Artificial Intelligence and Machine Learning in Composite Manufacturing: Recent Advances, Emerging Trends, Digital Twins, and Future Perspectives

1. Introduction

Composite industries offer high strength, high stiffness, corrosion resistance, design flexibility, bespoke fiber architecture, and applications in aerospace, automotive, wind energy, hydrogen energy storage, marine, infrastructure, and high-performance sports ^1–3. However, this personalization also creates production control challenges that depend on the enabling process’s sensitivity ^4–6. Defects can include voids, dry patches, wrinkles, tow gaps, overlaps, porosity, delamination, fiber waviness, and dimensional deviations, and can result from variations in fiber placement, resin viscosity, permeability, temperature, pressure, consolidation, cure kinetics, and environmental conditions ^4,7–9. Quantitative data from fiber-reinforced thermoplastic extrusion production indicate that, in some cases, there may be a porosity of up to 10-15%, and manufacturing errors may affect tensile strength by up to 30% and fatigue life by up to 20% ¹⁰. Automated manufacturing is increasing, adding more heterogeneous process data from AFP, ATL, RTM, resin infusion, and NDE systems. They include machine-state, thermal, force, vision, geometry, pressure, flow-front, cure, ultrasonic, thermographic, radiographic, and optical measures ^11–13. Conventional empirical models and high-fidelity simulations remain important tools for process knowledge and prediction, but complex multiscale and transient behavior can challenge modeling of automated fiber installation. AI/ML offers complementary data-driven approaches to identify non-linear relationships between process parameters, sensor measurements, manufacturing defects, and final material attributes ^14,15. Predictive modeling has also been applied to optimize machining performance during the drilling of graphene oxide/carbon fiber-reinforced polymer nanocomposites ¹⁶. Recently, several studies have explored AI/ML applications for composite defect detection, mechanical property prediction, manufacturing process optimization, microstructure analysis, and material design ^17,18. The 2024 review, ML Applications in the Inspection of Composite Defects, reviews ML techniques for defect identification and prediction, including supervised learning, deep learning, and other data-driven approaches. Quantitative results for manufacturing fiber-reinforced thermoplastic materials further show that, as reported in the surveyed literature, ML defect-detection performance may improve by about 25–30% in several experimental situations ¹⁰. Figure 1 presents the integrated framework for intelligent composite manufacturing. It is an AI/ML-based system linking material architecture, process parameters, sensor/NDE data, forecasting, and manufacturing decisions. The framework begins with selecting the composite material and fiber architecture, then tracks and controls key process parameters, including temperature, speed, feed, pressure, and processing time. The AI/ML models use process and quality data from the sensor and NDE systems to predict problems, assess quality, and optimize operations. The model outputs support manufacturing decisions, such as selecting optimal parameters, ensuring quality, and improving process control. A feedback loop enables ongoing parameter adjustments and defect mitigation as forecast outcomes are fed back into the production process, supporting adaptive control, improved product quality, and more dependable composite manufacturing ^10,12. However, the results are transferable only to a certain degree, depending on the material system, sensor type, data volume, defect type, and validation technique. Therefore, this review focuses on AI/ML applications in composite production and highlights the links between process monitoring, defect prediction, process optimization, and intelligent process control. The following four questions were examined: (i) What are the contributions of AI/ML approaches in manufacturing? (ii) Which algorithms and sensing modalities are most suitable for particular composite production processes? (iii) What are some of the issues that hinder the transition from lab demonstration to a robust manufacturing system? (iv) What are some new research directions that will shape smart manufacturing composites?

This review highlights recent advances and trends in applying AI/ML to composites manufacturing. Section 1 presents an overview of the background, importance, and challenges of composite manufacturing and the role of AI/ML in process monitoring, defect prediction, and intelligent process control. Section 2 describes the AI/ML environment for composite manufacturing, including supervised learning, deep learning, reinforcement learning, physics-informed machine learning, and digital twin technologies. Section 3 describes AI/ML applications in major composite manufacturing processes, including AFP, ATL, RTM, resin infusion, compression molding, pultrusion, and additive manufacturing (AM) of fiber-reinforced composites. Section 4 reviews intelligent inspection and flaw identification using machine vision, infrared thermography, ultrasonic testing, and other non-destructive evaluation methods. Section 5 discusses emerging AI trends, including physics-informed AI, transfer learning, multimodal learning, explainable AI, federated learning, digital twins, surrogate models, and closed-loop control. Section 6 highlights problems and research gaps in data availability, domain shift, model interpretability, sensor integration, validation, and industrial implementation. Finally, the paper describes major findings and recommendations for dependable, sustainable, and intelligent composite manufacturing systems and suggests future research topics.

