Information Flow in the PEANUT models

Overview of the PEANUT model architecture and components.

Subsections

• PEANUT model architecture
  • Input for the model
  • Steps per atom \(i\) and message-passing iteration \(t = 1, \dots, T\)
  • Energy prediction
• PEANUT dual cutoff model architecture
• Peanut DualCutoff LearnedBoundary Architecture
  • Graph construction
  • Learned atom-pairwise cutoff distances
  • Soft distance-based gating
  • Dual edge message functions
  • Distance-weighted message combination
• PEANUT mini model architecture

The models use a star graph architecture to efficiently capture interactions between atoms in a molecular system. In this architecture, each atom is represented as a node, and edges connect each atom to a central “star” node that aggregates information from all atoms.

For each atom \(i\), a cutoff sphere of radius \(r_c\) is defined, within which neighboring atoms contribute to the atom’s representation. The star graph structure allows for efficient message passing and information aggregation, enabling the model to learn complex

Additionally, an attention mechanism can be employed to weigh the contributions of neighboring atoms based on their relevance to the target atom. This has not been done yet. Exemplary sketch for an attention layer is shown below: