Maxwell's Demon, Landauer's Limit, and Quantum Information Heat Engines: From Microscopic Models to a Macroscopic Self-Powered Energy Harvester

We present a unified theoretical, numerical, and experimental investigation into autonomous information heat
engines (Maxwell’s demons), bridging microscopic open quantum thermodynamics, stochastic continuous moni-
toring, deep reinforcement learning, and practical energy harvesting engineering. First, addressing widespread
numerical pitfalls in quantum thermodynamics simulations, we identify an operator sign inversion in detailed-
balance jump rates that previously simulated an unphysical “perpetual motion machine” yielding −19.29 kB T .
Once corrected via adaptive Runge-Kutta integration, we prove that single-step work extraction from maximally
entangled Bell states is strictly bounded by kB T ln 2, and that the cyclic entropy accounting of entanglement re-
generation imposes a net non-negative work cost (Wnet ≥ 0), establishing quantum entanglement as a consumed
thermodynamic fuel rather than a violation of the Second Law. Second, using positive-preserving Rouchon-Ralph
stochastic master equation mappings across 48 trajectory ensembles, we evaluate an Information Conveyor Belt
in a 10-dot chain under continuous homodyne detection. We demonstrate that while an isolated cryogenic demon
reservoir (TM = 4.2 K) enables positive extracted power, incorporating the mandatory Carnot refrigeration over-
head required to maintain the cryostat exactly cancels all efficiency advantages, yielding an overall system deficit
matching the isothermal baseline. Third, we address autonomous control in partially observable Markov deci-
sion processes (POMDPs) by deploying Recurrent Proximal Policy Optimization (Recurrent PPO with LSTM),
enabling the demon to autonomously overcome measurement backaction and discover synchronized conveyor trans-
port policies superior to classical heuristics. Finally, we design and verify Configuration D, a body-heat-powered
thermoelectric energy harvester operating an ESP32-C3 microcontroller demon. By eliminating idle wake-time
delays and overlapping 9-bit digital sensor conversion with sub-millisecond USB polling, firmware v2 slashes ac-
tive controller consumption in our calibrated design model by a factor of 8.7 (from ≈ 208.5 mJ to ≈ 24.1 mJ),
establishing a net-positive self-powered operational cycle (+65.8 mJ/cycle). We quantify the fundamental gap of
≈ 8.4 × 1018 (≈ 19 orders of magnitude) between the Landauer erasure limit (2.87 × 10−21 J) and real-world
microcontrollers, delineating the rigorous boundary between physical information thermodynamics and real-world
energy harvesting.