Source-linked AI summary
Closed-loop AI achieves certifiable engineering design
Tianyi Yu, Chengxing Tao, Haoxuan Shen, Huiyang Li, Rugang Chen, Long Teng, Lilin Wang, Yan Li, Qingbin Chen, Chaogang Xu, Lizhong Wang
TL;DR
Complex physical engineering design lacks an automated approach that can coordinate heterogeneous tools while satisfying simultaneous fluid, structural, stability, and practical constraints. The AI Engineer closes this loop by coupling LLM orchestration to deterministic optimization and simulation, with an internally calibrated review gate. Its certified foundation reduced steel intensity and unit capital cost by 8.1% versus TuQiang and passed CCS Approval in Principle.
Problem
Complex physical engineering design must satisfy simultaneous fluid-dynamics, solid-mechanics, and structural-stability constraints, but current workflows remain heavily dependent on human orchestration.
Method
The AI Engineer couples LLMs to deterministic engineering backends in a closed loop and uses an Automated Reviewer calibrated against real floating-wind projects to govern certification-ready termination.
Results
8.1% lower steel intensity and 8.1% lower unit capital cost than the TuQiang baseline were achieved while the 20 MW foundation passed CCS Approval in Principle.
Takeaways & Limitations
The results establish a verification-closed regime in which agentic orchestration is filtered by deterministic physics and codified limit states before designs proceed.
Takeaways & Limitations
Detailed design and construction remain human responsibilities because fabrication constraints are currently prompt-approximated rather than CAD-CAM exact.
Abstract
from arXiv · showhide
Agentic AI has automated parts of scientific discovery, including paper generation, expert-level coding, therapeutic proposal, and autonomous experimentation. Complex physical engineering design remains a gap, because candidates must satisfy simultaneous constraints in fluid dynamics, solid mechanics, and structural stability. We introduce The AI Engineer, an agentic framework that couples large language models (LLMs) to deterministic engineering backends in a closed loop: natural-language requirements are converted into design-domain geometry and mesh; topology is optimized with bi-directional evolutionary structural optimization (BESO) coupled to the CalculiX solver; and member sizes are refined with particle swarm optimization (PSO) coupled to Zwind under offshore aero-hydro-servo-elastic load cases. To explore many designs without per-candidate certification cost, an Automated Reviewer scores each candidate on five dimensions (capacity, steel intensity, unit cost, constructability, and fatigue life) using piecewise-linear functions calibrated on 11 real floating-wind projects. Search terminates only when a candidate reaches a composite score $S \ge 85$ (grade A) with no subscore below 60. We validated this gate by submitting the top-scoring design to the China Classification Society (CCS) for Approval in Principle (AIP), which it passed; AIP is thus an external check that the reviewer tracks professional judgment, not the daily objective. The certified design outperforms the human-optimized TuQiang baseline, reducing steel mass and unit capital cost by 8.1% each while meeting all AIP criteria. This verification-closed regime, in which every proposal is judged by deterministic physics and codified limit states, distinguishes The AI Engineer from open-ended generative systems. Remaining limits include detailed design and fabrication-hard constraints.
Introduction
Complex physical engineering design must satisfy coupled physics, manufacturability, cost, and reliability constraints, while current workflows remain human-intensive. The AI Engineer addresses this orchestration gap with a verification-closed loop connecting language requirements, deterministic solvers, optimization, and regulatory-calibrated review.
- Introduction: Floating offshore wind design balances energy capture, fatigue, stability, manufacturing, deployment, and cost under high-dimensional, non-convex constraints.High-fidelity design iterations can require hours or days on computing clusters.
- Introduction: Existing deterministic engineering solvers provide high-fidelity physics predictions, but manual translation and orchestration across CAD, meshing, and analysis tools remain bottlenecks.The limitation lies in coordinating heterogeneous tools rather than in the fidelity of individual solvers.
- Introduction: The AI Engineer converts natural-language requirements into structured specifications and autonomously orchestrates topology, size, simulation, drawing, and reporting stages.Its closed loop combines LLM orchestration with deterministic engineering backends and monitors solver diagnostics for replanning.
- Introduction: The system replans after mesh degradation, semantic divergence, or structural unity-margin violations instead of treating optimization as a fixed simulation pipeline.This addresses geometric, numerical, and physical failures during engineering execution.
- Introduction: The Automated Reviewer governs termination only after candidates satisfy five engineering dimensions and a certification-readiness threshold calibrated against regulatory benchmarks.The internal gate is externally checked through a one-time CCS Approval in Principle submission.
Results
The AI-generated floating-wind foundation reduced material and unit-cost metrics relative to the TuQiang baseline while satisfying dynamic, strength, fatigue, and regulatory requirements. Automated Reviewer scores were also aligned with independent regulatory evaluations, and the highest-scoring candidate passed CCS Approval in Principle.
- Results: 8.1% lower steel consumption and 8.1% lower unit capital cost were achieved versus the TuQiang baseline.The AI-generated foundation used 255.5 t·MW⁻¹ versus 278 t·MW⁻¹ and ¥18197 kW⁻¹ versus ¥19,800 kW⁻¹.
- Results: The AI-generated design avoided turbine-excitation resonance, maintained allowable strength unity checks, and kept fatigue damage within the 25-year design life.Extreme-sea-state platform motions were comparable to or lower than the baseline, with a Design Fatigue Factor of 3.0.
