Development of Data-Driven Predictive Framework for Nanofluid-based Solar Thermal Collector: A Machine Learning Approach

The conventional approach for solar collector design often requires performing large-scale experimentations or computationally intensive simulations which hinders the comprehensive screening and optimization of process design. This creates a strong rationale for developing a computationally efficient framework capable of leveraging a relatively small number of samples to generate a machine-learning model with sufficiently high fidelity. In this regard, the present study aims to integrate the concepts of random sampling, Gaussian process(GP), and Bayesian optimization for developing a computationally efficient data-driven framework for capturing the complete continuous domain of the parametric variation and predicting the desired performance measure. The proposed framework is rigorously tested at different stages with the help of unknown samples (out-of-fold test samples) to ensure the sound generalization capability of the constructed model. The model assessment revealed that the increase in sample size for training the GP model from 35 samples to 105 samples resulted in ≈ 56% reduction in root mean square error (RMSE), which further reduces to ≈ 96.5% after performing Bayesian optimization based hyperparameter tuning. The proposed framework will be extremely helpful in designing the highly efficient nanofluid-based solar thermal collector, by preventing the need of performing large-scale experimentations/simulations for screening purpose.