Empowering phase change material thermodynamics via physics-consistent neural network-enabled advanced building control

The increasing frequency of climate extremes and the global imperative to reduce building-related emissions call for materials and control strategies that operate synergistically. This study proposes a physics-consistent neural network (PCNN)-enabled differentiable predictive control (DPC) framework that integrates bio-based phase change materials (PCMs) with advanced building energy management. A modularized encoder-decoder surrogate model was developed to emulate nonlinear zone-level thermodynamics grounded in the first law of thermodynamics, embedding latent PCM behavior through physically informed neural modules. Based on this surrogate, a differentiable control policy was optimized to minimize energy cost, peak load, and comfort violations across dynamic climate and pricing scenarios. The framework was validated via closed-loop EnergyPlus co-simulations of a five-zone office with PCM-integrated envelopes, spanning multiple bio-based PCMs, six U.S. climate zones, and future extreme weather scenarios. Across the typical summer week, PCNN-DPC enabled Q25 PCMs to achieve up to 25.06 % energy-cost savings and 37.84 % peak-load reduction without compromising thermal comfort. Under extreme-heat events, PCMs with DPC policy reduced indoor peak temperature by as much as 5.31 °C, evidencing substantial thermal resilience. These findings recast PCMs from passive media into programmable thermal assets and provide a blueprint for co-optimizing envelope materials and predictive control in next-generation climate-adaptive buildings.