ConNAS4ML: Gradient-Based Hardware-Constrained NAS Framework for Real-Time Machine Learning in High-Energy Physics

Chi-Jui Chen^1*, Bo-Cheng Lai²

¹Graduate Degree Program of College of Electrical and Computer Engineering, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
²Institute of Electronics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan

* Presenter:Chi-Jui Chen, email:silencekugel.ee11@nycu.edu.tw

At the CERN Large Hadron Collider (LHC), achieving sub-microsecond latency is crucial for filtering interested events, enabling discoveries of new physics phenomena. Machine learning (ML) methods have become essential in enhancing event selection, with hls4ml providing an efficient pathway for deploying ML on FPGA hardware. However, balancing performance with stringent resource constraints presents complex trade-offs that make effective deployment challenging. We introduce ConNAS4ML, a gradient-based, constraint-aware neural architecture search (NAS) framework optimized for hardware-aware FPGA deployment. ConNAS4ML integrates practical resource metrics directly into the search process and adapts dynamically to different HLS configurations, FPGA models, and tool versions. It enables fine-grained, layer-specific optimization, allowing high-sensitivity layers to retain precision while staying within resource limits. ConNAS4ML’s contributions include (1) a customizable search interface for user-defined constraints, (2) seamless integration with the hls4ml workflow, and (3) efficient, one-shot architecture optimization that minimizes post-search fine-tuning. Validation on high-energy physics tasks demonstrates ConNAS4ML’s potential: a CNN for calorimeter energy reconstruction achieved a 48% parameter reduction with minimal accuracy impact, while a Jet Tagging task retained accuracy within 0.37% of baseline under strict resource constraints. This framework supports real-time ML deployments in high-energy physics and can be expanded to various other resource-constrained fields, including edge computing and autonomous systems.

Keywords: Machine learning , FPGA, hls4ml , high-energy physics, particle physics