Neural Architecture Codesign for Fast Physics Applications

← Back to all benchmarks

Date: 2025-01-09

Name: Neural Architecture Codesign for Fast Physics Applications

Domain: High Energy Physics

Focus: Automated neural architecture search and hardware-efficient model codesign for fast physics applications

Task Types: Classification, Peak finding

Metrics: Accuracy, Latency, Resource utilization

Models: NAC-based BraggNN, NAC-optimized Deep Sets (jet)

AI/ML Motif: Classification

Resources

Benchmark: Visit

Results: Gemini LLM Deep Research

Keywords

neural architecture search FPGA deployment quantization pruning hls4ml

Citation

• Jason Weitz, Dmitri Demler, Luke McDermott, Nhan Tran, and Javier Duarte. Neural architecture codesign for fast physics applications. 2025. URL: https://arxiv.org/abs/2501.05515, arXiv:2501.05515.

@misc{weitz2025neuralarchitecturecodesignfast,
  archiveprefix={arXiv},
  author={Jason Weitz and Dmitri Demler and Luke McDermott and Nhan Tran and Javier Duarte},
  eprint={2501.05515},
  primaryclass={cs.LG},
  title={Neural Architecture Codesign for Fast Physics Applications},
  url={https://arxiv.org/abs/2501.05515},
  year={2025}
}

Ratings

CategoryRating

Software

3.00

Toolchain (hls4ml, nac-opt) described but not yet containerized or fully packaged

Specification

5.00

Fully specified task with constraints and target deployment; includes hardware context

Dataset

2.00

Simulated datasets referenced but not publicly available or FAIR-compliant

Metrics

5.00

Clear, quantitative metrics aligned with task goals and hardware evaluation

Reference Solution

4.00

Models tested on hardware with source code references; full training pipeline not yet released

Documentation

4.00

Detailed paper and tools described; open repo planned but not yet complete

Average rating: 3.83/5

Radar plot

Edit: edit this entry