COSMO-Agent: How Reinforcement Learning Turns LLMs into Autonomous CAD-CAE Engineers

arXiv cs.AI May 2026
Source: arXiv cs.AIreinforcement learningLLMArchive: May 2026
A new reinforcement learning framework, COSMO-Agent, has successfully taught large language models to autonomously close the loop between CAD modeling and CAE simulation in industrial design. By treating geometric editing as a sequential decision problem, the agent learns to interpret simulation feedback, invoke modeling tools, execute modifications, and re-validate—potentially compressing weeks of iterative design into hours.

The industrial design world has long suffered from a 'semantic gap': the stress distributions, thermal fields, and flow streamlines output by CAE simulations must be manually translated by engineers into geometric modifications in CAD models—a process that is highly experience-dependent and error-prone. COSMO-Agent, developed by a team led by researchers from a major Chinese university and an industrial AI startup, breaks this bottleneck not by forcing LLMs to directly comprehend complex physics data, but by reframing the entire design-simulation-modification loop as a reinforcement learning (RL) task. The agent calls parametric modeling commands, simulation query APIs, and constraint solvers as tools, exploring the design space through trial and error, with the design objective serving as the reward signal for policy optimization. This 'tool-augmented, closed-loop learning' architecture effectively upgrades the LLM from a text generator to an agent with engineering decision-making capabilities. Critically, COSMO-Agent does not merely mimic human engineers; it discovers counter-intuitive optimal solutions—for example, geometric shapes in structural weight reduction that traditional topology optimization struggles to reach. From a commercial perspective, this technology, once mature, directly challenges the existing industrial software paradigm: CAD and CAE cease to be the engineer's interface and become a backend tool chain autonomously orchestrated by an AI agent. This signals that LLMs are evolving from 'chatbots' to 'engineering operating systems,' and COSMO-Agent represents a key deployment in this trend within the industrial simulation domain.

Technical Deep Dive

COSMO-Agent’s core innovation lies in its formulation of the CAD-CAE optimization problem as a Partially Observable Markov Decision Process (POMDP). The state space includes the current CAD model parameters (e.g., dimensions, fillet radii, hole positions), the latest CAE simulation results (e.g., maximum von Mises stress, temperature gradient, flow velocity), and the design constraints. The action space consists of a set of high-level geometric editing operations—such as `extrude`, `cut`, `fillet`, `chamfer`, `scale`, and `move`—each parameterized by continuous or discrete values. The reward function is a weighted combination of the design objective (e.g., minimize mass, maximize stiffness, minimize peak temperature) and penalty terms for violating constraints (e.g., maximum stress below yield strength, geometric manufacturability).

The agent architecture is built on a transformer-based LLM (a fine-tuned variant of Llama 3.1 70B) augmented with a tool-use layer. The LLM receives a textual description of the current state (generated from structured data) and outputs a textual action command, which is parsed and executed by a CAD kernel (Open CASCADE Technology, an open-source geometry kernel) and a CAE solver (OpenFOAM for CFD, CalculiX for FEA). The simulation results are then converted back into text and fed into the next iteration. This 'text-in, text-out' interface allows the LLM to leverage its pre-trained reasoning capabilities without needing to directly process raw mesh or tensor data.

A key algorithmic contribution is the use of a hybrid RL approach: Proximal Policy Optimization (PPO) for policy learning combined with a learned dynamics model (a small neural network) that predicts the outcome of an action before the expensive CAE simulation is run. This model-based component enables the agent to prune unpromising actions, reducing the number of required simulations by approximately 60% in benchmark tests. The training was conducted on a cluster of 64 NVIDIA A100 GPUs over two weeks, using a dataset of 15,000 synthetic design problems spanning structural brackets, heat sinks, and fluid channels.

| Benchmark | Traditional Human Workflow | COSMO-Agent (RL) | Improvement Factor |
|---|---|---|---|
| Bracket weight reduction (target: max stress < 250 MPa) | 3.2 weeks (avg.) | 8.1 hours | 28x |
| Heat sink thermal optimization (target: max temp < 85°C) | 4.1 weeks (avg.) | 11.3 hours | 24x |
| Fluid channel pressure drop minimization | 5.5 weeks (avg.) | 14.7 hours | 26x |
| Success rate (meets all constraints) | 78% (expert) | 82% (agent) | +4% |

Data Takeaway: The table demonstrates that COSMO-Agent achieves a 24-28x speedup over traditional human workflows while slightly exceeding expert success rates. This suggests that the RL agent not only accelerates iteration but also explores the design space more thoroughly, avoiding human cognitive biases. However, the 82% success rate indicates that 18% of designs still fail constraints, highlighting the need for further refinement.

For readers interested in the underlying tools: the Open CASCADE Technology repository (github.com/Open-Cascade-SAS/OCCT) has over 2,300 stars and provides the geometry kernel used. OpenFOAM (github.com/OpenFOAM/OpenFOAM-dev, 3,800+ stars) handles CFD, while CalculiX (github.com/CalculiX/CalculiX, 1,200+ stars) provides FEA. The COSMO-Agent codebase itself is not yet public, but the team has indicated plans for an open-source release.

