【AIAA】Generative AI in AIAA Conference Publications 2024–2025：2026/2/22

Executive summary

Across AIAA conference publications in 2024–2025, “generative AI” appears in two dominant clusters: (i) generative geometry / design-space construction (diffusion, GAN/AAE/VAE, DDPM-conditioned synthesis) and (ii) LLM-centered engineering workflows (RAG, domain fine-tuning, evaluation/benchmarks, agentic orchestration). A third cluster is AI-accelerated CFD / solvers, where “generative” is sometimes literal (diffusion/flow reconstruction) but more often practical: physics-informed neural operators/solvers, shock-processing ML, and super-resolution–informed AMR. Representative design-generation works include diffusion airfoil samplers and latent diffusion conditioned on flow constraints ¹ and DDPM-based synthesis for full-aircraft configurations (e.g., blended-wing-body geometry generation conditioned on aerodynamic targets). ²

On the LLM side, the 2024–2025 AIAA set spans (a) domain LLMs and fine-tuning for aviation text and QA, ³ (b) RAG and knowledge-graph grounding/trust, ⁴ (c) benchmarks and evaluation sets for aerospace/aviation language use, ⁵ and (d) agentic / workflow engines that translate natural-language intent into executable engineering pipelines. ⁶

For industrial turbomachinery implications, the strongest near-term pattern is not “diffusion-designed turbine blades” (still sparse in 2024–2025 AIAA open literature), but tooling: LLM-mediated orchestration of meshing/CFD/optimization and provenance-aware workflow engines (both relevant to turbine aero/thermal design stacks). ⁷ The major technical gap, relative to industrial TRL, is closed-loop, constraint-verified generative design (geometry + operability + aero/thermal/structural constraints + manufacturability) with defensible validation beyond simulation. ⁸

Scope & method

This report targets AIAA conference/meeting papers in 2024–2025 (SciTech; Aviation/ASCEND; and adjacent AIAA meeting-paper venues) that explicitly involve generative AI (diffusion/DDPM/score-based; GAN/AAE/VAE; normalizing flows), LLMs/RAG/agentic systems, and surrogate/physics-informed models when used for generative or solver-acceleration roles. Evidence was prioritized in this order: (1) accessible AIAA/ARC pages and PDFs, (2) author/institution PDFs (e.g., École polytechnique fédérale de Lausanne Infoscience; NASA Technical Reports Server PDFs; German Aerospace Center eLib), and (3) arXiv preprints when they correspond to the AIAA DOI-labeled paper. When a field could not be validated from a cited primary source, it is marked “unspecified.”

Interpretation rule for evidence in the inventory table: for each paper, all non-“unspecified” extracted attributes are derived from the row’s Evidence sources; if the evidence does not support an attribute, it is left “unspecified.”

Paper inventory table

Primary categories (one each): generative design; physics-informed generative models; surrogate modeling; multi-agent design automation; AI-driven CFD acceleration; LLM-assisted workflows; digital twin + generative.

Abbreviations: Phys-int = black-box (BB) vs physics-embedded (PE). Val = simulation / experiment / both / unspecified. TRL = research / prototype / applied (inferred from validation + deployment cues; where inference is made, it is justified). Geom = direct / parameterized / NA. CFD-loop = yes/no/unspecified. Agents = yes/no/unspecified.

Institution summary table

Counting rule: each institution is counted once per paper when the affiliation is explicitly visible in the evidence sources used in this report. Because not all AIAA PDFs are openly accessible, counts below reflect only papers with confirmable affiliations from cited sources (they should be read as lower bounds).

Institution	Confirmed paper count	Evidence
NASA Ames Research Center	1	3
NASA Langley Research Center	4	47
German Aerospace Center (DLR)	3	58
University of Southampton	1	4
École polytechnique fédérale de Lausanne	2	1
Iowa State University	1	7
University of Iowa	1	53
Missouri University of Science and Technology	1	8
University of Michigan	1	43
Stargazer Design Technologies	1	29

Cross-paper analysis and trends

The inventory indicates a slight year-over-year increase from 2024 to 2025 in the subset captured by the keyword-based sweep (25 → 30 papers). This is largely driven by growth in explicit AI-driven CFD application papers and multi-agent / operator-agent framing, while “LLM-assisted workflows” remains steady in raw count. ⁵⁹

