nvalchemi-dynamics-implementation

To build a custom dynamics integrator for GPU atomic simulations, subclass BaseDynamics and override the pre_update() and post_update() methods to define deterministic per-step logic for updating positions and velocities.

A custom dynamics integrator coordinates inputs and outputs with neural potentials by declaring __needs_keys__ and __provides_keys__, ensuring proper model I/O coordination during GPU-accelerated atomic simulation workflows.

Yes, you can integrate custom dynamics with existing forward models and convergence logic by leveraging inherited methods such as compute(), step(), and run() to build end-to-end simulation pipelines on a prepared Batch.

The best way to define deterministic per-step logic in a GPU simulation is by implementing pre_update() and post_update() within a BaseDynamics subclass, updating positions and velocities while relying on inherited methods for execution.

Yes, you need Python to implement custom dynamics, as the integrator requires subclassing BaseDynamics and overriding Python methods like __init__, pre_update(), and post_update() for GPU-accelerated atomic simulation workflows.