UC Davis

The material and geometric characteristics of every particle in a system of two or more particles affect the mechanical interaction between them and consequently affect the mechanical behavior of the entire system. The Finite Element Method (FEM) or the Discrete Element Method (DEM) are two popular strategies for numerical evaluation of the mechanical behavior of the system of particles. However, both of these numerical techniques are computationally expensive or can be difficult to implement in some practical applications. This dissertation utilizes methods from Deep Learning (DL) to extensively reduce the computational time of two different problems associated with mechanical interactions between similar and dissimilar particles. In the first problem, two different Convolutional Neural Network (CNN) models are used for the sub-second prediction of thermo-mechanical interaction between particles in powder beds used in Selective Laser Sintering (SLS) Additive Manufacturing (AM) process. In the second problem, an Artificial Neural Network (ANN) model is developed for the sub-second evaluation of the dynamic response of structures to the reaction force that arises during the collision of dissimilar material bodies.

Physical modeling of the transient temperature throughout the SLS manufacturing process is critical for determining the quality of Additively Manufactured structure (AM structure). Conventional numerical models for simulating the thermal field of AM structures, however, are time-consuming and cannot be easily applied to the development of a real-time simulation system. In this dissertation, a sophisticated existing conventional physics-based simulation is utilized to generate a dataset with thousands of two-dimensional (2D) position-time representations of the laser head with various process parameters and their associated heat-map of AM structures for DL training purposes. This dataset is used to train a deep encoder-decoder CNN model capable of sub-second prediction of the heat-map of AM structure. It is shown that, on average, the proposed DL model is more than 41,000 times faster than the physics-based model.

The physical modeling of powder-based AM structures aims to relate the macroscopic effective mechanical properties of these structures quickly and accurately to their microscopic structural features. For DL training purposes, the DEM simulation is used to simulate powder particle interactions and evaluate macro-level elastic properties of hundreds of AM structures. In this dissertation, an accurate CNN model for predicting the effective elastic properties of SLS manufactured structures is introduced, with an average error of less than 4%. The sub-second level computational time of this CNN model may be viewed as a step toward in developing a real-time system capable of predicting the effective properties of powder beds.

A comprehensive understanding of the nonlinear compliance behavior of colliding bodies throughout the collision process is required to assess the dynamic behavior of many multi-body granular systems such as powder-based AM processes, pharmaceutical manufacturing processes, and mineral operations. However, the conventional numerical models for evaluating the nonlinear force-displacement behavior of dissimilar colliding material bodies often use time-consuming iterative algorithms, which may considerably slow down multi-body granular simulations with a high number of colliding material bodies. In this dissertation, a time-consuming nonlinear lumped parameter model based on the strain-hardening power law is utilized to create a dataset with thousands of force-displacement curves for colliding bodies with different material and geometrical properties and various relative impact velocities. The generated dataset is used to train an ANN model for sub-second prediction of force-displacement behavior of colliding material bodies during elastoplastic impact. The ANN model results in lowering the computational cost of multi-body dynamic analysis and granular flow simulations significantly.