Application of Autoencoders for Additive Manufacturing

The construction of complicated designs via AM is made possible by the fact that elements are "constructed" from the ground up enabling the engineer to specify the design problem in terms of the intended functionality. While there is a high rate of acceptance in academia and industry as a result of the research field's growth, there is a driving demand for new algorithms in the design process for advanced mechanics and innovative production techniques in order to enable the full potential of the AM. One of the key drawbacks has been the intricacy of the solution topologies, with many critics claiming that, while theoretically optimal, the solutions are frequently difficult or impossible to produce resulting often in compromised functionality of the product. We propose AI-based tool commonly referred to as Autoencoder (AE) composed of two separate neural networks for overcoming these shortcomings, first an encoder for compressing the complex design to a lower dimension, and the second one, a decoder for reconstruction of the compressed design. The goal is to construct an AI-based model, or the mathematical representation of the design and the complex physical system that can then be used for instant predictions of new designs given a set of new conditions.

A typical design space of a given geometry (Kaveto) is shown in Fig.1 on the LHS whereas the optimized geometry, for the given set of boundary conditions (BC:s), is shown on the RHS.

The conventional topology optimization as shown in Fig.1 is a powerful free-form tool based on Finite Element Method (FEM) which is commonly utilized in engineering for evaluation of a technical design. While the resulting geometry fulfils the BC:s, here a set of mechanical loads, the manufacturability of the part is excluded and additional reverse engineering is needed for a final part to be manufacturable. Ideally, AM-operating conditions should be included in order to prevent subsequent part failures. These operating conditions are essential across different manufacturing method but especially important for metal powder 3D printing where the heating and cooling during the printing process could cause severe damages due to development of internal stresses. These may also include laser power, scan speed, layer- and hatch extensions, which all can have fundamental importance on the final part. The design may also require additional, sacrificial support for anchoring the part adding additional time and economical cost to the AM-process.

The first step towards fully integrated AI-driven AM process is inclusion of the computer-aided design (CAD) model into the process. AI, or more specifically ML, are set of methods for learning and making inference given a set of training dataset. Among ML methods, Neural Networks (NN) are one of the most popular ML techniques which are currently undergoing rapid development due to the vast amount of data that is currently available, the abundance of processing resources, and its sophisticated algorithm structure with applications ranging from computer vision, natural language processing and voice recognition to autonomous driving.

As an initial step for incorporating the physical process of AM, we utilize autoencoder (AE) which are a specific type of NN to automatically parameterise the complete data of the 3D design as shown in Fig.2. The AE then learns how to compress the geometric shape information into a “compact summary” or a set of latent parameters which are the key aspect of a given design, the DNA of a 3D-design. As each design has specific set of latent parameters, i.e., the most representative set of parameters these can then be used to further assessment and generating new geometries, commonly referred to as generative autoencoders. In this case, the 3D mesh is represented by a structured grid of points with are then used for training the NN. AEs are considered as unsupervised learning technique since it does not require explicit labels to train on and depth of the NN can be adjusted as well as the number of time-steps and latent parameters. Here, two layers are used for the depth of the encoder and the decoder with total of 3000 training steps.

The performance of the model is shown in Fig.3 where the loss function, key latent parameters and visual assessment is shown. The loss functions, or reconstruction loss in this case, quantifies the reconstruction of the resulting design when compared to the original design. The key latent parameters are shown in the middle of same figure allowing the designer to study the impact of each individual feature, here, the location of the widest arm, location of the largest load and the thickness of these two. These are of high importance is quantification of the contribution of each variable enabling the designer to understand why the design functions as it does On the RHS in Fig.3 a visual assessment the encoders capability of regenerating the original design is shown, indicating that additional training of the model could improve the accuracy of the reconstruction.

Since the quantified model is still well represented it can now be used for coupling the geometry to latent parameters which in turn can be used to generate completely new geometries. These latent variables may also have combined influence on the surface morphology. What is evident is that the initial surface morphology can be captured with the model allowing for machine (operating) parameters to be added and quantified as well as the physical processes. This approach is thus paving the way for incorporating the operating conditions as well as the physics-driven methods. In order to validate the current method a case study is currently performed at the RISE AM-centre assessing effectiveness of the method.