Abstract:This study investigates the pyrolysis characteristics of reed bamboo and red fir wood to predict their fast pyrolysis product yields. Initially, the physicochemical properties of the raw biomass, including their thermogravimetric behavior, were analyzed. Fast pyrolysis experiments were subsequently performed in a horizontal fixed-bed reactor to examine the influence of temperature, residence time, and feedstock particle size on the yields of solid, liquid, and gaseous products. Finally, a neural network-based model was developed to rapidly and accurately predict the three-phase product yields under various reaction conditions, integrating data from this study with previously reported findings for different biomass types. The results reveal that the pyrolysis of reed bamboo and red fir wood proceeds through four main stages: drying; decomposition of hemicellulose and cellulose; decomposition of lignin; and subsequent low decomposition of the char residue. During the primary decomposition stage, macromolecules like cellulose break down into smaller molecules, with levoglucosan as a major product. As the pyrolysis temperature increases from 350 to 650 ℃, the solid char yields for both biomass types decrease significantly. At higher temperatures, secondary cracking of pyrolysis vapor is intensified, leading to a lower liquid yiled and a higher gas yield due to the formation of non-condensable gases. Furthermore, extending the pyrolysis residence time enhanced the extent of reaction, promoting further cracking of chemical bonds within the char matrix and residual polymeric materials, thereby releasing more volatile gases. Consequently, the solid yield decreases and the gas yield increases. Additionally, an increase in feedstock particle size impedes intra-particle heat transfer, resulting in incomplete pyrolysis. As a result, higher solid char yields and lower liquid and gas yields are observed, as the particle core does not reach the optimal pyrolysis temperature. An artificial neural network (ANN) model was developed using a combined dataset comprising 28 experimental runs from this study and 62 data points from the literature. The data were split into training and testing sets with a 7:3 ratio. The model architecture comprised a single hidden layer with 14 neurons, utilizing the logistic-sigmoid activation function. The RMSprop optimizer was employed for training, with a learning rate of 0.0009 and a smoothing parameter of 0.9. The developed ANN model demonstrated high accuracy, with regression coefficients (R2) for the predicted solid, liquid, and gas yields at 0.971, 0.966, and 0.974, respectively. When validated against the experimental data from this study, the average relative errors between the predicted and actual yields for the three phases were 5.25%, 5.44%, and 5.23%, respectively, confirming the model′s strong predictive capability. Close-