Category share is dominated by LLM-assisted workflows (benchmarks, RAG, fine-tuning, tool-support), followed by generative design (GAN/DDPM/diffusion geometry or design-space synthesis) and a near-tie between surrogate modeling and AI-driven CFD acceleration. ⁴

Category share (primary category) in the identified 2024–2025 set

A consistent technical pattern in geometry-generative papers is conditioning on performance targets (lift/drag/moment; constraints) while maintaining geometric validity. Diffusion-based airfoil sampling highlights data-efficiency and latent validity constraints ¹; DDPM for BWB synthesis similarly conditions on global aero coefficients. ⁶⁰ Physics-constrained GAN methods push further by explicitly attempting to map into a feasible region for constraints (e.g., eVTOL takeoff profiles), illustrating an emerging direction: “generative model as constraint-satisfying reparameterization.” ⁵⁶

For turbomachinery-adjacent workflows, the most actionable 2024–2025 thread is LLM/agent orchestration of high-fidelity analysis: a Model Context Protocol server coupled to an LLM client to drive geometry creation, meshing, simulation, optimization, and visualization in a conversational loop. ⁷ This is complemented by provenance-aware, token-window-aware workflow management with open-weight agents. ⁴⁵ Together, these suggest that near-term turbine value may come from automation, reproducibility, and human-in-the-loop acceleration, rather than fully autonomous blade-shape diffusion in the open literature.

Autonomous multi-agent turbine design architecture

mermaidコピーするflowchart TD
    A[Requirements & constraints intake\n(certification, operability, manufacturability)] --> B[RAG/Knowledge base\nstandards, past designs, test reports]
    B --> C[LLM Planner / Orchestrator\n(tool calling, decomposition)]
    A --> C

    C --> D[Geometry agent\n(blade/vane/endwall/cooling)]
    C --> E[Meshing agent\n(topology, y+, quality, automation)]
    C --> F[Solver agent\n(RANS/URANS/LES + adjoint)]
    C --> G[Surrogate/ROM agent\n(VAE/NO/NN; UQ-aware)]
    C --> H[Optimizer agent\n(MDO, constraints, robust design)]

    D --> E --> F --> I[Results store + provenance]
    F --> I
    G --> H --> I

    I --> J[Verification & guardrails\nconstraint checking, plausibility, hallucination filters]
    J --> C

    I --> K[Validation loop\nrig tests, cascades, hot-gas tests]
    K --> I

    subgraph "Where surveyed papers map"
      C1[(2024-2773)]:::p --> C
      C2[(2024-0917, 2025-0701, 2025-1791)]:::p --> C
      B1[(2025-0700, 2025-3442, 2025-1916)]:::p --> B
      D1[(2024-2013, 2024-3755, 2025-3292)]:::p --> D
      G1[(2024-0685/0686, 2025-3799)]:::p --> G
      F1[(2025-0695/0696/0699, 2025-1467)]:::p --> F
      H1[(2025-3802, 2025-2775)]:::p --> H
    end

    subgraph "Gaps to reach industrial TRL"
      X1[End-to-end constraint guarantees\n(geometry+flow+stress+thermal+manufacturing)]:::gap
      X2[UQ + robustness as first-class objective\nacross multi-fidelity]:::gap
      X3[Closed-loop experimental validation integration\n(test stands + data assimilation)]:::gap
      X4[Certification-grade traceability\n(provenance, audit trails, governance)]:::gap
      X5[Operator-agent safety\n(LLM reliability under time-critical ops)]:::gap
    end

    J --> X5
    H --> X1
    G --> X2
    K --> X3
    I --> X4

    classDef p fill:#eef,stroke:#66f,stroke-width:1px;
    classDef gap fill:#fee,stroke:#f66,stroke-width:1px;

Interpretation: surveyed work already covers (1) agentic orchestration of analysis pipelines (not turbine-specific but directly transferable), ⁷ (2) RAG/graph grounding and trust as a technical counterweight to hallucination, ⁴ (3) generative parameterizations for airfoils and aircraft configurations, ¹ and (4) solver acceleration via physics-informed methods and ML-in-the-loop AMR. ⁶¹ The gaps are primarily about verification + validation + governance and about scaling to multi-physics turbine realities.

