Imagine asking an AI to design a virtual replica of a robot exploring the moon. This “digital twin” could then be used in simulations to predict real-world performance. That’s the promise of digital twin technology, and researchers are pushing the boundaries of what AI can achieve in this space. A new benchmark called SimBench tests how well large language models (LLMs) can generate these digital twins. SimBench uses a clever approach: a “judge LLM” evaluates code created by “student LLMs” attempting to design digital twins for various physics simulations. These simulations range from simple mechanisms like a lever to complex systems like self-driving cars navigating a highway. The results are eye-opening. While LLMs have shown potential, they’re far from perfect. Even after being trained on vast amounts of code, the best-performing models only achieved a 13% success rate on SimBench’s challenging tasks. Why the struggle? Building a digital twin requires more than just stringing together lines of code. It demands a deep understanding of physics, engineering principles, and the specific quirks of different simulation software. LLMs excel at pattern recognition, but they often lack the reasoning abilities to translate a high-level request (like “design a lunar rover”) into a functional simulation. SimBench’s multi-turn interactions reveal another weakness: LLMs sometimes struggle to adapt and refine their code based on feedback. The research highlights the limitations of current LLMs, but also points toward future improvements. More sophisticated judging techniques, better training datasets, and refined LLM architectures could pave the way for truly AI-powered digital twin generation. This breakthrough would dramatically accelerate the development of complex systems, allowing engineers to test and refine designs virtually before building physical prototypes. From optimizing robot designs to simulating disaster scenarios, the possibilities are vast.