TurboAgent: An LLM-Driven Autonomous Multi-Agent Framework for Turbomachinery Aerodynamic Design

Abstract: The aerodynamic design of turbomachinery is a complex and tightly coupled multi-stage process involving geometry generation, performance prediction, optimization, and high-fidelity physical validation. Existing intelligent design approaches typically focus on individual stages or rely on loosely coupled pipelines, making fully autonomous end-to-end design challenging. To address this issue, this study proposes TurboAgent, a LLM-driven autonomous multi-agent framework for turbomachinery aerodynamic design and optimization. The LLM serves as the core for task planning and coordination, while specialized agents handle generative design, rapid performance prediction, multi-objective optimization, and physics-based validation. The framework transforms traditional trial-and-error design into a data-driven collaborative workflow, with high-fidelity simulations retained for final verification. A transonic single-rotor compressor is used for validation. The results show strong agreement between target performance, generated designs, and CFD simulations. The coefficients of determination for mass flow rate, total pressure ratio, and isentropic efficiency all exceed 0.91, with normalized RMSE values below 8%. The optimization agent further improves isentropic efficiency by 1.61% and total pressure ratio by 3.02%. The complete workflow can be executed within approximately 30 minutes under parallel computing. These results demonstrate that TurboAgent enables an autonomous closed-loop design process from natural language requirements to final design generation, providing an efficient and scalable paradigm for turbomachinery aerodynamic design.

• The paper presents an LLM-driven multi-agent framework that autonomously integrates generative inverse design, surrogate performance prediction, and physics validation for turbomachinery aerodynamic design.
• It employs a cDDPM for diverse blade geometry generation and transformer-based models achieving R² > 0.99 in performance prediction, while LLM optimization improves efficiency and design outcomes.
• Experimental results demonstrate a 95% success rate in CFD validation and a complete design cycle reduced to approximately 30 minutes, evidencing a practical shift from traditional methods.

TurboAgent: Autonomous Multi-Agent Framework for Turbomachinery Aerodynamic Design

Motivation and Context

TurboAgent proposes an LLM-driven framework to address the high complexity, nonlinear coupling, and multi-stage dependencies inherent in turbomachinery aerodynamic design. Traditional methods, dominated by expert-led forward/inverse design, simulation-driven iteration, and reliance on surrogate models, suffer from inefficiency, high computational cost, and limited explicit knowledge reuse. While generative AI—GANs, VAEs, DPMs—have enabled rapid geometry generation and performance mapping, most prior work remains siloed at single stages or consists of rigid, static tool integrations. The critical gap is the absence of an autonomous, flexible, closed-loop design system capable of dynamic cross-stage planning and collaborative execution, with final physical validation. TurboAgent responds by introducing a multi-agent, LLM-planned paradigm, tightly integrating generative inverse design, surrogate performance prediction, agentic optimization, and automated high-fidelity simulation.

System Architecture

TurboAgent leverages the LangGraph agent framework, supporting graph-structured orchestration of tasks and tool invocation. The architecture comprises distinct functional agents:

• Task Planning Agent: LLM-powered, parses natural language requirements, plans global workflow, dynamically adapts task ordering and invokes appropriate specialized agents.
• Generative Design Agent: Conditional Denoising Diffusion Probabilistic Model (cDDPM)-driven, produces diverse, high-quality blade geometries parametrized by performance targets (mass flow $m$, total pressure ratio $T_t$, isentropic efficiency $\eta$), with transformation to 3D CAD models for downstream validation.
• Performance Prediction Agent: Transformer-based surrogate, enables millisecond-level evaluation of key aerodynamic metrics from design parameters, facilitating rapid candidate screening and sensitivity analysis.
• Optimization Agent: Integrates meta-prompt-driven LLM optimization with traditional algorithms (GA, PSO); supports adaptive search strategies, multi-objective reward formulation, and semantic constraint handling.
• Physics Validation Agent: Automates CFD and FEA via natural language interaction, configuring mesh generation, boundary conditions, and physical solver execution; enables batch parallel simulation analysis.
• Knowledge Synthesis Agent: Aggregates outputs, supports domain-specific Q&A, report generation, and semantic summarization for interpretability.

A Flask-based front-end ensures seamless user interaction, visualization, and real-time feedback.

Methodological Foundations

Parameterization and Database

An industrial compressor blade dataset underpins the generative and prediction agents. Blade geometry is parameterized using 21 key variables across hub, mid-span, and tip sections, based on NURBS representations. This enables high-dimensional inverse design and supports effective surrogate modeling for performance prediction.

