Why Artificial Intelligence Can't Design a Jet Engine

The article analyzes the gap between AI's ability to generate visually convincing 3D models and the genuine understanding of the physical world required to design functional machines. It maps technological progress in generative 3D models (Hunyuan3D, TRELLIS), the systematic failures of language models in spatial reasoning, and research efforts to create spatial intelligence capable of understanding physics and mechanics.

Between generating pretty three-dimensional models and genuinely understanding the physical world lies a chasm that no one has yet bridged. For the first time, though, we know where the bridge should run.

Imagine asking an AI for a three-dimensional model of a turbine blade. In ten seconds you get a visually convincing object — smooth surfaces, the right shape, plausible textures. Except that blade doesn't have the correct flow profile. Its dimensions match no tolerance class. It won't fit into the rotor. And, above all — the artificial intelligence has no idea why it should fit there at all.

This contradiction defines the current state of artificial intelligence in the three-dimensional world. On one side stands dizzying progress in shape generation — models like Hunyuan3D or TRELLIS produce textured objects in seconds. On the other side remains the chasm between what AI can manufacture and what it actually understands. The chasm between a triangle mesh and a working machine.

This analysis maps the full breadth of the field: from the fundamental limitations of language models, through the revolution in generating three-dimensional bodies, to advanced attempts at teaching artificial intelligence to grasp physics. The key question reads: when — if ever — will AI be able to independently design a functional mechanical system?

Today's large language models — GPT-4, Claude, Gemini — process the world through two modalities that are flat by their very nature: sequential text tokens and raster image crops. A survey paper for the IJCAI 2025 conference (Zha et al.) identified the core of the problem: text is inherently imprecise and ambiguous about the size, shape, and relative placement of objects unless it is robustly coupled with spatial understanding (IJCAI 2025).

Benchmark tests quantify this gap mercilessly. The BLINK task suite (Fu et al., ECCV 2024) reformulated 14 classic computer-vision tasks — depth estimation, multi-view reasoning, spatial relationships — as multiple choice. Humans achieved an average accuracy of 95.7%. GPT-4V a mere 51.3%. Gemini 45.7%. That is roughly 13 and just under 8 percentage points above random guessing — on tasks a human solves at a single glance (arXiv: 2404.12390).

The "Thinking in Space" study (Yang et al., CVPR 2025) delivered another disquieting finding. On more than 5,000 questions from real-world indoor videos, the best model, Gemini-1.5 Pro, reached an overall accuracy of around 46%, whereas humans reached 79%. In a sub-experiment with cognitive maps the model did achieve 64% accuracy in placing adjacent objects, but for greater distances the accuracy fell off sharply. Moreover, techniques such as chain-of-thought, which usually help with logical reasoning, paradoxically did not improve spatial performance — in some cases they even slightly worsened it (arXiv: 2412.14171, project page).

The "Stuck in the Matrix" study (Cohen et al., arXiv 2025) revealed how quickly language models' spatial reasoning collapses as complexity grows. On a small 3×3 grid all models work reliably. As the grid grows, accuracy plummets — the average drop is 42.7%, in the worst cases up to 84%. Every test that started with accuracy above 50% recorded a drop of at least 48 percentage points (arXiv: 2510.20198).

MIT research (Wu et al., 2024) then offered an explanation: language models excel in familiar scenarios, but under a minimal change of conditions — altered arithmetic, chess with different starting positions — they drop to the level of random guessing. The performance is largely a consequence of memorization, not general reasoning (MIT News).

Particularly troubling is the work "Revisiting 3D LLM Benchmarks" (Jin et al., ACL Findings 2025). Vision-language models using only two-dimensional renderings of point clouds significantly outperformed the best specialized three-dimensional language models on some benchmarks. The authors called this the problem of "cheating with two-dimensional information" — many supposedly three-dimensional tasks can be solved from surface-level two-dimensional information (arXiv: 2502.08503).

What does it actually mean to "understand" a jet engine? It requires reasoning simultaneously about thousands of components with micron tolerances, about flow across the compressor stages, about the thermal expansion of various alloys at various temperatures, about rotor dynamics and vibration modes. A language model can recite learned facts — "a turbofan engine has a fan, a compressor, a combustion chamber, and a turbine." But it cannot mentally rotate components, reason about tolerance chains, grasp the influence of thermal expansion on clearances, or trace load paths through the structure. This is the fundamental difference between declarative knowledge — facts about something — and procedural spatial knowledge — understanding how it works.

