By Nidhi DhullReviewed by Susha Cheriyedath, M.Sc.Sep 4 2024A recent article published in Applied Sciences proposed integrating Multilayer Perceptron (MLP) and a Generative Adversarial Network (GAN) to predict the thermal conductivity (TC) of concrete based on its mass composition and density. Experimental and synthetic data were used to compare the performance of the developed model.
Methodology framework. Image Credit: https://www.mdpi.com/2076-3417/14/17/7598
Background
With rising concerns about energy efficiency in buildings, the demand for the development of materials with better thermal performance in buildings has increased. Such materials can preserve indoor thermal comfort despite fluctuations in the outdoor environment and reduce energy consumption.
Concrete is the most commonly used building material. Its thermal conductivity depends on several factors, including the constituents’ type and weight. Accurately predicting concrete's thermal conductivity is crucial to improving buildings’ energy efficiency.
A machine learning-based prediction model can help avoid several inefficient experiments during concrete production’s design and material selection phases. Artificial neural networks (ANNs) are one such model commonly employed to solve complex problems in various fields. However, effectively training an ANN requires a huge amount of data. Using a GAN helps overcome this problem.
Thus, this study focused on developing an ANN model integrated with a data augmentation model to predict the thermal properties of concrete comprising different types of materials such as slag, aggregates, fibers, etc.
Methods
The proposed model comprised two primary components: an MLP for material property prediction based on its features and a GAN for data augmentation to enhance the dataset and improve the accuracy of MLP. The developed model was then applied to a broad case study on concrete mixtures.
A comprehensive literature survey was performed to create a database of 200 points representing different types of concrete, such as recycled aggregate, lightweight, foamed concrete, moderate-strength, and vermiculite-comprising concrete. Accordingly, 18 input parameters influencing the output (TC) were considered.
The compiled dataset was randomly split into training (80%) and validation subsets (20%) to train the model. Subsequently, the reliability and reproducibility of the ANN model were evaluated on a new test dataset. Additionally, the predictive capabilities of the model different performance metrics were used, including coefficient of determination (R2) and root mean squared error (RMSE).
The MLP model's hyperparameters were initially predefined and later fine-tuned to improve prediction accuracy. The number of hidden layers and neurons was systematically varied to optimize the MLP model. Notably, a rigorous trial-and-error method was followed to identify the most effective number of hidden layers and neurons in each layer.
Finally, the CopulaGAN synthesizer, a type of conditional GAN, was employed to augment data and improve the overall efficacy of the MLP-based model. It used a cumulative distribution function-based transformation like Gaussian copulas, making the learning process easier.
Results and Discussion
The thorough optimization of the MLP model using the trial-and-error approach yielded a two-hidden-layer structure with 200 and 100 neurons as the most proficient model. This structure exhibited an RMSE of 0.111 W/mK and an R2 of 0.984 for the training dataset. The corresponding values were 0.183 and 0.960 for the validation dataset, respectively.
Notably, increasing the number of layers could not enhance the model’s learning capacity or performance. Moreover, the model exhibited overfit with increasing hidden layers with a small dataset, and thus, the loss functions exhibited higher values. Therefore, CopulaGAN was implemented for data augmentation, which could create novel synthetic data from the real data used for making the MLP model.
Two distinct sets of synthetic data, one with 200 data points and the other with 1000 data points were created to assess the model’s performance with this diverse training approach. Additionally, two metrics were utilized to analyze the quality of the synthetic data: Correlation Matrix Comparison and Kolmogorov-Smirnov (KS) Test. While the former examined the consistency of the relationships between variables in the artificial and real datasets, the latter compared the two datasets to identify if they were derived from the same underlying distribution.
The synthetic data effectively replicated the correlation structure in the real data with some deviations in the exact correlation values. This could be resolved with more refinement of the artificial data generation process.
Conclusion
Overall, the researchers successfully demonstrated the efficacy of using GAN for data augmentation during the development of an MLP model. The developed model demonstrated satisfactory performance with a highly accurate prediction of concrete TC across the following training scenarios for the MLP: exclusively with the real database, with 200 synthetic data points followed by real data, and with 1000 synthetic data points followed by real data. However, incorporating synthetic data in the training remarkably enhanced the accuracy and reduced errors within the test dataset.
The results of this study highlight the importance of diverse datasets for training ANN models to ensure accuracy and the potential of synthetic data to augment training when real data is scarce. Such accurate prediction models for a building material’s thermal properties can advance energy efficiency in buildings.
Journal Reference
Rosa, A. C., Elomari, Y., Calderón, A., Mateu, C., Haddad, A., & Boer, D. (2024). Methodology for the Prediction of the Thermal Conductivity of Concrete by Using Neural Networks. Applied Sciences, 14(17), 7598. DOI: 10.3390/app14177598, https://www.mdpi.com/2076-3417/14/17/7598
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.