necessity of AI/ML integration to AM processes is due to the contemporary need for reduced labor cost and time, digi- tization in AM, and massive data availability (Kumar et al. 2023). AI/ML can be integrated into different sectors of man- ufacturing. In design, AI/ML increases acceptance of novel approaches and saves time and resources. In production, application of AI/ML saves time and energy and avoids waste. Finally, smart manufacturing can be interpreted as application of AI/ML in assembly processes to adjust any error in real time. Addin et al. (2007) demonstrated the potential application of ML in material science and design. In their paper, the Naïve Bayes classification is used for deterioration detection in con- struction. Jin et al. (2020) indicated that an ML model based on real-time camera images and deep learning algorithms can detect different levels of delamination conditions in FDM and determine the tendency of warping before it actually occurs. This paper aims to survey the application of AI and ML for data processing in acoustic-based in situ monitoring of AM processes. First, an overview of the acoustic emission NDT method for in situ monitoring of AM processes is presented. Then, various AI/ML techniques used by different researchers and the outcome of their analyses are described. The paper concludes with a summary of the discussion, existing chal- lenges, and potential future work. Acoustic Emission for In Situ Monitoring in AM Acoustic emission (AE), also known as acoustic emission testing (AET), as a monitoring technology has been explored by several research groups (Koester et al. 2018a Wasmer et al. 2019 Wu et al. 2016). AE refers to the generation of elastic (mechanical) waves released by materials when subjected to an external impetus, such as raising the gas pressure inside a cylinder, stim- ulating a given structure will cause deformation inside of it, such as crack growth. Consequently, this will trigger the rapid release of stored strain energy as transient elastic waves, typ- ically from a localized source. Formally, AE refers to both the generation mechanism and the waves themselves (ASTM 2020). Rapid melting and solidification occurring during the AM pro- cesses is a significant potential source of elastic waves that AE can hypothetically detect (Morales et al. 2022). Rapid generation of defects, such as cracks or porosity, can also produce elastic waves in the form of AE. A standard AE setup includes a set of piezoelectric transducers coupled to a structure, connected via cables to a monitoring system that performs data acquisition and processing. The data is stored on a computer and can be visual- ized in real time for further analysis after testing is complete. For the sake of brevity, this paper will not go into further technical details of AE fundamentals (Hossain et al. 2020). Most AE systems use a hit-based mode, which identifies transient waves in the signal and extracts features from them. A small set of parameters can describe discrete AE, which is digital (Taheri et al. 2013). The most commonly used parame- ters are rise time, peak amplitude, duration, MARSE (measured area under the rectified signal envelope) energy, and (ring- down) counts, as highlighted in Figure 1. The rise time is the time it takes for the signal to reach its peak amplitude after the first threshold crossing (defined by the operator), measured in microseconds. The duration of the hit is the time measured (usually in microseconds) from the first to the last crossing of the threshold, after which the AE hit will remain below the signal detection threshold, which the user identifies. The duration is often measured in microseconds. Given reflection and other mechanisms in a specimen, AE systems use different timing parameters to compute rise time and duration. The burst signal energy, or MARSE, is computed by taking the integral over time of the squared electrical signal over its duration. Finally, ring-down counts are the number of thresh- old crossings of an AE signal. It is another valuable parameter to help distinguish between AE signals and background noise. Combined with other signal features, some or all of these parameters can be correlated with the AM process condition through statistical signal processing and ML techniques and used to identify potential discontinuities in the manufactured parts (Bond et al. 2019 Taheri et al. 2019). For instance, Li et al. (2021) observed that the AE signals collected over a laser-cladding AM process where cracks exist in the parts have larger amplitude and energy than AE signals collected over a normal cladding process. Hossain and Taheri (2021a) discussed the potentials, limitations, and opportunities of acoustic techniques for process monitoring of AM. In this paper, the authors highlighted the capability of acoustic tech- niques for volumetric quality identification and adaptability to various manufacturing techniques as the major promising features of acoustic techniques for in situ process monitoring for AM. These abilities have been investigated in various man- ufacturing processes, including but not limited to AM, by other researchers. Ramalho et al. (2022) showed that the influence of contamination in WAAM can be identified through the analysis of the acoustic spectrum of the process. Ramalho et al. aimed to establish a microphone-based acoustic sensing method for WAAM quality monitoring. WAAM parts were fabricated with ME |AI/ML Figure 1. A burst-type AE signal and associated features (from nde-ed.org). Rise time Counts MARSE Threshold Time Duration Threshold-crossing pulses out Comparator circuit A E s i g n a l i n T h r e s h o l d i n 52 M A T E R I A L S E V A L U A T I O N • J U L Y 2 0 2 3 2307 ME July dup.indd 52 6/19/23 3:41 PM Volts Amplitude CREDIT: CNDE