Despite these fundamental limitations, the field of generative three-dimensional models underwent a dramatic transformation in 2024–2025. For comparison: Point-E (OpenAI, 2022) generated rough point clouds in one to two minutes; DreamFusion (Google, 2022) optimized a neural radiance field (NeRF) for a single object for roughly an hour and a half. Modern feed-forward models produce textured three-dimensional polygonal meshes in seconds.

Hunyuan3D (Tencent) is among the most advanced open-source platforms. Version 2.0, from January 2025, combines shape generation via a diffusion transformer with texture synthesis producing physically based materials — albedo, metalness, roughness, normal maps — in 10 to 25 seconds. Version 2.5, from June 2025, increased the fidelity of detail (GitHub, hy-3d.com).

TRELLIS (Microsoft, CVPR 2025, Spotlight honor), with its TRELLIS.2 version, reached 4 billion parameters and introduced O-Voxel — a sparse voxel structure handling complex topologies at resolutions up to 1536³ voxels in roughly 60 seconds on an H100 graphics processor (GitHub, Datameister).

The Large Reconstruction Model family (Hong et al., ICLR 2024), with 500 million parameters, predicts a three-dimensional object from a single image in roughly 5 seconds. The GS-LRM variant achieves 0.23 seconds from two to four images. InstantMesh (TencentARC, 2024) combines a multi-view diffusion model with an LRM architecture and differentiable iso-surface extraction (LRM project, InstantMesh).

Gaussian Splatting (Kerbl et al., SIGGRAPH 2023) has become the standard for real-time rendering. Representing scenes with millions of anisotropic Gaussian primitives, it achieves real-time rendering at 30 to 361 frames per second with quality matching or exceeding NeRF. Standardization efforts by the Khronos and OGC groups in 2025 add Gaussian splats to the glTF format ecosystem. The DUSt3R/MASt3R family (Naver Labs, CVPR 2024/2025) fundamentally advanced dense three-dimensional reconstruction without camera calibration — MASt3R-SLAM runs at 15 frames per second (MASt3R-SLAM).

The visual results are impressive. But for engineering these models are fundamentally inadequate. The generated polygonal meshes are "triangle soups" — without a parametric CAD model with constraints, dimensions, or a feature tree. Testing by Xometry showed that tools such as Zoo Design worked for simple flanges and brackets but failed on moderately complex gears. According to Meshy, roughly one in ten generations is usable for a client without manual edits, with the system struggling with complex multi-part assemblies (Leo AI). The difference between 10.0 mm and 10.05 mm is the difference between a working part and scrap.

If current language models do not understand space, and generative models create only visual shells without physical content, is there a way forward? Several research groups believe there is — and they call it spatial intelligence.

Fei-Fei Li, a Stanford professor and one of the key figures of computer vision, defines spatial intelligence as an artificial intelligence's ability to "perceive, model, reason about, and act in physical or geometric space." In her November 2025 essay "From Words to Worlds," she argued that large language models are in a sense wordsmiths in the dark — eloquent but inexperienced; knowledgeable but ungrounded in the real world (Fei-Fei Li Substack, Fast Company).

Li co-founded World Labs with Justin Johnson, Christoph Lassner, and Ben Mildenhall (a co-creator of NeRF). The company raised 230 million dollars in a round led by Andreessen Horowitz, NEA, and Radical Ventures in September 2024, with participation from NVIDIA and others. In early 2026, it then raised roughly another 1 billion dollars from AMD, NVIDIA, Autodesk, Fidelity, and others, at a valuation of around 5 billion dollars (World Labs blog, TechCrunch).

Their product Marble, launched in November 2025, is a generative multimodal world model creating persistent, walkable three-dimensional worlds from text, images, video, or rough spatial layouts — exportable as Gaussian splats, polygonal meshes, or videos, compatible with Vision Pro and Quest 3 (World Labs blog, TechCrunch). World Labs further introduced the World API, integrated with Isaac Sim and MuJoCo, and the open-source Spark rendering library (World Labs: API, Bigger Worlds).

