Robust biochar yield and composition prediction via uncertainty-aware ResNet-based autoencoder

Abstract With growing demand for energy and fossil fuels, biomass and biochar are gaining attention due to their abundance and sustainability. However, there is a crucial need for optimizing production conditions and minimizing environmental risks associated with biochar production. Machine learning is an emerging strategy for predicting biochar yield and composition, optimizing production conditions and minimizing environmental risks. This paper presents a ResNet-based autoencoder model that utilizes biomass properties and pyrolysis conditions to more accurately and robustly predict biochar yield and composition. The developed model has the advantage of addressing the common data uncertainty concerns in training data. Our model outperforms commonly used baseline methods, including MLP-NN (mean R 2 = 0.907), Random Forest (RF, mean R 2 = 0.798), XGBoost (XGB, mean R 2 = 0.826), and Gaussian Process (GP, mean R 2 = 0.786), by achieving a mean R 2 of 0.974. The performance of the model was further improved by incorporating previously discarded data with high missing rates, achieving an average R 2 of 0.983. The addition of the three newly collected covariates resulted in an average R 2 of 0.985. Additionally, robust sensitivity analyses of the input covariates revealed the impact of data uncertainty on the performance of the model, emphasizing the robustness of the model. In advancing the application of machine learning in biochar research, this study provides a reliable method to determine optimal production conditions. Graphical Abstract