introduced material contaminations in Ramalho et al.’s work, and acoustic signals were recorded during the manufacturing process. Power spectral density (PSD) and short time Fourier transform (STFT) were used to pinpoint the location of dis- continuity formation (Ramalho et al. 2022). Active acoustic methods, or ultrasonic, have also been studied for in situ mon- itoring of the WAAM process. Hossain et al. (2020) designed a fixture to connect an ultrasonic transducer to the build plate of the WAAM system and keep it in constant contact during the manufacturing process. The features extracted from ultrasonic signals showed that there is a detectable difference between the values of root mean square (RMS), root-sum-of-square (RSSQ), and peak magnitude-to-RMS ratio (P2R), which was interpreted as the indication in process deviation from the typical window of WAAM (Hossain et al. 2020). The features extracted from AE signals can be correlated with the AM process condition through statistical signal processing and ML techniques or be used to identify the potential discontinuities in the manufactured parts. Despite the large amount of infor- mation that can be extracted from AE signals, challenges exist in interpreting the signals due to the potentially low signal-to- noise ratio (SNR) and significant variation in the magnitude or frequency of the AE signal over the monitoring period of the AM process. The literature discussed previously reveals that AE shows a promising ability to distinguish variations in the oper- ating conditions of AM systems, known as process conditions. The contrast between AM process conditions is the main cause of quality variation and changes in AM parts. Studies have also shown that AE not only distinguishes between contrary AM processing conditions, which potentially cause different types of defects, but also differentiates various levels of defects. As an example, Shevchik et al. (2019) showed that three levels of quality categories of AM parts manufactured by LPBF can be identified by detecting AE signals analyzed by ML techniques. In their study, quality categories are defined as high, medium, and poor corresponding to various levels of porosity of 0.07%, 0.30%, and 1.42%, respectively (Shevchik et al. 2019). Machine Learning Techniques for Acoustic Data Processing Massive datasets are ubiquitous across scientific and engi- neering disciplines in the current era, and this trend can be attributed to the meteoric rise in computing power over the past few decades. Consequently, applying ML algorithms to infer patterns and gain insight from these datasets has become a new mode of scientific inquiry (Brunton et al. 2020). The NDT industry is no exception to this trend, especially for AET. ML is a subset of AI and is usually divided into three main categories: supervised, unsupervised, and reinforcement learning. Several learning algorithms fall under each of these categories, and in the context of NDT, the fundamental task is to discover or find discontinuities in the specimen of interest. This section aims to avoid discussing ML jargon for brevity. Instead, this paper will elucidate the workings of selected ML algorithms relevant to AE testing as applied to AM. This paper will explain mathematical concepts with analogies, where nec- essary, to reach a wider audience. One of the challenges in AE signal processing is the high level of dependency on human expert participation. However, this could be a major limiting factor when AE is used for in situ monitoring and control of the manufacturing processes. Specifically, this can be an issue when instant and accurate feedback is desired. AE is a data-intensive technology and using ML algorithms to analyze large datasets is of consider- able interest to researchers and practitioners. Additionally, uti- lizing ML algorithms makes the technique more quantitative and less vulnerable to subjective judgments made by techni- cians and engineers when analyzing AE test data. However, despite the large amount of information that can be extracted from AE signals, challenges exist in interpreting the signals due to the potentially low SNR and a considerable variation in the magnitude or frequency of an AE signal over the monitoring period of the AM process. The forthcoming sections briefly discuss how classifiers using various ML techniques are built to help sort AE data obtained from AE systems in the context of AM. ML methods can handle these situations with reasonable efficiency. However, there are still some challenges associated with various ML techniques that must be resolved. Supervised Classification of AM Process States Supervised learning refers to a learning paradigm that requires prior knowledge of the answers to the problem at hand, which implies providing both the input data and the correspond- ing output labels when training the ML model. The model then learns a pattern to better predict or classify future data based on the knowledge from the examples during training. Supervised learning is analogous to a pupil learning a subject by studying a set of questions and their corresponding answers. Classes of problems that require supervised learning include regression and classification problems. Neural Networks This section provides an overview of neural networks, includ- ing the differences between artificial neural networks (ANNs), convolutional neural networks (CNNs), spectral convolutional neural networks (SCNNs), reinforcement learning (RL), and support vector machines (SVMs). ARTIFICIAL NEURAL NETWORKS ANNs are a commonly utilized ML architecture, modeled loosely on the human brain, mimicking how biological neurons communicate with one another. The perceptron, demonstrated by Frank Rosenblatt of Cornell in 1958, was the first trainable neural network (NN) (Rosenblatt 1958). However, it consisted of only a single layer, as opposed to the modern iteration of neural nets (also known as feedforward NNs), which have multiple layers of neurons (multilayer percep- tron, or MLP). Figure 2 shows a sample ANN with one input layer (with five neurons), two hidden layers (each with four neurons), and one output layer with two neurons. J U L Y 2 0 2 3 • M A T E R I A L S E V A L U A T I O N 53 2307 ME July dup.indd 53 6/19/23 3:41 PM