Examples

Use an LLM to Control Software

The LLM generates configuration files, script commands, and codes in Python/Julia, C++/C, and domain-specific languages to control the operational flow.

• In a finite element solver, the main workflow is fixed, but many options are selected at each step.

• Unlike the fixed solver core, pre/post-processing workflows are dynamic and highly customizable.

Type	Flow Control	Readability	Examples
Configuration Files	Fixed (all steps)	Easy	Palace (.msh and .json)
Script Commands	Fixed (selectable steps)	Simple	Gmsh, LAMMPS/Kalmelo, Ansys APDL
Python/Julia	Dynamic (new flow)	Moderate coding capability	Gmsh (.jl and .py)
C++/C	Dynamic (new flow)	Advanced coding capability	Gmsh, OCCT, VTK
Domain-Specific Languages	Dynamic (new flow)	Closed to human language	FEniCS (UFL), KittyCAD/Zoo (KCL)

For LLM-driven software flow control, generating Python or Julia scripts is the recommended approach. Input the following content into DeepSeek:

Write a Julia script using Gmsh to create three spheres:
one large sphere centered at (0, 0, 0) with radius 1,
and two small spheres centered at (0.2, 0, 0) and (-0.2, 0, 0) respectively, each with radius 0.1.
Then, subtract the two small spheres from the large sphere and generate the mesh.

The generated Julia script is available at this link.