In parallel, three-dimensional scene understanding has advanced. ConceptGraphs (Gu et al., ICRA 2024) build an open-vocabulary, graph-structured scene representation by fusing two-dimensional foundation models (SAM, CLIP) into three-dimensional space via multi-view association, with GPT-4 generating object descriptions and inter-object relationships (ConceptGraphs). LERF (Kerr et al., ICCV 2023, UC Berkeley) embeds CLIP language features into neural radiance fields, thereby enabling language queries in three-dimensional space. 3D-LLM (Hong et al., NeurIPS 2023), the first language model accepting three-dimensional point clouds as input, outperformed two-dimensional vision-language models on the ScanQA benchmark (arXiv: 2307.12981).

The fundamental question for designing machines is not whether AI can create a shape, but whether it grasps why things move, break, and deform.

LLMPhy (Cherian et al., NeurIPS 2024) demonstrated a promising approach: combining a language model with a physics compute kernel produces state-of-the-art physical reasoning without prior training on the given task. The key insight is that a language model alone cannot predict many-body dynamics, but it can generate hypotheses testable in a simulator. The language model thus serves as the hypothesis generator, the physics kernel as the verifier. This suggests that language models need external physical simulation in place of internal physical understanding (OpenReview: LLMPhy, arXiv: 2411.08027).

The "How Far is Video Generation from World Model" study (Kang et al., ICML 2025, ByteDance) systematically demonstrated that video-generating models cannot generalize physical rules. Instead, they exhibit nearest-training-example imitation. The authors' conclusion is unambiguous: merely scaling up models is not enough to uncover fundamental physical laws (OpenReview: ICML 2025).

Even Genie 3 (Google DeepMind, August 2025) — the first interactive world model generating three-dimensional worlds in real time at 720p resolution and 24 frames per second — has limitations in physics accuracy and consistency lasting only a few minutes (TechCrunch).

In the area of understanding mechanical assemblies, specialized systems have brought advances. DYNAMO (2025) predicts motion in geared mechanical assemblies — spur, helical, and bevel gears, worm drives, rack-and-pinion, and planetary gear sets (arXiv: DYNAMO). MechaFormer (Bolanos et al., AAAI 2025, Autodesk Research) automates the design of kinematic mechanisms — generating both topology and joint coordinates from constrained curves for four-, six-, and eight-bar mechanisms (arXiv: MechaFormer). Articulate-Anything (Le et al., ICLR 2025) achieves 75% accuracy in the automatic modeling of articulated three-dimensional objects, versus the 8.7–12.2% of earlier methods (arXiv).

These are real advances. But they are still isolated demonstrations on relatively simple mechanisms — not complex machine assemblies with hundreds of interacting components.

The practice of commercial AI-enabled CAD tools shows a realistic picture of current capabilities — and of their limits.

In topology optimization, AI functions as an accelerator: neural networks replacing the finite element method (FEM) within the optimization loop (Deng et al., Nature Communications, 2022) reduce the share of FEM computation to roughly 40% of total time thanks to self-supervised online learning (Nature Communications). The latest 2025 approach, based on reinforcement learning, integrates the PPO algorithm with FEM for generative lightweight structures with manufacturing constraints (PubMed Central). A review in the Journal of Computational Design and Engineering maps three avenues: surrogate models, neural FEM substitutes, and implicit neural representations (Oxford Academic).

As of early 2026, commercial tools are in the "AI assistant" phase. Autodesk Assistant in Fusion automates sketch constraints and machining toolpaths. Siemens NX Copilot offers context-aware design assistance. PTC Creo has generative topology optimization running inside the parametric history tree. Onshape AI Advisor assists with modeling (Autodesk blog, Siemens NX blog, DEVELOP3D). At the 3DEXPERIENCE World 2026 conference, SOLIDWORKS demonstrated the conversion of a two-dimensional sketch into a fully parametric three-dimensional model — with a planned release in mid-2026 (Machine Design).

