A comparative study of deep learning neural network architectures and sensitivity analyses for the prediction of color changes in biochar

The objective of this research is to model the “L” color parameter that gives the dark color of biochar to the soil, with machine learning methods. In the study, 24 different deep learning neural networks (DLNN) and 8 different artificial neural networks architectures were utilized. At the modeling stage, models were produced for 3 different input sets by using carbonization temperature, holding time, gas flow rate, “a” and “b” color parameters with different combinations. “a” and “b” color parameters are indicators of redness-greenness, and blueness-yellowness, respectively. Sensitivity analyses were performed on network architectures that gave the best results. According to the results, the best models were obtained in DLNN architectures. L color parameter was modeled with the accuracy of 76.5%, 96%, and 98% using first, second, and third input sets respectively. In the sensitivity analysis of input set 3, it was determined that the effect of “a” and “b” values on the “L” color change of biochar was more than 50%. In this input set, the effects of carbonization temperature, gas flow rate, and holding time on the “L” color parameter of biochar were determined as 22.1%, 18.5%, and 9.5%, respectively. According to elemental analysis results, the C ratio changed directly proportional to carbonization temperature and inversely proportional to holding time and gas flow rate. H and N ratios decreased depending on the increase in the carbonization temperature, holding time, and gas flow rate. Thanks to this study, both the best models that predict the color parameter L, which is an indicator of the dark color of biochar and the impacts of the factors used in the production phase on the color of biochar are determined.