A novel macro-scale machine learning prediction based on high-fidelity CFD simulations: A case study on the pore-scale porous Trombe wall with phase change material capsulation

Abstract

In the present study, multi-layer calculations loops based on machine learning used to simulate the flow pattern and the thermal behavior through pore scale porous media (PSPM) walls, including phase-change materials (PCMs) in Trombe walls. A study of the thermal behavior of PSPM walls over a 24 hr period was conducted with a steady-state flow passing through the wall by applying the Monte-Carlo radiation model for the system. Constant-concentration PCMs help to reduce the temperature gradients between day and night. On the basis of porosity, solar radiation, heat flux, and time, a novel solution for prediction and optimizing the concentration of PCM in the Trombe wall was presented. For small components of PSPM, this method considers a high-fidelity CFD-based calculation. Following thorough validation, the values of velocity, pressure, and temperature for the fluid and solid zones were used as a training data set for machine learning. High-order optimization helps us to find the optimum PSPM wall porosity and PCM concentration (10%<ε<90%,0<C*PCM<9×103capsules.cm-3). The results indicated that the optimal combination of the PCM concentration and the wall porosity was ε=48%,C*PCM=7.1×103capsules.cm-3. According to the results, PCMs can reduce the temperature of PSPM wall (outer side) by 5.2% compared to the without PCM state over day and night. Furthermore, the temperature gradient over the course of the day was 6.34% lower than it is when PCMs are not used. For the period from 17:00 to 06:00, the average temperature of the PSPM wall with PCMs is up to 6.64% higher than it is without PCMs. We demonstrate how a combination of machine learning and numerical simulation can be used to predict the flow behavior and thermal pattern of a large PSPM Trombe wall. This solution may provide a framework to understand flow behaviour through complex geometry based on a micro-scale approach and long-term prediction for the macro-scale domain.