2. AI/ML Landscape in Composite Manufacturing

AI in manufacturing can be seen as having several levels of intelligence. The descriptive level classifies or visualizes process states using algorithms. Predictive systems forecast future process states, such as defect probability, the resin-flow front, void fraction, and temperature history. In prescriptive systems, the process parameters are recommended, and in closed-loop systems, the process parameters are communicated to the equipment, and the recommendation is updated continually. Separating predictive and validated process-control actions matters because many published composite studies remain at the predictive level, although industrial value increases when predictions are linked to verified process-control activities. The main families of ML are conventional supervised learning, deep learning, unsupervised learning, reinforcement learning, and hybrid physics-data techniques. If data are limited and the input-output relationship is clear, supervised regression models such as random forests, gradient boosting, support vector regression (SVR), Gaussian processes, or multilayer perceptrons are attractive. Recurrent networks and transformers can exploit temporal sensor sequences, whereas convolutional networks are particularly well-suited to images and spatial maps. Unsupervised approaches can be useful for ultrasonic or thermographic measurements where labeled defect information is minimal. Reinforcement learning has become an increasingly appropriate approach to adaptive control, where the process behavior is learned from a reward function. One major shift is from isolated models to integrated digital workflows. Recently, researchers developed a PixelRNN metamodel to predict RTM filling patterns, achieving up to 97.35% prediction accuracy while reducing computation compared to exhaustive physics-based simulations ¹⁹. After training, these surrogate models can quickly evaluate thousands of parameter combinations, making this viable. The next stage is to insert the surrogate into a digital twin that continuously exchanges data with the actual machine and the virtual process model ²⁰. Physics-informed ML is especially crucial for composites, since purely data-driven models may extrapolate poorly when composite material batches, shapes, sensors, or processing methods change ^20,21. Recent manufacturing studies show that physics-informed models can improve data efficiency and physical consistency by embedding governing relations or constraints within learning systems ²¹. These constraints for composite production can include relations governing heat transfer, resin flow, cure kinetics, fiber placement geometry, conservation laws, or relationships between process parameters and defect generation, as described in the literature ²¹.

The association between AI/ML approaches, process data, physics-informed models, and intelligent control in composite production is shown in Figure 2. The framework progresses from descriptive intelligence (manufacturing condition identification and visualization) to predictive intelligence (defect prediction and process behavior), prescriptive intelligence (optimal process parameter recommendation), and closed-loop control (continuous process adjustment). AFP/ATL, RTM, infusion, autoclave curing, and NDE systems provide diverse data for AI/ML technologies, including supervised learning, deep learning, unsupervised learning, reinforcement learning, and hybrid physics–data models ^19,21. These data support feature engineering, model construction and surrogate modeling, defect prediction, process optimization, and quality assessment. Physics-informed ML and digital twins also allow the addition of controlling principles such as heat transfer, resin flow, cure kinetics, and fiber-placement geometry, leading to more realistic and valid models and better process understanding ^20,22. Table 1 presents a typical role of AI/ML in composite manufacturing.

3. AI/ML Applications Across Composite Manufacturing Processes

3.1. AFP and ATL

AFP and ATL are among the most strategically important processes because of their many process variables and the high-value composite structures they produce. Steering, gaps, overlaps, tow waviness and wrinkles, insufficient consolidation, heat variations, and machine tolerances can all create faults. The AFP manufacturing review shows that key manufacturing parameters such as consolidation force, lay-up speed, temperature, and steering-control flaws affect part quality, and it emphasizes the importance of in-situ monitoring and ML-based prediction ^13,37. Computer vision may be particularly useful for visually detectable gaps and overlaps. ML inspection can provide learned representations for categorizing defect classes under different lighting, geometry, and texture conditions, rather than relying on hard-coded lighting, geometry, and texture criteria. The thermal sensor also provides supplemental information, as the thermal state of the input tape and substrate affects heating and consolidation conditions ³⁸. In recent years, research has increasingly focused on multimodal surrogate ML for automated composite manufacturing, including reinforcement-learning-based thermal control, ML-driven laser process control, process-structure-property modeling, point-cloud defect identification, and physics-informed digital twins ¹². The long-term goal is a closed-loop AFP cell. Here, a vision system identifies a gap or overlap, a thermal model correlates this gap/overlap with local process parameters, and a controller adjusts lay-up speed, compaction, heating power, or path strategy. The key challenges are not the ML algorithm itself but its seamless integration with machine control and its consistent performance across new geometries and material batches.

