In robotics the URDF format can be used to specify and simulate rigid-body systems.
You can also check various physics engine input file formats like MuJoCo (MJCF format). (It's just some variant of XML).
If you want soft-body dynamics, you can have decompose your objects into particles of different types which interact, (examples are liquidfun or powder toy). (It's just a list of particles and a particle-particle interaction matrix) (More or less base on Smoothed Particle Hydrodynamics).
If you want something more serious, you'll have to look at finite-element based multi-physics tools, like Ansys or Comsol, which will allow you to specify the various domains and solve the differential equations on each domains.
If you want to go very small scale you have Molecular Dynamics tools with standards like GROMACS.
If you are more into the topological optimization, you can use neural networks to represent the quantities which are of interest to you and Physically Inspired Neural Networks (PINN), or neural ODE, or "sim-to-real" tools, in which case you'll probably be using pytorch.
Or text if you use WorldSim like LLM tools and only need an abstract representation. Agents playing with CAD software tools will probably standardize some convenient way to solve engineering tasks.
If you want soft-body dynamics, you can have decompose your objects into particles of different types which interact, (examples are liquidfun or powder toy). (It's just a list of particles and a particle-particle interaction matrix) (More or less base on Smoothed Particle Hydrodynamics).
If you want something more serious, you'll have to look at finite-element based multi-physics tools, like Ansys or Comsol, which will allow you to specify the various domains and solve the differential equations on each domains.
If you want to go very small scale you have Molecular Dynamics tools with standards like GROMACS.
If you are more into the topological optimization, you can use neural networks to represent the quantities which are of interest to you and Physically Inspired Neural Networks (PINN), or neural ODE, or "sim-to-real" tools, in which case you'll probably be using pytorch.
Or text if you use WorldSim like LLM tools and only need an abstract representation. Agents playing with CAD software tools will probably standardize some convenient way to solve engineering tasks.