Generative Design via cDDPM

The generative agent employs a cDDPM with 1D U-Net backbone, guided by hybrid condition vectors incorporating physical targets and geometric priors. Sampling and model inference are fast, supporting interactive design sessions and automated conversion to 3D CAD geometry for simulation.

Surrogate Performance Prediction

Transformer encoders capture global dependencies among geometric parameters, achieving high-fidelity surrogate predictions for $m$, $T_t$, and $\eta$, thus bypassing expensive CFD for preliminary screening.

LLM-Driven Multi-Objective Optimization

The optimization agent utilizes meta-prompts embedding design variables, targets, and constraints in natural language, inferring probabilistic mean vectors for candidate generation. Design vectors are sampled from $N(\mathbf{x}, \sigma^2 I)$ distributions, with adaptive variance and iterative feedback. Scalar rewards aggregate weighted performance scores and normalized constraint penalties.

Automated Physics Validation

The physics validation agent interfaces with NUMECA and ANSYS, automating mesh generation and solver configuration. CFD utilizes SST turbulence models; FEA supports material property modification and stress analysis. Key physical quantities and convergence histories are extracted and visualized.

Experimental Validation and Results

Task Planning Evaluation

TurboAgent demonstrates robust semantic reasoning and logical workflow construction, mapping natural language queries into executable pipelines with conditional branching, multi-agent tasking, and feedback-driven iteration.

Generative Design and Diversity

For 100 out-of-distribution targets, the generative agent produces 10 diverse designs per target; geometric profiles are physically plausible and satisfy constraints. Interactive requirement completion is supported.

Performance Prediction Accuracy

Surrogate evaluations on generated candidates yield $R^2 > 0.99$, nRMSE $<2\%$ for all metrics—mass flow, pressure ratio, efficiency—indicative of highly accurate rapid assessment.

Physics Validation Consistency

Automated CFD on 124 valid cases (95% success rate) demonstrates strong agreement: $R^2$ values for $T_t$0, $T_t$1, and $T_t$2 reach 0.9775, 0.9795, and 0.9158; nRMSE $T_t$3. Surrogate predictions are validated against full simulation; ARE distributions corroborate the fidelity.

Optimization Performance

Meta-prompt LLM optimization outperforms GA and PSO in convergence speed and reward attainment. Notably, isentropic efficiency increases by 1.61% and total pressure ratio by 3.02% relative to baseline. Optimization is adaptive, does not require explicit gradient computation, and operates within the unified multi-agent workflow.

Closed-Loop Workflow Operation

A full compressor design case validates end-to-end workflow. TurboAgent autonomously executes requirement interpretation, generation, evaluation, optimization, high-fidelity simulation, and final trade-off analysis. Four best candidates are selected based on aerodynamic, structural, and stability criteria; final recommendations integrate CFD, FEA, and off-design performance.

Computational Cost

Total token usage per workflow is approximately 80,000–100,000. End-to-end execution under parallelism (30 cores) completes in ~30 minutes, an order-of-magnitude reduction compared to traditional workflows spanning weeks.

Theoretical and Practical Implications

TurboAgent establishes a scalable autonomous workflow, transitioning turbomachinery design from expert-driven, simulation-heavy loops to LLM-planned, collaborative agent execution. The system demonstrates that unified orchestration and closed-loop validation can facilitate rapid, diverse candidate generation, efficient screening, and robust final selection, while preserving high-fidelity physical constraints. The integration of LLM-driven optimization provides evidence that semantic understanding and adaptive search can practically accelerate multi-objective engineering tasks.

The framework promotes enhanced generalization, knowledge reuse, and interpretability. Further research should focus on mitigating data dependency, transferring across disparate configurations, and increasing decision transparency in highly complex or multidisciplinary design scenarios. Future developments may extend toward real-time, cross-domain engineering tasks, integrating broader multimodal input and output modalities.

Conclusion

TurboAgent delineates a comprehensive LLM-driven autonomous multi-agent system, achieving efficient, scalable, and physically valid turbomachinery aerodynamic design. It significantly reduces design cycle time, increases candidate diversity, and improves optimization efficacy. The framework signals an evolution from “model-assisted” to “autonomous intelligent” design practices, offering broad applicability to high-complexity engineering problems. As LLMs and agentic AI mature, such integrated systems will likely redefine workflows across computational design, optimization, and physical validation in aerospace and beyond (2604.06747).