kha8128/NPS — reverse-engineered prompt

Build me a Python research package called Neural Phase Simulation that can train deep learning surrogate models for materials simulations and then run fast predictions from them. I want it focused on microstructure evolution and accelerated molecular dynamics, with examples for things like grain growth, nucleation and growth, spinodal decomposition, dendrite growth, dislocation dynamics, and periodic 2D or 3D simulations.

It should have a clear command line entry point where I can run training mode or prediction mode, plus simple example data or tutorial scripts so a new researcher can try it without guessing. Support common neural model types used for this area, like convolutional networks, recurrent convolutional models, graph neural networks, equivariant models, and diffusion style models if practical. Please include plotting or visualization helpers for results, configuration options for experiments, and clean documentation explaining installation, training, prediction, and how to add new simulation types.

Use Python with PyTorch and scientific Python tools. Look up current docs online if you need to.