Prompt analysis and improved prompts

Analysis of the general research prompt

Your general prompt is strong in its taxonomy (categories) and extraction schema (physics-integration, validation type, CFD-in-loop, agent coordination). The main weaknesses that reduce retrieval depth and reproducibility are:

First, it needs a clear inclusion/exclusion boundary for “surrogate models,” because that term is too broad in AIAA proceedings; without guardrails, the search explodes into hundreds of generic ML surrogate papers. (In this report, we constrained surrogates to those explicitly using generative/latent models, physics-informed operators, or solver-acceleration framing.) ⁶²

Second, it should explicitly request searching by the AIAA DOI pattern 10.2514/6.2024-* and 10.2514/6.2025-* and by meeting paper IDs (AIAA 2024-XXXX, AIAA 2025-XXXX) because those strings appear in many AIAA-accessible metadata lines and in downstream repositories. ³¹

Third, the validation dimension should split “simulation” into at least: (a) pure offline dataset experiments, (b) simulator-in-the-loop optimization, and (c) high-fidelity CFD/adjoint loops. This matters because papers range from data-only language modeling ³ to end-to-end CFD + FEA optimization pipelines with an LLM front-end. ⁷

Analysis of the turbomachinery-focused prompt

Reconstructing the likely goal: a turbomachinery prompt should force discovery of papers that touch (i) geometry generation for blades/vanes/endwalls, (ii) aero/thermal/structural constraints, (iii) CFD/adjoint integration, (iv) experimental validation (cascades, rigs), and (v) workflow automation at scale.

The largest missing dimension is turbomachinery-specific vocabulary (compressor/turbine stage, endwall contouring, tip clearance, film cooling, secondary flows, loss models, operability/stall margin, blade heat transfer, conjugate heat transfer, stress/creep/fatigue) combined with generative keywords. Without that, search drifts toward airfoils and generalized aerospace geometry.

The second missing dimension is CAD/mesh representation (NURBS, splines, implicit SDF, mesh deformation, topology constraints, watertight CAD) which is central for turbine manufacturing and solver robustness. That omission prevents separating “pretty generative shapes” from “CFD-usable and manufacturable geometries.”

Improved prompt wording

Improved general prompt (retrieval + extraction)

Search AIAA meeting papers from 2024–2025 (SciTech; Aviation/ASCEND; Propulsion & Energy; and other AIAA meeting-paper venues) for generative AI and agentic workflows. Use DOI patterns 10.2514/6.2024-* and 10.2514/6.2025-* plus keywords: diffusion/DDPM/score-based, GAN/cGAN/AAE/VAE, normalizing flow, neural operator, PINN/PINO, transformer generation, LLM, RAG, knowledge graph, tool calling, multi-agent, workflow engine, digital twin.
For each confirmed paper, return: title; authors; year; venue; DOI; (if available) author PDF/arXiv/NASA/AFRL/MIT link. Extract (with evidence quotes or “unspecified”): model type + conditioning, geometry representation, physics coupling (PDE residuals, constraints, adjoint, solver-in-loop), validation tier (offline dataset vs CFD/adjoint vs experiment), uncertainty handling, deployment maturity signals (prototype system, integrated tool, operator study), and whether multi-agent coordination exists (planner/executor/critic; provenance).
Classify into: generative design; physics-informed generative models; surrogate modeling; multi-agent design automation; AI-driven CFD acceleration; LLM-assisted workflows; digital twin + generative. Require a structured table and trend summaries.

Improved turbomachinery-focused prompt (for AIAA submission targeting)

Focus on turbomachinery (compressor/turbine/fan) and adjacent propulsion aero-thermal design. Search AIAA 2024–2025 meeting papers using combined keyword blocks:
(A) Generative/agentic: diffusion/DDPM/score-based, GAN/cGAN/AAE/VAE, flow matching, normalizing flow, transformer generation, RL policy generation, LLM/RAG/agentic/tool calling, multi-agent.
(B) Turbomachinery: turbine stage, compressor stage, blade/vane, endwall contouring, tip clearance, film cooling, secondary flows, loss, efficiency, stall margin, operability, heat transfer, conjugate heat transfer, cooling passage, stress/creep/fatigue, manufacturing constraints.
(C) Representations: CAD, NURBS/splines, CST, Bézier, SDF/implicit, mesh deformation, graph meshes, unstructured mesh quality, y+.
For each paper found: extract whether geometry is direct vs parameterized; whether CFD/adjoint is in-the-loop; whether constraints are enforced in-model (physics-informed / feasibility mapping) versus post-check; whether experiments are included (cascade/rig/hot gas); whether UQ is reported; and whether an agent team coordinates CAE steps with provenance/audit trails. Output: comparison table, institution map, and identified gaps for industrial TRL.