The most ambitious direction is Neural CAD — generating parametric CAD models from text. DeepCAD (Wu et al., ICCV 2021) laid the foundations: a Transformer-type autoencoder treating CAD command sequences much like natural language, trained on 178,238 CAD models from Onshape (TheCVF). Text2CAD (Khan et al., NeurIPS 2024 Spotlight) outperforms DeepCAD by 18.6% on intermediate-level prompts (Text2CAD). Zoo/KittyCAD offers a production-ready text-to-CAD interface generating STEP files, but limited to simpler geometries (3D Printing Industry). In November 2025, MIT introduced an AI agent that learns to use CAD to create three-dimensional objects from sketches (MIT News).

In practice, the deployment of artificial intelligence is further along in optimization and inspection than in independent design. GE Aerospace uses AI inspection of turbine blades in more than 12 repair facilities. Boeing has over 70 generative-AI applications in production (Boeing Innovation). Airbus uses generative design for components (Aero Mag). NVIDIA PhysicsNeMo achieves up to a 500-fold speedup by combining GPU-accelerated solvers with AI surrogate models (NVIDIA Developer). Siemens and NVIDIA are building a partnership for the first artificial-intelligence-driven manufacturing plants, beginning in 2026 (NVIDIA Newsroom). At CES 2026, Siemens introduced the Digital Twin Composer tool (Siemens). ANSYS SimAI shortened automotive aerodynamic analysis from 50 hours to less than one hour with accuracy above 95% (AI Buzz).

According to Mordor Intelligence, the generative-design market is expected to reach 8.58 billion dollars by 2030, with an average annual growth of 14.82% (Mordor Intelligence). But the key detail remains: AI does not yet design entire machines — it optimizes individual components within parameters defined by a human engineer.

Robots take a different route to spatial understanding. Physical interaction yields qualitatively different knowledge than the passive observation of images.

The WoW paper (September 2025) trained a 14-billion-parameter generative world model on 2 million robotic interaction trajectories. It showed that models trained on interactions exhibit emergent physical understanding that passive video models lack (arXiv). EmbodiedMAE (May 2025) created the DROID-3D dataset with 76,000 trajectories representing 350 hours of interaction data, and outperformed all previous visual foundation models (arXiv).

DreamerV3 (Hafner et al.), published in Nature in April 2025, reached a milestone — a single algorithm with fixed hyperparameters mastering more than 150 diverse tasks across continuous control, Atari, and Minecraft (Nature). V-JEPA 2 (Meta, June 2025) is a 1.2-billion-parameter world model trained on more than 1 million hours of video and 1 million images by self-supervised learning, with fine-tuning on only about 62 hours of robotic data. It achieves 65–80% success rates in pick-and-place tasks in unseen environments with direct transfer and no prior training (The Robot Report, TechCrunch).

Foundation models for robotics have undergone turbulent development. π0 (Physical Intelligence, October 2024) demonstrated multi-step tasks such as folding laundry. OpenVLA (Stanford, June 2024) outperformed RT-2-X by 16.5% in absolute terms. GR00T N1.5 (NVIDIA, June 2025) is an open-source model with 3 billion parameters for humanoid robots (Wikipedia: VLA, Emergent Mind). Tactile sensors achieve resolution 5× higher than human fingertips; Figure 03 detects forces as small as 3 grams (Analog Devices, Figure AI).

In transfer from simulation to the real world: DexGraspNet 2.0 demonstrated a 90.7% success rate in real dexterous grasping via direct transfer. NVIDIA SRSA achieved a 90% success rate in real assembly tasks. Google DeepMind DemoStart achieved a 97% success rate in real cube reorientation (Google DeepMind blog, Edge AI).

Even so, the assembly of complex mechanical systems remains a frontier challenge. Robots manage simple assemblies — peg in a hole, nut and bolt. Multi-step mechanical assemblies requiring an understanding of kinematic chains, tolerances, and sequential dependencies exceed current capabilities.

The fundamental chasm between current capabilities and genuine three-dimensional understanding has several dimensions.

Compositionality — assembling components into working systems — requires understanding not only individual shapes but their functional relationships and constraint conditions. PhysPart (ICRA 2025) introduced differentiable physics simulation to ensure smooth motion of the generated components. The key finding: physical constraints must be explicitly modeled — they do not arise from pattern recognition (ICRA 2025). PhyCAGE (2024) reached the same conclusion for object interpenetration problems (arXiv).

