Physics-Integrated Inference for Signal Recovery in Non-Gaussian Regimes

TL;DR

A physics-integrated inference framework that decouples signal morphology from stochastic transients using a hierarchical 1D CNN-GRU topology that recovers deterministic signals from fluctuation-dominated regimes, enabling near-ideal detection limits in noisy edge environments is introduced.

Abstract

High-performance room-temperature sensing is often limited by non-stationary $1/f$ fluctuations and non-Gaussian stochasticity. In spintronic devices, thermally activated Néel switching creates heavy-tailed noise that masks weak signals, defeating linear filters optimized for Gaussian statistics. Here, we introduce a physics-integrated inference framework that decouples signal morphology from stochastic transients using a hierarchical 1D CNN-GRU topology. By learning the temporal signatures of Néel relaxation, this architecture reduces the Noise Equivalent Differential Temperature (NEDT) of spintronic Poisson bolometers by a factor of six (233.78 mK to 40.44 mK), effectively elevating room-temperature sensitivity toward cryogenic limits. We demonstrate the framework's universality across the electromagnetic and biological spectrum, achieving a 9-fold error suppression in Radar tracking, a 40\% uncertainty reduction in LiDAR, and a 15.56 dB SNR enhancement in ECG. This hardware-inference coupling recovers deterministic signals from fluctuation-dominated regimes, enabling near-ideal detection limits in noisy edge environments.

Figures (4)

• Figure 1: General signal restoration. Physics-integrated inference (red) resolves deterministic signals from stochastic inputs (grey), outperforming heuristic baselines (yellow). (a) Radiometric recovery in HADAR, reducing NEDT from 234 mK to 40 mK and avoiding heuristic signal attenuation. (b) LiDAR path reconstruction reducing MaxPE to 0.8 m. (c) Radar jamming mitigation maintaining 1.9$^\circ$ RMSE stability against electronic bursts.
• Figure 2: Stochastic physics of spintronic transduction. (a) Time-domain dynamics: Raw transients containing low-amplitude Néel signal events (red zone) that are misclassified and lost via standard heuristic clipping. (b) Thermally activated reversal: Switching probability follows an exponential Arrhenius signature (red curve), confirming the signal's physical origin. (c) Radiometric transfer function: The Noise Equivalent Differential Temperature (NEDT) is defined by the ratio of the RMS noise floor ($V_n$) to the linear sensor responsivity ($dV/dT$, dashed line).
• Figure 3: Hierarchical Spatiotemporal Reasoning Core. (Top) Skip connection preserves signal integrity. (Center) 1D-CNN extracts transient signatures into feature map $F \in \mathbb{R}^{C \times W'}$ (where $C$ is channel depth, $W'$ is encoded time). (Bottom) Bi-GRU utilizes global hidden state $h_t$ to resolve history-dependent stochasticity. A dense layer fuses these streams to recover physical estimate $\hat{s}(t)$ from input $\mathbf{x} \in \mathbb{R}^{W \times 1}$.
• Figure 4: Performance sensitivity. (a) Ablation study confirming the CNN-GRU hybrid is required for <45 mK noise. (b) Residual error density $\epsilon$; inference core (red) Gaussianizes the heavy-tailed transients seen in heuristic filtering (yellow). (c) Latency vs. efficiency; green zone marks real-time operation ($L < 1$ ms).