TRIADS for Catalyst Zero Research

Catalyst Zero Research fork/copy of TRIADS: Tiny Recursive Information-Attention with Deep Supervision.

TRIADS is a compact recursive neural architecture for materials-property prediction in small-data regimes. This fork exists so Catalyst Zero Research has a clean public research base for materials ML work, future Catalyst integrations, and materials-science benchmark workflows.

Original work: Rtx09x/TRIADS
Model artifacts: huggingface.co/Rtx09x/TRIADS
Preprint: 10.5281/zenodo.19200579

Why This Lives Here

Catalyst Zero Research is focused on materials science, scientific machine learning, and AI-driven computational research.

TRIADS fits that direction because it is a small-data materials ML system built around:

• composition-aware and structure-aware featurization,
• attention over physically meaningful descriptors,
• recursive shared-weight reasoning,
• deep supervision across iterative prediction steps,
• Matbench-style evaluation workflows.

Included Tasks

This repository includes TRIADS work for:

• matbench_steels
• matbench_expt_gap
• matbench_jdft2d
• matbench_phonons
• classification tasks under matbench_classification

The code and archived experiments are preserved from the original public TRIADS release so the research path stays inspectable.

Direction

This fork is intended to become the Catalyst-facing home for:

• TRIADS-based materials-property prediction,
• Matbench reproducibility work,
• Catalyst knowledge-graph feature experiments,
• structure/property reasoning workflows,
• future scientific ML baselines under Catalyst Zero Research.

Status

Research artifact / public fork.

The original TRIADS results and paper should be treated as the current reference. Future Catalyst-specific changes should be documented here as separate branches, notes, or releases rather than silently replacing the original research artifact.

Citation

@misc{tiwari2026triads,
  title = {TRIADS: Tiny Recursive Information-Attention with Deep Supervision},
  author = {Tiwari, Rudra},
  year = {2026},
  doi = {10.5281/zenodo.19200579},
  url = {https://doi.org/10.5281/zenodo.19200579}
}

Links

• Catalyst Zero Research: github.com/Catalyst-Zero-Research
• Original TRIADS repo: github.com/Rtx09x/TRIADS
• Portfolio: rtx09x.github.io