Research work · 2026

Physics-Guided or Physics-Limited? Diagnosing Where Physics Works in Deep Radar Emitter Recognition

Elsevier BV

Abstract

Radar emitter recognition is physics-rich: each waveform class follows a precise modulation law, and inter-class confusion is governed by matched-filter cross-correlation, not data statistics. Yet most deep-learning approaches treat classes as arbitrary labels, ignoring this structure. We ask when physics should guide the training signal versus architecture, testing three injection strategies for 15 waveform classes across four families (SNR ∈ [-20,+10] dB) using a physics similarity matrix S (15×15) built from waveform taxonomy and matched-filter cross-correlation. Standard cross-entropy reaches r = 0.39 alignment with S. As a relational metric loss (Strategy 1), S raises alignment to r = 0.79 — physics is learnable — yet fine-class accuracy is unchanged (49.3±0.5% vs. 47.9±0.9%), since the 15-class S is near-PSD (λ_min= -0.1795), an ill-posed metric target. Family supervision through the loss alone (Strategy 2) recovers 73.3±0.5% family accuracy but leaves fine-class flat, so the architectural change is decisive. The 4×4 family projection is perfectly PSD (λ_min = +0.682); a hierarchical model (Strategy 3) attains 75.6±0.3% family versus 49.3±0.6% fine at -8 dB — a 26.3 pp gain explained by the Cramér–Rao bound: P-code fine-class discrimination needs SNR* = 25.6 dB, 15.6 dB above the dataset maximum. A cascaded diagnostic confirms P-code expert accuracy stays near-random (24.4% at -8 dB, 23.6% at +10 dB) — a physical law, not an architectural deficiency. The contribution is a data-independent rule — the sign of λ_min(S) — for whether physics should enter as loss or architecture in similarity-based recognizers; the application is low-SNR radar emitter recognition.

DOI: 10.2139/ssrn.7411570