3.2. RTM, Resin Infusion, and Liquid Composite Molding

RTM and infusion are greatly influenced by the resin flow, permeability, viscosity, temperature, pressure, and cure kinetics. Traditional simulation can be time-consuming when considering many combinations of gate positions, injection conditions, tool shape, and material parameters ³⁹. ML surrogate models can approximate the mapping from process inputs to filling patterns, pressure histories, and/or defect probability ⁴⁰. The 2024 PixelRNN investigation exemplifies this, using image-based ML to simulate mold filling. The recorded prediction accuracy reached 97.35% using 50% of the available training data and at nearly half the cost of comprehensive simulations ¹⁹. This means ML can serve as a tool for rapid exploration of the design space, rather than replacing physics-based simulation. Hybrid workflows are extremely tempting: a smaller set of high-fidelity simulations/experiments provides training data, and the surrogate model then expedites optimization. AI can also be used in resin infusion to track the flow front with cameras or dielectric/pressure sensors, detect infilling abnormalities, and assess the likelihood of dry areas. A potential opportunity is multimodal learning, where flow-front images, pressure, temperature, resin viscosity, and permeability estimates are combined into a single model. Such models could convert the infusion from an open-loop monitoring system to an adaptive infusion control system ^41,42.

3.3. Compression Molding and Thermoplastic Composite Processing

Compression molding is a manufacturing method that involves substantial interactions between temperature, pressure, fiber content, material viscosity, consolidation time, cooling rate, and tooling ^43–45. In recent years, research has focused on optimizing the process through pressure, temperature, clamp tonnage, cooling rate, automation, and AI-based optimization to ensure process control and part quality ⁴⁶. ML can anticipate thickness, void content, warpage, mechanical performance, and cycle time, and optimize multiple targets simultaneously ^47,48. Thermoplastic composites with heating and cooling histories provide further opportunities. From these data-driven models, we can correlate thermal histories with end-of-process parameters and establish processing windows. The model, however, should account for batch-to-batch fluctuations and the machine’s thermal behavior. Consistent datasets that record material grade, fiber architecture, temperature history, pressure, cooling rate, and measured quality indicators considerably improve transferability ⁴³.

3.4. Pultrusion and Continuous Composite Manufacturing

Pultrusion is a continuous process where fiber reinforcement is impregnated, molded, and cured through a die as it is pulled ^49,50. Process stability depends on pulling speed, resin temperature, die temperature, resin viscosity, fiber volume fraction, tension, and cure state ^51,52. In continuous processes, sensors can capture long time-series data, and the manufacturing objective of maintaining process state makes this well suited to AI/ML ¹¹. These can be utilized to forecast dimensional variation, cure quality, pull force, surface flaws, and mechanical performance ⁵³. Anomaly-detection algorithms can alert operators to unanticipated departures from normal process operation before lengths begin to fail. Longitudinal validation studies are in high demand because models must be evaluated across large production runs and multiple raw-material batches, not just on randomly split laboratory data sets ¹¹.

3.5. Additive Manufacturing of Fiber-Reinforced Composites

In the developing field of composite additive manufacturing (CAM), AI can help address the process-structure-property uncertainty challenges associated with reinforcement orientation, interlayer bonding, voids, residual stress, thermal history, and process-induced anisotropy. A 2026 evaluation of CAM identifies nonuniform reinforcement dispersion and orientation drift, inadequate interlayer bonding, voids, residual stress, and property scatter as key industrial constraints. It also highlights transfer learning, domain adaptability, and self-supervised learning as potential avenues for addressing data scarcity ⁵⁴. AI can help in optimizing deposition parameters, bead geometry, raster orientation, temperature conditions, and tool path planning. It can also incorporate infrared imaging, machine vision, acoustic signals, and dimensional measurement to detect printing problems. In 2026, deep learning was combined with infrared thermography to quantify faults in additively built carbon-fiber composites, achieving mAP values of 99.3% for instance segmentation and image classification, with defect-area error within 6% of experimental settings. These results are encouraging, but they do not support a general conclusion about industrial accuracy, as lab data, controlled flaws, and constrained geometries can yield optimistic results ^55–57.

Figure 3 shows the performance of multiple composite manufacturing and AI/ML-based inspection approaches, such as RTM, CFRP NDE, DenseNet121, VGG19, and CAM IR. The results indicate that the methods perform very close to 100%, with CFRP NDE near 100%, DenseNet121 at 99%, and CAM IR at 98.5%. The RTM value is lower, at about 97%, compared with the others ^58–60. The error bars show the range of reported measurements and indicate variations in the performance of the methodologies studied. The results suggest that deep learning models and infrared inspection can assess composite manufacturing quality. However, the percentages reported may not be directly comparable, as each study uses specific datasets, evaluation criteria, experimental settings, and validation techniques.

