PEANUT dual cutoff model architecture
=====================================

Overview of the PEANUT dual cutoff model architecture and components.


To potentially improve learning, a variant of the PEANUT dual cutoff model separates neighbor pairs into close and non-close neighbors. 
After constructing the neighbor list, pairs are divided based on a distance threshold :math:`0 < B \le r_c`:

.. math::
    CN_B = \{(i,j) \mid d_{ij} < B\}, \quad
    NCN_B = \{(i,j) \mid d_{ij} \ge B\}

while the overall architecture and feature encoding (Behler-Parinello symmetry functions and spherical harmonics) remain the same. 
This separation, implemented via two edge MLPs, allows the model to assign greater importance to close (mostly bonded) atom pairs and test whether this improves its ability to distinguish between bonded and non-bonded interactions. 
The number of parameters increases if both edge MLPs have separate weights, but inference speed is not affected since each atom pair is processed independently through its respective part of the network.