PhysRes-MPC: Physical Hysteretic Reservoir Computing for Embedded Control of Liquid-to-Gas Phase-Change Actuators
TL;DR
This letter integrates a two-model PhysRes predictor with sampling-based Model Predictive Control (MPC) and deploys it on an edge platform, achieving the lowest measured desktop one-step inference latency and lowest mean absolute error.
Abstract
Liquid-to-gas phase-change actuators provide compact, compliant actuation, but their nonlinear pressure dynamics, thermal delay, and asymmetric heating-cooling response make precise control difficult. Building on an established Physical Hysteretic Reservoir Computing (PhysRes) architecture, this letter integrates a two-model PhysRes predictor with sampling-based Model Predictive Control (MPC) and deploys it on an edge platform. Separately trained scalar pressure and joint two-output temperature models use shifted-sigmoid reservoirs, with only their linear readouts fitted by closed-form ridge regression. With 1,050 training samples, PhysRes predicts chamber pressure, Peltier temperature, and liquid temperature with <inline-formula><tex-math notation="LaTeX">$R^{2} \geq 0.9977$</tex-math></inline-formula> for each state. Among all evaluated predictors, PhysRes achieves the lowest measured desktop one-step inference latency of <inline-formula><tex-math notation="LaTeX">$0.013 \,\mathrm{m}\mathrm{s}$</tex-math></inline-formula>, 12.0× faster than a fully trained long short-term memory (FT-LSTM) network. Across experiments at 120, 130, and <inline-formula><tex-math notation="LaTeX">$140 \,\mathrm{k}\mathrm{Pa}$</tex-math></inline-formula>, PhysRes-MPC achieves a mean absolute error (MAE) of <inline-formula><tex-math notation="LaTeX">$1.95 \,\mathrm{k}\mathrm{Pa}$</tex-math></inline-formula>, <inline-formula><tex-math notation="LaTeX">$32.2 \,\%$</tex-math></inline-formula> lower than proportional-integral-derivative (PID) control, and the lowest maximum overshoot at each target. On the ESP32, frozen PhysRes and FT-LSTM pressure predictors require 157.0 and <inline-formula><tex-math notation="LaTeX">$2090.2 \,\mathrm{\upmu }\mathrm{s}$</tex-math></inline-formula> per inference, respectively; their corresponding 200-candidate, 10-step rollouts require 316.5 and <inline-formula><tex-math notation="LaTeX">$4205.5 \,\mathrm{m}\mathrm{s}$</tex-math></inline-formula>.