Description
Advanced carbon fiber reinforced polymeric composite (CFRP) materials possess high strength and sti↵ness properties and are lightweight. CFRP materials allow spatial tailoring of material sti↵ness and strength to meet the load requirements. Such spatially tailored material properties can be constructed by tow steering using Automated Fiber Placement (AFP) manufacturing. Parametrization of the material orientations for designing tow steered composites require a large number of design variables and result in non-convex optimization problems. Fiber or tow paths are generated from the optimized material orientation fields, as a post processing step, using tow placement path planning algorithms. Curvilinear fiber paths develop overlaps and gaps due to the difference of the curvatures mismatched between fiber tows. These gaps and overlaps decrease the strength of the composite structure. Optimization of composites for tow steered curvilinear paths therefore require manufacturing constraints that limit the overlaps and gaps. This thesis investigated Deep Neural Networks for developing surrogate models of the manufacturing constraints (overlaps and gaps) to replace the using of path planning algorithms in the design optimization directly.Feed Forward Neural Network (FFNN), Convolution Neural Network (CNN) and Recurrent Neural Networks (RNN) were investigated to determine the most appropriate model for the manufacturing constraint approximation. The input variables are the fiber orientation angles used to develop a spatially varying material orientation and the outputs are the areas of overlap and gap developed by tow steering. The methods are demonstrated for two examples; a flat rectangular plate (16 variables) and a rectangular plate with an elliptical hole (24 input variables). For the first example the RNN had the best prediction accuracy. For the second example, the accuracy of all DNN models were significantly less accurate compared to the first. However, both the FFNN and RNN had similar accuracy. The decreased prediction accuracy in the latter case is due to the sparsity of the sampling used and the highly non-linear and noisy nature of the output function. The current results from the flat plate problem demonstrate that Neural Networks can provide a computationally efficient and accurate surrogate model to implement manufacturing constraints for tow steered composites.