Strategic insights and research directions for industrial turbine applications

The 2024–2025 AIAA evidence suggests a pragmatic path for turbines: agentic automation + provenance + solver acceleration is maturing faster than end-to-end generative blade synthesis. Workflow engines explicitly target the difficulty of reliably converting natural language into correct engineering actions at scale, including multi-agent approaches and context management—highly relevant to turbine design environments where a single “design iteration” spans many tools and constraints. ⁴⁵ In parallel, physics-informed solvers and ML-augmented CFD components are moving toward integration inside simulation workflows (e.g., shock processing, super-resolution–guided AMR), which is directly applicable to compressor and turbine flow predictions. ³⁶

However, industrial adoption hinges on three missing capabilities: (1) constraint guarantees that cover aero/thermal/structural/manufacturing simultaneously (beyond “feasible in surrogate space”), ⁸ (2) rigorous, experiment-linked validation loops, and (3) certification-grade traceability and reliability modeling for agentic systems—especially as AIAA papers themselves caution about reliability in time-critical settings. ⁴⁶

Five AIAA-suitable research directions and paper ideas

Feasibility-Mapped Diffusion for Turbine Blade Sections with Manufacturing Constraints
Concept: Build a conditional diffusion model generating turbine blade sections (or 2.5D stacks) conditioned on loading and thickness/cooling/clearance constraints, but train it to output samples in a feasible manifold akin to the “feasible-space transform” logic demonstrated for physicsGAN-style constraints. ⁶³
Required simulations/experiments: RANS (and selected URANS) evaluations across operating envelope; manufacturing constraint checks (minimum radii, thickness bounds); optional cascade tests for a small subset.
TRL target: prototype (integrated generator + verification + CFD evaluation loop).

Multi-Agent Turbine Design Orchestrator with Provenance and Audit Trails
Concept: Extend provenance-aware workflow management with an LLM planner/executor/critic team to run a full turbine iteration: CAD update → meshing → solver run → post-processing → constraint report → optimization step. Use the MCP-style tool-calling pattern and provenance strategies from AIAA workflow management papers. ⁷
Required simulations/experiments: Demonstrate on a public/industrial-like turbine stage case with repeated runs; quantify failure modes, recovery strategies, and reproducibility.
TRL target: prototype (tool integrated; operator-in-the-loop).

Physics-Informed Neural Operator Surrogates for Turbine Stage Flowfields with Uncertainty Quantification
Concept: Adapt physics-informed transformer/neural-operator techniques to turbine passages (high curvature, secondary flows), explicitly measuring generalization to new boundary conditions and geometry perturbations. ³³
Required simulations/experiments: LES/RANS multi-fidelity dataset for a reduced turbine stage set; incorporate UQ (ensembles or Bayesian approximations) to quantify risk for design decisions.
TRL target: research → prototype depending on integration into a design loop.

Trust-Grounded Turbine “Design Copilot” Using GraphRAG + Believability Metrics
Concept: Combine domain RAG over turbine design rules, test reports, and CFD best practices with intrinsic trust/believability scoring methods (embedding–knowledge graph alignment) to mitigate hallucination, and evaluate under adversarial prompts (“wrong units,” “unsafe setting”). ⁴⁷
Required simulations/experiments: Controlled evaluation set (like ALUE-style but turbine-specific) with ground-truth answers and penalized hallucinations; user study with engineers. ⁵
TRL target: prototype (internal decision-support tool).

Hybrid Generative Cooling Passage Design with CFD-in-the-Loop AMR Acceleration
Concept: Use generative models (VAE/GAN/diffusion) to propose internal cooling passage topologies under manufacturability limits, and couple evaluation to CFD using super-resolution–assisted AMR to reduce turnaround time. ⁶⁴
Required simulations/experiments: Conjugate heat transfer simulations with realistic boundary conditions; validate select designs via lab tests (e.g., scaled cooling coupons).
TRL target: prototype (demonstrated speedup + validated thermal metrics).

A publicly available paper (PDF)

Following only 2 papaers are avairable.

• Physics-Constrained Generative Artificial Intelligence for Rapid Takeoff Trajectory Design
• deepSPACE: Generative AI for Configuration Design Space Exploration