Evaluation of properties of bio-composite with interpretable machine learning approaches: optimization and hyper tuning

Abstract

Hemp bio-composite (HBC) is a sustainable material that can be considered as a “carbon negative” or “better-than-zero-carbon” because it absorbs more carbon from its surrounding than it releases during its manufacturing and application on the field. Because of their eco-friendliness, these bio-composites not only improve structure thermal efficiency but also encourage sustainable construction. However, due to its heterogeneous nature, the majority of the published research has experimentally investigated the characteristics of bio-composites. While the mathematical models remain still a challenge for academics. Therefore, three machine learning methods known as deep neural network (DNN), gene expression programming (GEP), and optimizable Gaussian process regressor (OGPR) are used to build up prediction models for compressive strength (CS) and thermal conductivity (TC) of hemp bio-composite. A total of 86 and 159 experimental values for TC and CS, respectively, were collected from the literature. To formulate the models, the ten most dominant input variables were chosen. To evaluate the performance of the models, various statistical metrics were employed. Based on statistical indicators, the increasing order of prediction capabilities of the established models is DNN>OGPR>GEP. The value of the correlation coefficient (R) for CS models is 0.9893 (GEP), 0.9997 (DNN), and 0.9995 (OGPR). While that for TC models are 0.9847 (GEP), 0.9999 (DNN), and 0.9999 (OGPR). The values of the performance index (ρ) and objective function (OF) are approaching 0. Thus, indicating the outburst performance of the developed models. The comparison of machine learning (ML) models with traditional regression models further confirms the accuracy of the developed models. Furthermore, contour maps analysis was conducted to get an insight into the ranges of input variables that were employed by global researchers to gain higher strength and thermal properties. For the interpretation of the models, Shapley Additive explanations (SHAP) analysis was carried out. The result revealed that the influencing trend of input variables is in close agreement with experimental studies. Thus, practitioners and designers to avoid hectic experimental tests to save time and cost can utilize these ML-based developed models in field projects. However, it is recommended to use more updated and extensive data to further validate the models and utilize other ML algorithms and post hoc interpretation techniques for the model's explanations.

Keywords

• Hemp bio-composite
• Thermal conductivity
• Compressive strength
• Machine learning
• Deep neural network
• Gene expression programming