PhysicsNeMo Functionals

PhysicsNeMo functionals follow the torch.nn.functional pattern: stateless operations designed for direct use in model code, training loops, and pre/post-processing pipelines. They are intended to be easy to compose and to behave consistently across CPU and GPU execution paths.

Many functionals are optimized for NVIDIA GPUs and can dispatch to accelerated implementations when those backends are installed. For operations with multiple implementations, PhysicsNeMo selects a preferred implementation by default and falls back to another supported one when needed, emitting a warning so behavior is explicit. Functionals with multiple implementations have plots available in the documentation for performance comparisons.

PhysicsNeMo Functionals

• Neighbor Functionals
  • KNN
  • Radius Search
• Geometry Functionals
  • signed_distance_field()
• FFT Functionals
  • rfft()
  • rfft2()
  • irfft()
  • irfft2()
  • view_as_complex()
  • real()
  • imag()
• Regularization and Parameterization Functionals
  • drop_path()
  • weight_fact()
• Interpolation Functionals
  • interpolation()