- Results: Automated Reviewer and regulatory scores showed Spearman correlation 0.717, mean absolute deviation 3.9 points, and 89% high agreement across the reference fleet.High agreement required both scores to be at least 60 with an absolute difference no greater than 15.
- Results: The highest-scoring candidate, with composite score S = 91, passed the China Classification Society’s Approval in Principle review.The submission followed the CCS review checklist and underwent preparation, application, independent appraisal, and final certification.
Discussion
The AI Engineer couples agentic orchestration with deterministic solvers, threshold monitoring, and failure-driven replanning to generate certifiable physical designs. Its FOWT demonstration passed CCS AIP and surpassed the TuQiang baseline, while broader deployment remains bounded by domain, validation, fabrication, reproducibility, and benchmarking constraints.
- Core contribution: The AI Engineer autonomously generates certifiable physical designs through deterministic simulation, rule-based threshold monitoring, and failure-driven replanning.The framework connects natural-language requirements to engineering workflows and uses hard physics and limit states to filter proposals.
- Demonstrated outcome: 8.1% lower steel intensity and 8.1% lower unit capital cost were achieved against the human-optimized TuQiang baseline while satisfying coupled limit-state constraints.The generated 20 MW floating offshore wind foundation passed CCS Approval in Principle.
- Verification-closed regime: Verification-closed operation evaluates every proposal against deterministic physics and codified limit states, with failures triggering numerically grounded replanning rather than textual caveats.This distinguishes the framework from scientific-discovery agents whose proposals require external experimental validation.
- Applicability to other engineering domains: The framework is transferable to other engineering domains when deterministic solvers, parameterized designs, and codified regulatory thresholds are available.Translation requires redefining geometry parameterization, solver toolchains, and threshold sets, with approximately three person-months reported for the FOWT implementation.
- Limitations and future directions: Commercial deployment remains limited by incomplete candidate ranking, unresolved extreme hydrodynamics, prompt-approximated fabrication constraints, LLM stochasticity, and moving human-optimized baselines.The paper identifies these as limitations requiring improved optimization, physical or CFD validation, manufacturing integration, reproducibility controls, and periodic benchmarking.
- Paradigm shift and ethical considerations: Transparent audit trails and immutable content hashes enable third-party reconstruction of the computational chain for safety-critical design submissions.The paper presents auditability as critical infrastructure for computational trust.
Methods
The AI Engineer uses schema-validated deterministic backends in a closed loop, with diagnostic-driven replanning and termination gates for certifiable physical design. The workflow combines topology optimization, geometric scaling, structural sizing, and independent reviewer validation.
- The AI Engineer orchestrator: Natural-language requirements become structured job descriptors, while schema validation restricts tool use and preserves state across design phases.The orchestrator validates solver outputs, monitors diagnostics, and replans after geometric, numerical, or physical-limit-state failures.
- Termination criteria: The loop terminates only when phase retries are inactive, sizing reaches n = 100 generations without meaningful improvement, and all coupled limit-state violations are zero.The certification gate additionally requires composite score S ≥ 85 and every dimensional subscore to be at least 60.
- Certification gate: Certifiability is a non-negotiable objective: the threshold S = 85 was calibrated to the 90th percentile of reference projects that passed independent regulatory reviews.The selected candidate achieved S = 91 and passed all CCS AIP review items.
- Topology optimization: BESO coupled to CalculiX minimizes static compliance under a prescribed volume fraction by removing low-stress material and adding material in highly stressed void regions.A soft-kill lower bound x_min = 10^-3 prevents stiffness-matrix singularities while effectively removing void elements.
- Parametric upscaling: Anisotropic fixed-draft scaling preserves vertical dimensions while uniformly scaling horizontal geometry, then refines the multiplier to minimize specific steel consumption.The sizing envelope also enforces static pitch θ ≤ 5° under maximum rated aerodynamic thrust and evaluates hydrostatic restoring stiffness.
Statistics and reproducibility
The paper reports single deterministic solver runs for flagship metrics and provides integrity and reproducibility materials alongside supplementary validation analyses.
- Reproducibility: Flagship metrics come from single deterministic solver runs unless otherwise noted.The paper states this as the default reporting basis for its headline measurements.
- Reproducibility: A SHA-256 integrity manifest accompanies the artifact bundle, while additional convergence histories, sensitivity analyses, and extended validation results appear in supplementary information.The paper also states that code and data are available as described in its availability section.
Data and code availability
The paper provides an open-source orchestration platform and describes planned deposition of simulation data and flagship run bundles, with anonymized artifacts available on request.
- Availability: The orchestration platform is available on GitHub under an open-source license, while high-fidelity datasets and flagship run bundles are planned for Zenodo deposition with a DOI upon publication.Third-party solvers follow their respective licenses, and anonymized artifacts are available from the corresponding author on reasonable request.
Declaration of interests
The authors disclose that they are applying for national and international patents.
- Declaration of interests: The authors are applying for national and international patents.
Supplementary Information
The supplementary information details implementation methods for the BESO solver, Automated Reviewer, and Zwind’s aero-hydro-servo-elastic coupling. It also covers parametric upscaling sensitivity analysis and supporting mathematical derivations.
- The supplementary methods detail the BESO solver’s softkill formulation, filter-radius evolution, and adaptive mesh-refinement strategies.
- The materials include parametric upscaling sensitivity analysis and Automated Reviewer scoring functions calibrated with benchmark fleet data.
- The supplement provides mathematical derivations for Zwind 2.5’s aero-hydro-servo-elastic multiphysics coupling.