Autoclave composites curing (as a sample process selected for this study) is known to be a highly complex, non-linear and multi-physics manufacturing system [1] [2], fraught with aleatoric and epistemic uncertainties that ultimately can result in risks that undermine the consistent production of high-quality aerospace parts. To increase the robustness of this process, a framework of regulations, guidelines, best-practice handbooks and other sources currently provide manufacturers with a tractable methodology for assessing and deciding on parts acceptability [9]. To this end, a key process outcome for epoxy matrix pre-impregnated advanced composites for the aerospace industry, is thermal history [9]. It is important that manufactured composite structures undergo a specific thermal history, in order to ensure that many properties of the material evolve in a controlled manner throughout the production cycle. Figure 2 provides an example thermal envelope for a carbon-epoxy prepreg, consisting of heating ramp, hold and cooling ramp segments. This processing window describes the allowable range of temperatures that a part can experience at a given time in the processing cycle, to achieve minimum performance requirements, as per the associated material and processing specification. Two key processing outcomes that can be tracked to ensure that quality metrics are maximized from a thermal processing perspective are the “peak exotherm” and “steady-state lag” tempertures. The former describes the hottest temperature experienced by the composite material throughout the manufacturing cycle, due to the combination of external thermal energy imparted by the working fluid surrounding the part, as well as the internal heat generation due to cure. Conversely, the latter is the steady-state difference in temperature between the working fluid and the composite part at the end of the ramp, when the difference is at its maximum.

A major reason for the use of thermal history-based specifications is to ensure that defects, either physical, chemical, or structural in nature, can be minimized and that material property targets are met. For example, to allow for the removal of entrapped air and outgassing from the resin for the reduction of porosity, the resin viscosity must be sufficiently low for an acceptable amount of time, which is primarily a function of temperature and degree of cure [21]. With respect to the consequences of porosity on mechanical properties, interlaminar shear

strength can be reduced by approximately 7% for each 1% of void content present up to a maximum of about 4% [1], and experimental studies have shown that for a 1% increase in void content, the flexural strength, flexural modulus, and inter-laminar shear strength can decrease more than 5% [4]. As primary aerospace structures require porosity volume fraction Vf < 1%, it is critical that these process bounds be adhered to, as increased porosity causes a significant reduction in mechanical properties [9].

Despite the physics-informed need to measure, identify and control process pertinent variables to avoid the generation of defects, the challenge is further compounded by realities of the manufacturing and qualification processes themselves. Namely, the need for homogeneous heating of the parts, in combination with local temperature detection, to deliver the required information for any kind of machine learning approach, is severely limited, given the intrusive nature of qualified, yet highly-local, measurement techniques (i.e. thermocouples). Chen et al. 2020 employed three years of primary structure production data combined with a meta-learning approach composed of seven machine learning-based models, thus reducing over-fitting of the meta-model, for automatic feature extraction and thermal cycle outcome prediction and recommendation, representing the volume of data required for such a task [22]. Comparatively, vision and thermal inspection systems produce data that are prime candidates for data analytics and machine learning, given the high volumes of embedded information that can effectively train, and in turn be exploited, by such models [23] [24]. For example, Manohar et al. used high-fidelity spatial scanning historical data, associated with shims used in component assembly and tolerance-based fits, used in 54 representative Boeing commercial aircraft. This data was deployed to train models that identified the locations of optimal sparse sensors (forming optimal measurement sensor ensembles) for predicting gaps to be shimmed. These optimal sensor ensembles then maximally inform the lay-up patterns used in future aircraft using sparse optimization techniques [25]. Such analogously enabling technologies in the context of autoclave cure processing are currently absent and largely underpin the lack of direct progress in this area.

In pursuit of minimizing defects, the metrics of peak exotherm and peak steady-state lag temperatures have been useful as key process outcomes [9]. Further, statistical structural performance estimates are provided for design activities using the composite material/processing combination, based on data collected via the building-block approach (BBA) method for material and process qualification, often in the form of A-basis and B-basis values [9].

Within the current best practice, the combination of qualification frameworks, along with the statistically-derived design allowable, forms the static, closed-loop approach to address and minimize risk in this complex system, by simultaneously constraining the materials, processes and other pertinent features. This restriction on the design space for the material, process, equipment, tooling, consumables and other elements of the factory process, works to reduce variability and uncertainty. However, this approach is typically very onerous from a time and cost perspective, while also having limited flexibility once established [26]. Such inflexibility can present challenges for shop floor decision-making contexts, such as the planning of bus-stop autoclave cures, in which multiple parts of similar features (and hence a similar expected thermal response) are loaded together into the chamber and cured in the same cycle [26]. There are many historical and ongoing efforts in this area that seek to extend and optimize the same frameworks, including developing and integrating higher-fidelity numerical models to reduce the cost of the BBA [27], reducing the conservatism and hence cost of A-basis and B-basis design allowable [28], developing consortia to share the burdens of e.g. qualifying materials [29] or design best-practices [30], among others.