4. Intelligent Inspection and Defect Detection

One of the most sophisticated uses of AI in composite engineering is defect detection, because NDE produces image-like or signal-rich datasets. Common faults include voids, delamination, debonding, inclusions, fiber waviness, gaps, overlaps, porosity, impact damage, and lack of consolidation. Manual interpretation is important and time-consuming but also vulnerable to operator variation; traditional inspection cannot be replaced ^13,61,62. Data classification and segmentation, as well as AI-based anomaly and defect quantification. Researchers developed a new method integrating thermography with deep learning to detect composite faults ^17,63. Bang et al. ⁶⁴ proposed a strategy that includes a deep-learning framework based on thermographic images. They tested their method on composite specimens fabricated by RTM and thermoplastic injection molding. Recently, researchers tested nine ML models based on flash-pulse thermography for automated CFRP impact-damage detection, with the bagged-tree model reporting 99.8% balanced accuracy when using the recommended preprocessing on the examined dataset. The results show that classical ML approaches without feature engineering and deep neural networks may be important ⁶⁵. Table 2 presents an AI/ML tool for detecting common defects and its deployments.

A 2024 study demonstrated the capacity of unsupervised ML to help detect issues in automated ultrasonic inspection of CFRP, addressing the problem of large automated scans that are difficult to interpret manually. Transfer learning has also been studied using multimodal ultrasonic images: a 2024 study reported DenseNet121 and VGG19 achieving accuracies of 98.8% and 98.6% on two datasets, respectively ⁶⁶. A trained model may not generalize well to a different scanner, frequency, fault geometry, or laminate design. Thus, future NDE work should focus on cross-machine validation, out-of-distribution testing, uncertainty estimation, and calibration, rather than accuracy on a single random train/test split ^67,68. Explainability is also vital, as it provides documentation for inspectors and the certification authority that an algorithm responds to physically meaningful fault signatures ^69,70.

5. Digital Twins, Surrogate Models, and Closed-Loop Control

AI and digital twins go beyond prediction toward intelligent industrial systems. A digital twin is a digital copy of a physical object or process that is automatically updated with data from the original. ML can speed up the twin by employing surrogate models to predict unmeasured states, detect deviations, and forecast future quality. Key factors in intelligent manufacturing in contemporary industrial digital-twin studies include position sensing, modeling and prediction, and decision-making ⁸⁴. The twin could encompass the entire composite manufacturing process, including material characterization, lay-up or mold filling, curing, machining, NDE, and structural performance ^85,86. AI can be trained to distinguish between simulated and real process performance ⁸⁷. The resin flow may be predicted by a physics model, and the systematic deviation can be learned by an ML model from permeability uncertainty ^88,89. Similarly, IR-observed data can tune an AFP thermal model (Liang et al., 2026; Lu et al., 2023). Next, we consider closed-loop control. The system does not anticipate a defect; it adjusts a process variable to prevent one ⁹². This demands very fast inference, strong uncertainty modeling, machine interfaces, safe operating envelopes, and explicit fallback procedures ⁹³. Reinforcement learning can eventually optimize sequential activities; first, simulate and learn actions under constraints, then confirm them with humans on expensive composite equipment ⁹⁴.

6. Emerging AI Trends

6.1. Physics-Informed and Hybrid AI

Most composite manufacturing involves multiphysics phenomena, and physics-informed ML is predicted to be a major approach. ML models approximate expensive submodels or unknown constitutive behavior and preserve physical relationships in hybrid models. A new cross-sector study of manufacturing identifies data scarcity, physical inaccuracy, and interpretability as key issues for purely data-driven techniques, and shows that PIML provides a way to create more data-efficient and physically realistic models ²¹.

6.2. Transfer Learning, Self-Supervised Learning, and Domain Adaptation

Transfer learning is of special interest, since the knowledge to build a composite may be hard to come by, even if only in name. Pretrained image models can be adapted to thermographic or ultrasonic data, and self-supervised learning can leverage the large volume of unlabeled sensor data. The CAM review (2026) indicates that transfer learning, domain adaptation, and self-supervised learning address data paucity and that, because the source and target processes are physically distinct, negative transfer can occur ⁵⁴.

