SPINTRONIC POISSON BOLOMETERS: BROADBAND INFRARED SENSING AND NEUROMORPHIC COMPUTING

The sensitivity of room-temperature infrared sensing is fundamentally constrained by stochastic thermodynamics and non-stationary 1/f fluctuations. In high-performance regimes, approaching the fundamental thermal limit (kB T) typically necessitates bulky cryogenic cooling. This dissertation presents the Spintronic Poisson Bolometer, a room-temperature detector that leverages thermally activated magnetization switching to encode radiometric flux into stochastic bit-streams. We address intrinsic heavy-tailed Néel relaxation noise via a two-stage strategy. First, we characterize the experimental limits using heuristic adaptive algorithms that provide stabilization but are constrained by signal clipping. To overcome this, we introduce a physics-integrated CNN-GRU inference framework that decouples deterministic signals from stochastic transients. This reduces the Noise Equivalent Differential Temperature (NEDT) by a factor of 6 (from 233.78 mK to 40.44 mK), effectively bridging the gap to cryogenic detectors. The framework’s universality is validated across Radar, LiDAR, and ECG domains. Beyond noise mitigation, we expand the device’s spectral capabilities by integrating nanoplasmonic antennas, enabling continuous ultra-broadband detection from Short-Wave to Long-Wave Infrared (0.8–14 µm). Finally, we reframe the sensor’s operation for neuromorphic and event-based imaging by exploiting native stochasticity as a computational resource. This architecture facilitates high-speed, asynchronous in-pixel computing, shifting the paradigm from passive sensing to active edge processing.