Machine learning in additive manufacturing: State-of-the-art and perspectives

Abstract

Additive manufacturing (AM) has emerged as a disruptive digital manufacturing technology. However, its broad adoption in industry is still hindered by high entry barriers of design for additive manufacturing (DfAM), limited materials library, various processing defects, and inconsistent product quality. In recent years, machine learning (ML) has gained increasing attention in AM due to its unprecedented performance in data tasks such as classification, regression and clustering. This article provides a comprehensive review on the state-of-the-art of ML applications in a variety of AM domains. In the DfAM, ML can be leveraged to output new high-performance metamaterials and optimized topological designs. In AM processing, contemporary ML algorithms can help to optimize process parameters, and conduct examination of powder spreading and in-process defect monitoring. On the production of AM, ML is able to assist practitioners in pre-manufacturing planning, and product quality assessment and control. Moreover, there has been an increasing concern about data security in AM as data breaches could occur with the aid of ML techniques. Lastly, it concludes with a section summarizing the main findings from the literature and providing perspectives on some selected interesting applications of ML in research and development of AM.

Introduction

Additive manufacturing (AM) is a disruptive digital manufacturing technology to make 3D objects, usually layer upon layer, according to computer-aided design (CAD) models. Compared to conventional manufacturing technologies, it has the advantages of fabricating intricate parts with complex geometries and designs, unique microstructures and properties, as well as reduced lead time and cost. Therefore, in recent years, AM has attracted a great deal of research interest in both academic research and industrial applications worldwide.

According to the ASTM F42, AM processes can be broadly classified into 7 categories [1]. The AM techniques involving Machine Learning (ML) in this article mainly fall under 3 classes of technology, namely powder bed fusion (PBF), directed energy deposition (DED) and material extrusion, as they are currently the mainstream AM processes that have attracted great attention in both academic research and industrial applications (see Fig. 1). Although ML has also been applied in other AM processes such as materials jetting [2] and stereolithography [3], these research works are not very relevant to the focus of this article hence they are not specifically discussed here.

Under the PBF category, a laser or electron beam is used as the energy source to selectively melt powder bed which is uniformly spread by re-coating layer by layer [[4], [5], [6]]. In the DED process, a focused laser beam melts the continuous powder stream or wire which are fed from the deposition nozzle into the melt pool in order to fabricate near-net-shape parts [7,8]. One typical material extrusion process is fused deposition modelling (FDM). The filament is fed into the liquefier head by a driver gear and is subsequently heated to the glass transition state. The semi-liquid filament material is then deposited from the extrusion nozzle to print parts layer by layer [9].

ML is an artificial intelligence (AI) technique that allows a machine or system to learn from data automatically and make decisions or predictions without being explicitly programmed. In research, ML is gaining popularity in medical diagnostics [[10], [11], [12]], material property prediction [[13], [14], [15]], smart manufacturing [[16], [17], [18]], autonomous driving [[19], [20], [21]], natural language processing [[22], [23], [24]] and object recognition [[25], [26], [27]]. ML algorithms are commonly categorized as supervised, unsupervised and reinforcement learning.

Supervised learning enables a computer programme to learn from a set of labelled data in the training set so that it can identify unlabelled data from a test set with the highest possible accuracy [28]. The datasets can be in a variety of forms including forms of images [29,30], audio clips [31,32] or text [33]. There is an objective function known as cost function, which calculates the error between the predicted output values and the actual output values. In the training process, the parameters (or weights) between neurons in adjacent layers are updated in order to reduce the cost function after each iteration (or epoch) [34]. In the testing process, the previously unseen new data, i.e. test set, is introduced to provide an unbiased evaluation of the model’s accuracy.

Unsupervised learning infers from unlabelled data [35,36]. It is a data-driven ML technique which can uncover hidden patterns or group similar data together (i.e. clustering) in a given random dataset [37]. Unsupervised learning is widely used in anomaly detection [38], recommendations systems [39], and market segmentation [40,41].

Reinforcement learning is a semi-supervised ML paradigm which allows the model to interact with the environment and learn to take the best actions that can yield the greatest rewards [42]. It requires no training dataset, and the model learns from its own actions. Reinforcement learning is popularly adopted in robotic arms [43,44], autonomous cars [45,46], and AlphaGo [[47], [48], [49]].

The terminologies of the ML algorithms that are mentioned in this overview article can be found in the Appendix. Readers are suggested to refer to the recommended textbooks or research papers in this table to gain more insights into the details of each ML algorithm.

This paper aims to provide a state-of-the-art review of applications of ML techniques in various domains of AM production practices from the most recent literatures as well as our perspectives on some impactful research directions that are still on-going or may occur in the future. To clearly elucidate the benefits of using ML in AM, we broadly classify the applications into 3 categories, namely design for additive manufacturing (DfAM), AM process and AM production, as illustrated in Fig. 2. This is to reflect a logical sequence from design, process optimization and in-process monitoring, to manufacturing planning, product quality control and data security that are closely linked to overall production concerns.