns. The cases test numerical correction, transient diagnosis, QoI transfer, full-field confirmation, discrepancy-driven refinement and adaptive stopping. Reported speedups use the case-specific solver baselines in Fig. 2 b.
Figure 2: FastAgent validation across six public CFD benchmark campaigns. a , Representative error metrics or diagnostic deviations for the six cases. b , Measured speedups relative to expert-defined solver baselines. c , Reduction in normalised worst-QoI error and decision-relevant computational speedup relative to the matched Generic Agent control. Dashed lines mark the 5% accuracy-improvement and 1.25× speedup criteria. All six cases satisfy both. Labels BM01–BM04 denote the laminar channel, backward-facing step, heated channel and lid-driven cavity. RP01 and RP02 denote the Ghia cavity and Schäfer–Turek cylinder reproductions. The laminar-channel values are lower bounds, and the cylinder comparisons begin from a shared 4.0 s checkpoint.
Figure 3: FastAgent-guided startup screening identifies thermal nonuniformity as the dominant early mechanism. a , Time-resolved steam heat input, equivalent 5 mm shell heating, spatial temperature spread and combined stress proxy for six cases varying orientation and distance. b , Cross-case relations connecting heat dose to mean shell heating, temperature nonuniformity to the thermal-stress contribution and the combined stress proxy to the crack-driving index. c,d , Final frontal and side stress-proxy maps. e , Inlet-diameter sensitivity of shell heating, impact impulse and peak crack-driving index.
Figure 4: FastAgent-guided full-cycle fidelity allocation separates risk localisation from melt-completion confirmation. a , Three-round search of 16 candidates and selection of a stable candidate for confirmation. b , Timing accuracy of the selected FastFluent candidate against Fluent. c , Quantity-specific evaluation of risk timing by rapid calculation and melt completion by Fluent.
Figure 5: FastAgent-guided structural analysis links dewaxing loads to blade-edge fracture. a , Thermal and pressure states converted into regional shell loading and compared with independently calculated ProCAST contact pressure. b , Resulting maximum-principal-stress localisation and the observed blade-edge shell crack. c , XFEM crack initiation and edge-aligned growth. d , Mechanical histories describing stress, crack connectivity, opening, damage and energy.
Figure 6: FastAgent-guided outlet-layout design balances drainage performance and local fracture safety. a , Complete FastFluent design-space screening followed by selective high-fidelity CFD and structural analysis. b , Global ranking agreement across the eleven Fluent-evaluated layouts for the baseline, rank-calibrated and physics-informed stages (R6.1, R7R and R8M3). The fracture-aware point (R9F) instead reports local-risk agreement across the three structurally evaluated layouts. c , Consistent local-risk ordering from Fluent critical-region exposure and Abaqus crack opening; R11 denotes the exposure metric. d , Upper-8, Sparse-6 and Dense-15 layouts, labelled T2, V1 and T8 in the figure. e–g , Cohesive damage, activation delay and reductions in damage, opening and connected extent relative to the Original design.
Recent LLM-based agents have closed substantial portions of the scientific discovery loop in software-only machine-learning research, in chemistry, and in biology. Extending the same loop to high-fidelity physical simulators is harder, because solver completion does not imply physical validity and many failure modes appear only in field-level imagery rather than in solver logs. We present AI CFD Scientist, an open-source AI scientist for computational fluid dynamics (CFD) that, to our knowledge, is the first to span literature-grounded ideation, validated execution, vision-based physics verification, source-code modification, and figure-grounded writing within a single inspectable workflow. Three coupled pathways cover parameter sweeps within a fixed solver, case-local C++ library compilation for new physical models, and open-ended hypothesis search against a reference comparator, all running on OpenFOAM through Foam-Agent. At the center of the framework is a vision-language physics-verification gate that inspects rendered flow fields before any result is accepted, rerun, or written into a manuscript. On five tasks under a shared GPT-5.5 backbone, AI CFD Scientist autonomously discovers a Spalart-Allmaras runtime correction that reduces lower-wall Cf RMSE against DNS by 7.89% on the periodic hill at Reh=5600; under matched LLM cost, two strong general AI-scientist baselines (ARIS, DeepScientist) execute partial CFD workflows but lack the domain-specific validity gates needed to convert runs into defensible scientific claims; and a controlled planted-failure ablation shows that the vision-language gate detects 14 of 16 silent failures missed by solver-level checks. Code, prompts, and run artifacts are released at https://github.com/csml-rpi/cfd-scientist.
High-fidelity computational fluid dynamics (CFD) is widely used for thermal-fluid design, but repeated CFD solves remain expensive for design optimization, uncertainty analysis, and digital-twin workflows. Recently, our team has demonstrated that a proper orthogonal decomposition and neural-network (POD-NN) surrogate can predict two-dimensional thermal fields in an electronics-cooling cold plate with large inference speedups while preserving physically interpretable modal structure. Reproducing and extending such workflows, however, typically requires custom scripts for parameter sampling, Fluent automation, data extraction, reduced-order model construction, neural-network training, validation, and prediction. This paper introduces CFDTwin, an open-source Python package and optional desktop graphical user interface (GUI) that packages these steps into a reusable workflow for ANSYS Fluent simulations. CFDTwin allows users to define simulation inputs and output quantities, generate design-of-experiments samples, run and resume Fluent batch simulations, train POD-NN surrogate models for scalar, surface-field, and cell-zone outputs, inspect validation metrics, and evaluate trained models at new design points without re-running Fluent. The same workflow is exposed through a scriptable Python API and a GUI, supporting reproducible studies, user-facing model validation, and automated design exploration. CFDTwin extends the prior POD-NN modeling study from a case-specific research implementation to a reusable research-software platform for CFD surrogate modeling and digital-twin development.
Daniel Curl, Han Hu
Department of Mechanical Engineering, University of Arkansas, Fayetteville, AR 72701, USA
Computer-aided design (CAD) is the backbone of modern industrial design, yet learned CAD generators still fall short of real engineering pipelines: they neither iterate like engineers nor evaluate what engineering requires. Prior work has treated CAD generation as two disjoint steps, part synthesis and assembly, where the former is graded by proximity to a gold reference and the latter, when handled at all, is reduced to a separate constraint solving step. In this work, we introduce a more industry-native task formulation that requires a model to produce a fully assembled multi-part STEP file from a free-form engineering brief, which is then validated via finite element analysis (FEA). FEA validation reveals that Codex (GPT-5.5) and Claude Code (Opus-4.7) agents do not produce a single strict-passing artifact in the main first-attempt sweep, with the best configuration meeting only about 20% of typed requirements on average. Moreover, we introduce two additional supervision signals, a novel text-only blueprint schema and a 21-view image renderer that aids the agent's visual inspection, that better align the generation loop with how engineers iterate in practice. On S2O and Fusion360, the same feedback tools improve geometric reconstruction, with GPT-5.5/xhigh rising from 0.444 to 0.592 Box-IoU on S2O and from 0.397 to 0.505 on Fusion360. Together these signals move CAD programs toward artifacts that are not only visually plausible but also checked against physical and structural requirements.
Guijin Son, Jehyun Park, Seyeon Park +2
Seoul National University · OneLineAI · Sungkyunkwan University