Causal reasoning remains an unbridged chasm. The Transformer's autoregressive prediction mechanism is not inherently aligned with logical or physical causality. The CWMI framework (2025) explicitly addresses this deficit by introducing a Causal Physics Module, but the model still fails on substances, liquids, and deformable objects (arXiv).

Datasets of mechanical assemblies are scarce. PartNet-Mobility offers 2,700 articulated three-dimensional objects, but focuses predominantly on household items — not gearboxes and turbines. VideoCAD (MIT, NeurIPS 2025), with more than 41,000 examples of CAD interactions, represents a step in the right direction, but covers a fraction of the engineering space.

The most promising convergence is taking place at the intersection of three directions: World Labs and spatial intelligence creating persistent three-dimensional worlds; NVIDIA Newton — an open-source physics compute kernel with GPU acceleration, developed jointly by NVIDIA, Google DeepMind, and Disney Research, announced at the GTC conference in March 2025 and contributed to the Linux Foundation in September 2025 (NVIDIA Developer, Linux Foundation); and foundation models for robotics bringing embodied understanding.

It would be unfair not to consider the arguments of those who see a shorter path.

The emergence argument. Proponents of scaling object that spatial understanding may arise spontaneously in sufficiently large models trained on sufficiently rich data. GPT-4 does indeed solve some spatial tasks better than GPT-3.5. But the ByteDance study for ICML 2025 directly tested this hypothesis and concluded that scaling up models alone is not enough to uncover physical laws — models learn surface correlations, not causal structure.

The tool-sufficiency argument. Perhaps artificial intelligence does not need to "understand" physics — it is enough that it can correctly use a physics simulator. The LLMPhy approach partly confirms this. But to design a new mechanism it is not enough to test hypotheses — they must first be generated, and that requires at least an intuitive understanding of what might work.

The human-AI collaboration argument. The most realistic counter-position: artificial intelligence need not design machines on its own — it is enough that it dramatically accelerates the work of a human engineer. This argument is strong, and today's practice (Boeing, GE, Airbus) confirms it. The question we are posing, however, aims further — at autonomous understanding.

The development trajectory suggests four horizons with decreasing certainty.

2026–2027: Reliable conversion of text to parametric CAD for simple components. AI assistants as standard across all major CAD platforms. Surrogate models as a routine part of simulation workflows. Robots handling structured assemblies with known parts.

2028–2032: Foundation models combined with world models and tactile perception will enable adaptation to new assembly tasks with just a few demonstrations. AI-assisted design of complete subassemblies with automatic manufacturability verification.

2032+: Genuine universal mechanical assembly — a robot explores an unknown mechanism, infers its function, and performs precise assembly. This will probably require breakthroughs in causal physical reasoning and integrated three-dimensional world models with prediction of contacts and forces.

The fundamental challenge remains the transition from correlational pattern recognition to causal physical understanding. The solution will probably require a combination of multimodal language models for high-level semantic reasoning with world models for physically aware simulation — because language models alone overlook physical constraints, and world models alone lack high-level semantics.

Let us return to that turbine blade from the beginning. Today, AI can create a visual copy of it in 10 seconds, optimize its shape for aerodynamics in minutes, and shorten a flow simulation from 50 hours to one. But to design a functional turbine stage from scratch, where the blade works together with dozens of others in a rotating gear assembly at temperatures above 1,000 °C — that artificial intelligence still cannot do. Not because it lacks computing power. But because it lacks what engineers build over an entire career: a mental model of how things actually work in three dimensions.

For the first time, though, there exists a realistic roadmap for how to get there. The combination of spatial intelligence, differentiable physics, and embodied experience — robots that learn to understand the world by touching it — offers a path from flat thinking to three-dimensional understanding. The road is longer than the distance we have already traveled. But for the first time, the direction is clear.

Transparency of creation

The conception, structure, and editorial line of the article are the work of the author, who prepared the content outline, established the key theses, and directed the entire creative process. Generative AI (Claude Opus 4.6, Anthropic) was used as a tool for research, fact-checking, and elaborating the author's draft.

The author verified the key findings and approved the final wording. No part of the text was published without conscious authorial control. Factual data were verified against the publicly available sources cited in the text.

The procedure complies with the transparency principles of EU Regulation 2024/1689 (AI Act). #poweredByAI