Title: Rethinking Scientific Modeling: Toward Physically Consistent and Simulation-Executable Programmatic Generation

Abstract:

In computational engineering science, structural modeling serves as a foundational pillar, where even slight physical discrepancies or failures to meet specifications can compromise the validity of subsequent simulations. While recent studies have highlighted the capability of large language models (LLMs) to automatically generate modeling code, outputs that are either non-executable or physically inconsistent continue to be a significant challenge under strict engineering requirements. To address this, we propose a framework for the automatic generation of building models that ensures physical consistency. This approach integrates the construction of domain-specific knowledge, alignment with constraints, and evaluation driven by verification processes.

We introduce CivilInstruct, a specialized dataset designed to formalize structural engineering knowledge and constraint reasoning, thereby facilitating the creation of models ready for simulation. Furthermore, a two-stage fine-tuning methodology is implemented to guarantee adherence to constraints and compatibility with application programming interfaces (APIs), which significantly curtails the occurrence of hallucinated and non-compliant outputs. Additionally, we present MBEval, a benchmark focused on verification that assesses both executability and consistency in structural dynamics via closed-loop validation. Experimental findings indicate that our method yields consistent performance gains over baseline models across stringent verification metrics. The source code is publicly accessible at https://github.com/Jovanqing/AutoBM.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC