(87c) A Machine Learning Assisted Hopper Flow Design for Handling Granular Biomass Materials

The potential of biomass-derived energy is significant, yet its realization is hindered by handling and feeding challenges associated with granular biomass feedstock. For instance, the prevalent issues in biorefinery hoppers include arching, rate hole formation, and avalanche flow. These challenges stem from gaps in understanding the relationships between material properties, hopper design factors and flow performance.

In this study, we integrate physical experiments, validated numerical simulations, and data augmentation techniques to develop a machine learning-assisted hopper design tailored for flowing granular woody biomass materials. Initially, we employed experimental data from cyclic axial compression and ring shear tests to train two deep-learning networks: a sequential model utilizing Gated Recurrent Units as the encoder and feed-forward neural networks as the decoder, and an incremental model using feed-forward neural networks. Upon validation, these networks were utilized to enhance the laboratory-measured stress-strain behavior of pine chips across various moisture contents and mean particle sizes.

Subsequently, we employed a meticulous calibration process to adapt a modified hypoplastic model capable of accurately representing the intricate flow behavior of granular biomass materials. These calibrated constitutive models, in conjunction with a variety of initial packing configurations and hopper operational conditions (e.g., outlet opening size, wall friction, inclination), facilitated hopper flow simulations using a validated Smoothed Particle Hydrodynamics code. For each simulation, we defined and computed a set of performance indices, such as average flow rate, clogging potential, flow smoothness, and flow pattern coefficients, to quantitatively assess flow performance.

Finally, a feed-forward neural network was trained and validated to establish correlations between material attributes, initial packing, hopper operation parameters, and flow performance indices. This neural network serves as a comprehensive design guideline for effectively handling milled woody biomass materials in hoppers, benefiting stakeholders in biorefinery and equipment manufacturing sectors.