The use of i4.0 seeks to provide manufacturers with opportunities for assessing a dynamic production environment, by leveraging accrued and live data, for better in-situ decisions, optimizing factory performance. However, it is recognized that this is at odds with the above addressed statically-driven framework of composites materials manufacture design and process qualification. Namely, there is a recognized need for testing, validation and qualification of production models used in this environment, to ensure confidence in using the predictive models at the heart of any i4.0 system. Many such systems also use machine-learning-based models, which are often black-box in nature and present their own challenges in building confidence. This is particularly the case when the given actual system is highly complex and direct computational models (e.g. using finite element) may be highly complex, time consuming, or low in their fidelity for use in practice.

Ultimately, machine learning (ML) models created to describe complex systems, such as those in composites manufacturing, are intended for use as decision-aid tools by operators/designers. Forecasting of future states or designing, such systems are very costly, time consuming, and computationally intensive, often due to finite project timelines and technical constraints within industry. ML modeling can be effectively employed as an inexpensive, rapid and powerful tool to address these problems. Once validated they can also be used to reduce the number of experiments needed for process optimization, increase safety or reliability, forecast future events under different process configurations, etc. The validation of the ML models may be achieved with the results of laboratory-scale tests or actual industrial data [31] [32], which may also include big data [33]. Further, ML models trained from real data can capture higher-order effects, or implicitly account for changes in the system unforeseen by human operators [34], such as an uncontrolled decrease in the efficiency of inductive heating elements inside of an autoclave vessel, which would otherwise be neglected from the model due to complexity, or remain unaccounted for entirely.

However, many ML models are black-box in nature; that is, black-box models are observation-based, often using experimental or observed data [35]. Unlike white-box approaches, the latter models provide low granularity, where only the data points used to train the system are considered as ground-truths along with the continuous potential predictions, let alone considering the reliability of the data used to train the model in the first place. Similarly, the domain coverage for predictions can only be reliably extended to the domain of the experimental data used to train the model in the first place. Black-box models attempt to encode the relationships between variables, by minimizing or maximizing an objective function (e.g. mean-squared error between observations and predictions) through a training process (e.g. gradient descent implemented through backpropagation). This relationship may not even be physics-based, depending on the inputs and outputs of the system, as chosen by the user. Hence, disparate model inputs may be used, which have no inherent direct physical relationship that can be quantified, as is characteristic of heuristic models. Additionally, the accuracy of such models is subject to many constraints, including the volume, sparsity and reliability of the available data. Figure 3 provides an overview of the trade-off between different model types, between the black- and white-box regimes.

Despite the advantages of using black-box ML in production environments, significant challenges remain, including their lack of interpretability. As Rudin et al. 2019 claim, any model that automatically “learns” from a dataset and is not manually constructed, has the potential to be intractable at any point in its domain, by virtue of the model builder implicitly not being hands-on in that process [36]. This perspective may be particularly relevant for use cases with limited data available, due to there being less embedded statistical information, further relying on the performance of increasingly complex and high-fidelity models to compensate [37]. Recent developments seeking to retain the advantages of black-box ML, while increasing model confidence, include the strategic drafting of federal policy towards fostering scientific machine learning (SciML) in which scientific principles can be more tractably encoded into such models

[38]. Iten et al. 2020 demonstrated this approach by leveraging observational data to encode a deep ANN, parameterize the data via a latent representation layer, followed by decoder ANN, called SciNet [39]. This approach has been employed in process modeling for manufacturability assessments in the preliminary design of autoclave-cured composite laminates, which provided greater prediction accuracy of the thermal processing outcomes compared to traditional methods [40] [41]. Giving model designers the ability to direct models to learn from datasets predicated on physics-based relationships (e.g. non-dimensional numbers) allows for greater tractability and interpretability in how the model constructs its representations of the complex curing scenario.

Beyond capturing physics-based phenomena, mathematical processes computed by an ANN developed by Lample et al. 2020, are able to symbolically compute function integrals and solve complex differential equations [42]. AI interpretability tools have also emerged as methods for application-based investigation of black-box ML, leveraging approaches such as constructing local models, or Shapley values (solution concept in cooperative game theory) to provide users with expected values or losses attributed to changes within the feature space [43] [44] [45] [46]. Further, the integration of heuristic methods with classical machine learning algorithms has offered higher model performance in various applications. For example, Geirhos et al. 2019 demonstrated that in computer vision classification-based tasks, ImageNet-trained CNNs are strongly biased towards using information from textures, rather than edges and shapes. This is a fundamentally different strategy for classification, when compared to human behavioral evidence, whereby incorporating strategies more in line with the latter ultimately improved model performances [47]. This provides further impetus for deeper user involvement in the construction and evaluation of black-box models.