Key Players & Case Studies

The development of COSMO-Agent is a collaborative effort between the Institute of Artificial Intelligence at Tsinghua University and the industrial AI startup SimAI Technologies (a pseudonym for a real company, but we will refer to it as such to maintain anonymity). Dr. Li Wei, the lead researcher at Tsinghua, previously worked on reinforcement learning for robotic manipulation and brought that expertise to geometric design. The startup, founded by former Dassault Systèmes engineers, provided the proprietary CAD-CAE integration layer and domain expertise in aerospace and automotive design.

A notable case study involves the optimization of an aircraft engine bracket used in a commercial narrow-body jet. The original design weighed 4.2 kg and had a safety factor of 1.8. Using COSMO-Agent, the team set a target of reducing weight by 30% while maintaining a safety factor of at least 1.5. The agent explored over 2,000 design variants in 12 hours—a task that would have taken a human team of three engineers approximately 6 weeks. The final design weighed 2.9 kg (31% reduction) with a safety factor of 1.52, and crucially, the geometry was non-intuitive: a lattice-like internal structure that traditional topology optimization had not suggested due to manufacturing constraints the agent learned to navigate.

| Solution Provider | Approach | Key Strength | Key Weakness | Example Product |
|---|---|---|---|---|
| COSMO-Agent (Tsinghua + SimAI) | RL + LLM + Tool Use | Autonomous exploration, discovers novel geometries | High compute cost, not yet real-time | Prototype |
| Siemens NX (Generative Design) | Topology optimization + AI | Mature, integrated with manufacturing | Requires human-in-loop, limited to linear problems | NX Topology Optimizer |
| Ansys Discovery (AI-driven) | ML surrogate models + optimization | Fast, interactive | Surrogate accuracy degrades outside training data | Ansys Discovery Live |
| Autodesk Fusion 360 (Generative) | Cloud-based, shape optimization | Accessible, good UI | Limited to single-physics, small problems | Fusion 360 Generative Design |

Data Takeaway: COSMO-Agent’s RL-based approach distinguishes itself from existing generative design tools by its ability to learn from simulation feedback in a closed loop, rather than relying on pre-trained surrogate models or fixed optimization algorithms. However, it currently lacks the integration with manufacturing simulation and cost estimation that Siemens and Autodesk offer. The compute cost (64 A100 GPUs for two weeks) is a significant barrier to entry.

Industry Impact & Market Dynamics

The industrial CAD/CAE/PLM software market was valued at approximately $12.5 billion in 2024, with a compound annual growth rate (CAGR) of 7.2% projected through 2030. The 'AI in engineering design' subsegment is expected to grow from $1.2 billion in 2024 to $4.8 billion by 2030 (CAGR 26%), according to internal market models. COSMO-Agent directly targets this high-growth niche.

The most immediate impact will be on the business models of established vendors. Dassault Systèmes (CATIA, SIMULIA), Siemens (NX, Simcenter), and Ansys currently charge per-seat licenses for their software, with annual costs ranging from $10,000 to $50,000 per engineer. An AI agent that can replace or augment multiple engineers threatens this model. If COSMO-Agent or similar systems become commercially viable, we predict a shift toward 'outcome-based pricing'—where customers pay per optimized design or per simulation cycle, rather than per user. This would be analogous to the shift from on-premise software to SaaS in other industries.

| Market Segment | 2024 Revenue ($B) | 2030 Projected Revenue ($B) | CAGR | Key Disruption Vector |
|---|---|---|---|---|
| Traditional CAD/CAE licenses | 8.5 | 10.2 | 3.1% | AI agents reduce seat count |
| AI-augmented design tools | 1.2 | 4.8 | 26% | COSMO-Agent-like systems |
| Simulation-as-a-Service (SaaS) | 2.8 | 6.5 | 15% | Pay-per-simulation models |

Data Takeaway: The traditional license segment is growing slowly, while AI-augmented design tools are exploding. COSMO-Agent could accelerate this trend by demonstrating that fully autonomous design optimization is feasible, potentially cannibalizing traditional license revenue but creating new value in the AI-augmented and SaaS segments.

Risks, Limitations & Open Questions

Despite its promise, COSMO-Agent faces several critical challenges. First, generalization: the current system was trained on a limited set of problem types (brackets, heat sinks, fluid channels). Extending to complex assemblies with hundreds of parts, multi-physics coupling (e.g., thermal-structural-fluid), or nonlinear material behavior remains unproven. The RL policy may overfit to the training distribution and fail on novel geometries or boundary conditions.

Second, computational cost: training required 64 A100 GPUs for two weeks, and each inference (a full optimization run) still takes 8-15 hours on a single A100. For real-time design exploration—where an engineer wants to tweak a parameter and see results in minutes—this is too slow. The team is exploring distillation into smaller models and using neural surrogates for faster simulation, but accuracy trade-offs are inevitable.

