Abstract
Additive manufacturing is becoming transformative methodologies that can deliberate a flexible and dynamic production system. Additive manufacturing encompasses many technologies, such as computer-aided design, convergence technologies, and internet of the thing, to deliver the desired output by depositing material, layer upon layer, through thermo-mechanical cycles. Additive manufacturing systems play a vital role in meeting the fourth industrial revolution (industry 4.0) requirements. However, this process fuses multi-physics and multi-scale phenomena that need the expertise to build expensive lab experiments or extensive simulation time to achieve high quality, preventing its use in real-time monitoring and control. Moreover, the scarcity of quality assessment procedures leads to non-traditional quality control procedures. Accordingly, this study developed an inductive data-driven framework and implementable methods to predict, optimize, and simulate additive manufacturing part parameters. The developed frameworks consist of three data-driven methods: (1) Reduced-order modeling using the self-organizing map to compress the high-dimensional spatial field data into a reduced and manageable space. Self-organizing map expands space exploration and minimizes the computational efforts towards building a predictive model for the targeted field; (2) Group methods for data handling variants as the core engine in model predictions, presented in the polynomial neural network cascades. The extracted polynomial neural network can relate the input variables (additive manufacturing process parameters) to the high-dimensional outputs (temperature field, distortion field, and residual stress field, etc.); (3) Evolutionary algorithms variants to explore a more complex heuristic search for the best program to optimize the polynomial neural network architecture with minimal human interference. Moreover, the evolutionary algorithms derive prediction-based multi-objective optimization and compensation strategies. The results from different data-driven techniques correlate well with the physics-based high-fidelity finite element simulations. It should be noted that while the framework is currently applied to finite element simulation data, the integrated data-driven methods can also be deployed on experimentally measured 3D cloud data or other high throughput measurements. This study provides a feasible and reliable approach for additive manufacturing process modeling and optimizations.