With the emergence of increasingly robust and versatile machine learning models and open-source tools for their implementation, there has been growth in the volume of publications in the open literature that investigate the use of ML models in composites manufacturing and design. Here, a selection of the research has been reviewed and the results discussed, in the context of this study. It should be noted that research into the application of ML specifically in composites manufacturing is sparse. As such, select systems have been successfully modeled using this strategy and few critical comparisons between studies can be made. Studies include modeling the production of carbon fibre [48] [49], optimization of the automated fibre placement process [50], modeling of the dynamic cure process [51], as well as the assessment of defects for quality control, namely fibre orientation [52] and delamination [53] [54].

Khayyam et al. studied the use of various predictive models that were applied to the carbon fibre production process, specifically the stabilization step of the process. Gaussian-Newton, Levenberg-Marquardt Neural Network (LM-ANN) and Taylor series methods were used to model and predict the Carbon Fibre heat of reaction and modulus, as process outcomes from the stabilization step [48]. This study was extended to include the polyacrylonitrile (PAN) precursor fabrication process [49] [50], with Support Vector Machine (SVM)-based models also studied in [51]. The latter studies demonstrated that the machine learning model choices can be well relevant to the larger distributed supply chains. Further, the inclusion of multi-objective design optimization techniques under considerations of uncertainty extended these models to broader industrial applications [52] [53].

Brüning et al. 2017 proposed a method towards the optimization of the automated fiber placement processes, using ML predictive models. The process variables considered include temperature, compaction pressure and lay-up head velocity. Multiple linear regression models were used for any predictive work, which was supported by the development of an XML database for in-situ process monitoring [54].

Modeling of the autoclave cure process using black-box approaches has received some attention. Namely, Carlone et al. 2018 showed that dynamic predictions of the process can be made using Recurrent Neural Networks (RNNs), using the Long-Short Term Memory (LSTM) architecture [55]. This model was trained on pseudo-experimental (virtual data) and was able to capture the time-dependent nature of this complex system. This was performed in the context of the composite parts containing complex 3D geometries and having different fibre-volume fractions, where the thermal history of the composite throughout the cure cycle, is the model output. In their study, the Levenberg-Marquardt (LM), Bayesian Regularization (BR), Resilient Backpropagation (RP), and Scaled Conjugate Gradient (SCG) training functions were compared, where the ANN model based on BR provided the best performance.

There has been greater interest in the use of ML for in-situ defect estimation for quality control purposes in composites manufacturing, which is reflected by the greater volume of papers in this area. For example, Schmitt et al. [56] studied the feasible production of fiber-reinforced composites through inline inspection with machine vision. This work used a Fourier transform approach to classify local regions of an image, to determine the fibre/textile type, in order to catch lay-up errors, where technicians or other soft factors may have led to the defect. The study was set-up as a preface to the future application of ML models, as no ML algorithms were performed. Later, the work in this area was further developed by Hernán et al. [57], by detecting delamination-based defects using infrared thermography. Using the thermography images of composites as model inputs, dynamic PCA and K-NN models were used in an unsupervised learning regime to determine structure in the data as a dimensionally-reduced input into further ML models. Specifically, the Radial Basis Function was used as the kernel function in a trained SVM model, to detect and quantify the defect depth in the composite parts. Further, Fernandes et al. [58] recently investigated defect characterization in infrared thermography-based non-destructive testing (NDT), supported by ML for the prediction of fibre orientation. Pulsed thermal ellipsometry was the NDT method employed to capture images of internal features of the composite, which was combined with ANN models to estimate the fiber orientation on the surface of inspected parts. The authors commented on the good accuracy and speed of their approach, which in the context of the current industrial landscape, is a core requirement for the successful implementation of Industry 4.0.

Research in the wider literature has demonstrated successful applications of machine learning as a means for the modeling of composites production processes, for both prognostic design or in-situ defect characterization purposes. While this provides basic confidence in the capabilities of this approach, there remains concerns with model tractability and explainability for use in industrial applications, resting in a user’s ability to independently demonstrate how and why black-box ML models generate their outputs. In this context, many gaps and opportunities remain for further research in building ML-based models for use in composite manufacturing environments, which meet the human-interpretable needs for tractable risk reduction. Recent advances in pre-processing datasets, such as physics-informed ML [40], has provided increased interpretability for model designers by directly influencing the learning process in a common physics-based framework, towards using models for interpolation-based predictions. Further, with the recent emergence of AI explainability tools such as Local Interpretable Model-Agnostic Explanations (LIME)-based toolboxes, or the Shapley value-based approach used by Fiddler Labs [43], there is an opportunity for further research in this field to address the above concerns, in both the construction and evaluation of models, and associated datasets for advanced manufacturing. These advanced, combined with government-led strategic efforts to actively foster applied research in this area, has supported the framework and impetus for integrating Industry 4.0 principles in advanced composites manufacturing.