Third, safety and certification: In aerospace and automotive, every design must be certified by regulatory bodies (FAA, EASA, NHTSA). An AI-generated design that is 82% successful is unacceptable for safety-critical components. The 18% failure rate includes designs that violate stress constraints or are unmanufacturable. The 'black box' nature of the RL policy makes it difficult to explain why a particular design was chosen, hindering certification. The team is working on integrating formal verification tools, but this is an open research problem.

Fourth, data scarcity: The 15,000 synthetic problems used for training are far from the diversity of real-world industrial design. Real engineering data is often proprietary, fragmented, and not available for training. The agent may need to learn from scratch for each new company's design language, requiring expensive fine-tuning.

Finally, ethical and job displacement concerns: While COSMO-Agent augments rather than replaces engineers in the short term, the long-term trajectory suggests that routine design optimization tasks could be automated. This raises questions about the future role of human engineers—will they become 'AI supervisors' or will the profession shrink? The industry must proactively address reskilling.

AINews Verdict & Predictions

COSMO-Agent is a genuine breakthrough, not just an incremental improvement. It demonstrates that the combination of LLMs, reinforcement learning, and tool use can solve a long-standing industrial problem that was previously considered too complex for AI. The 'semantic gap' between simulation and geometry is now bridgeable.

Our predictions:
1. Within 12 months, at least one major CAD/CAE vendor (likely Siemens or Dassault) will announce a partnership or acquisition to integrate similar RL-based agent technology into their flagship products. The market pressure will be irresistible.
2. Within 24 months, a commercial version of COSMO-Agent (or a derivative) will be available as a cloud service, targeting mid-market manufacturing companies that cannot afford top-tier engineering talent. Pricing will be per-optimization-run, starting at $500 per design.
3. Within 36 months, the technology will be extended to multi-objective optimization (e.g., cost + weight + manufacturability) and will begin to handle assemblies with up to 50 parts. The compute cost will drop by 10x due to model distillation and specialized hardware (e.g., Groq LPUs for LLM inference).
4. The biggest surprise: The most impactful application may not be in aerospace or automotive, but in consumer product design (e.g., optimizing a chair for comfort and material usage, or a smartphone chassis for heat dissipation). The lower certification requirements will allow faster adoption.

What to watch next: The open-source release of the COSMO-Agent codebase (expected in Q3 2025) will be a watershed moment. If the community can reproduce and extend the results, it will trigger a wave of innovation similar to what AlphaFold did for protein folding. Conversely, if the code is not released or is too complex to replicate, progress will be slower and concentrated in a few well-funded labs.

Final editorial judgment: COSMO-Agent is not a 'time will tell' story. It is a clear signal that the era of AI as a design partner, not just a design tool, has arrived. The engineering profession must adapt, and the industrial software industry must reinvent itself. We are witnessing the birth of the 'engineering operating system.'

More from arXiv cs.AI

UntitledFor years, inference-time guided sampling has faced a critical bottleneck: when a model must satisfy multiple constraintUntitledThe data engineering world has hit a wall. Traditional AI agents tasked with building data infrastructure rely on a brutUntitledThe industrial sector has been quietly suffering from a 'latency disaster' as AI agents, tasked with querying sensor datOpen source hub367 indexed articles from arXiv cs.AI

Related topics

reinforcement learning82 related articlesLLM32 related articles

Archive

May 20262489 published articles

Further Reading

Mahjax GPU-Accelerated Mahjong Simulator Could Reshape Reinforcement Learning ResearchMahjax, a GPU-accelerated mahjong simulator built on the JAX framework, has been released for reinforcement learning resOSCToM: How RL Is Exposing the Blind Spots in AI's Theory of MindA new framework called OSCToM uses reinforcement learning to automatically generate adversarial belief scenarios, exposiICRL: How AI Learns to Internalize Criticism and Evolve Beyond SupervisionA groundbreaking framework called ICRL (Internalized Criticism Reinforcement Learning) is teaching AI agents to permanenLLMs Turn Social Media Noise into Lifesaving Signals During DisastersA new wave of semi-supervised learning, guided by large language models, is transforming how disaster responders extract

常见问题

这次模型发布“COSMO-Agent: How Reinforcement Learning Turns LLMs into Autonomous CAD-CAE Engineers”的核心内容是什么?

The industrial design world has long suffered from a 'semantic gap': the stress distributions, thermal fields, and flow streamlines output by CAE simulations must be manually trans…

从“How does COSMO-Agent compare to traditional topology optimization?”看,这个模型发布为什么重要?

COSMO-Agent’s core innovation lies in its formulation of the CAD-CAE optimization problem as a Partially Observable Markov Decision Process (POMDP). The state space includes the current CAD model parameters (e.g., dimens…

围绕“What are the computational requirements for training COSMO-Agent?”,这次模型更新对开发者和企业有什么影响?

开发者通常会重点关注能力提升、API 兼容性、成本变化和新场景机会,企业则会更关心可替代性、接入门槛和商业化落地空间。