A Hybrid Physics–AI Framework for Dynamic Pressure Prediction in Ammonia-Based Chemisorption Systems

This study presents a hybrid physics–AI framework that integrates thermochemical simulation and neural network modeling to predict the dynamic pressure behavior in ammonia-based chemisorption systems using manganese chloride (MnCl₂) and strontium chloride (SrCl₂) as working salts. A detailed numerical model was first developed to simulate the adsorption–desorption cycle of ammonia within a cylindrical chemisorption reactor, incorporating energy and mass balance equations based on the Clausius–Clapeyron relation and Arrhenius-type kinetics. The simulation generated high-resolution time-series data of temperature, pressure, and flow rate under operating conditions ranging from 100–300°C and 1–10 bar. Building on these datasets, this work develops a neural-network-based surrogate model for predicting reactor pressure using a hybrid input representation that combines high-dimensional time-series vectors with three pointwise operating parameters (inlet/outlet temperatures and mass flow rate). A three-layer fully connected neural network is trained on these features with appropriate shuffling, K-fold cross-validation, and standardization, achieving R 2 values close to 1 and very low prediction errors across all folds. Visualization of real-versus-predicted values and residual distributions confirms that the surrogate reliably reproduces the physics-based pressure trajectories, with only minor performance degradation in low-sample regions. The results demonstrate that integrating physics-based simulation with data-driven learning significantly enhances regression performance and provides a fast, accurate, and interpretable tool for optimizing thermochemical energy storage systems.