6.3. Multimodal and Foundation-Style Models

Data in composite manufacturing are multimodal. The same manufacturing event can be described by an image, a thermal map, an auditory signal, a machine status, a process parameter, a simulation output, and a material certificate. Compared to single-sensor models, multimodal models can learn associations not visible to a single sensor. The current research approach is not just a larger CNN but a model that can also reason across heterogeneous input streams and link them to process and quality outcomes ⁹⁵.

6.4. Explainable AI and Trustworthy Manufacturing

Explainability is vital for safety and essential for composite manufacturing. Recent manufacturing evaluations cite model-agnostic methods as an emerging approach to handling black-box behavior, which can hinder confidence, safety, and acceptance in manufacturing ^96–98. In composites, explanations must be linked to engineering parameters, including temperature excursions, compaction force, fiber angle, flow-front behavior, and defect shape. A model that provides a forecast and the dominant physical/process characteristics supporting the prediction is more valuable for engineering decision-making ^96,99,100.

6.5. Federated Learning and Cross-Factory Knowledge

Sometimes, manufacturers don’t provide raw production data because of intellectual property, contractual, or cybersecurity reasons. Federated learning trains models on multiple datasets owned by different businesses without sharing data locally. Federated learning is proposed as a solution for smart manufacturing and product lifecycle management in 2025, especially when data is distributed and not available for centralized learning. Federated learning could enable composite manufacturers to build models to detect problems or optimize operations without revealing proprietary process knowledge to other manufacturers ^101–103.

7. Challenges and Research Gaps

AI/ML has made significant progress but remains in the middle stage of industrial maturity in composite manufacturing ^12,17,54. Algorithm availability matters less than evidence quality ¹⁰⁴. Constructing a dataset is difficult because these manufacturing datasets are small, proprietary, inconsistently formatted, and generated under narrow conditions. The data distribution can depend on material batch, fiber architecture, resin chemistry, machine settings, sensor calibration, environment, and geometry. The second problem is interpolation and extrapolation. A model can achieve good accuracy on training and testing data within the same experiment, but poor accuracy on a fresh machine or material batch. Robust research should report external validation, leave-one-batch-out validation, leave-one-machine-out validation, uncertainty intervals, and out-of-distribution testing. Third, much research focuses on improving forecast accuracy without measuring its effects on production. A 1% increase in classification accuracy does not justify adding a second or an expensive sensor ¹⁰⁵. Further work is needed to relate ML metrics to manufacturing KPIs such as scrap rate, cycle time, energy usage, material use, rework, dimensional tolerance, certified structural quality, etc. Fourth, certification and responsibility remain unresolved. Aerospace and other safety-sensitive applications require traceable proof, validation, and regulated alterations. AI models should therefore be understood as part of a quality assurance system and not as an effective and independent authority. Human-in-the-loop architectures, version control, uncertainty estimation, audit trails, and predefined acceptability thresholds will matter.

AI-enabled composite production can contribute to sustainability by reducing scrap, rework, energy use, material overuse, and wasteful inspection. Predictive models can detect unstable process conditions before defective parts are produced, and optimization algorithms can minimize cycle time or energy consumption while maintaining the required quality level. Recent assessments on AI and sustainable manufacturing have identified optimization, resilience, real-time monitoring, anomaly detection, and predictive maintenance as significant prospects within Industry 5.0 ^106,107. Large models, high-performance processors, and sensors might increase energy and hardware requirements. So, the right move now is to optimize across the lifecycle: compare the AI system’s energy and material costs with the costs of preventing scrap, rework, inspection, and process inefficiency. This is especially true in composite production, where raw materials and processing costs are often considerable, and recycling is not trivial. Industry 5.0 also emphasizes human-centric production. Operators and engineers in composite production have implicit knowledge about material handling, process abnormalities, tooling, and quality. Knowledge should be augmented by AI, not replaced. Human-in-the-loop systems can combine model recommendations with human expertise, and subsequent decisions can continue to contribute labeled data for ongoing learning ^107,108. Table 3 represents the research gap and the recommended direction.

8. Conclusions and Recommendations

AI and ML are emerging as crucial enabling technologies for the next generation of composite manufacturing. The best state of the art is in automated inspection, defect classification, process surrogate modeling, and parameter optimization. Efforts also exist in AFP/ATL, RTM, compression molding, CAM, and intelligent NDE. The most compelling results suggest that standard AI can reduce computation time, automate analysis of vast sensor data, and enable fast exploration of process windows. However, the field should not confuse high lab prediction accuracy with industrial readiness. Data scarcity, domain shift, sensor variability, explainability, cybersecurity, certification, and closed-loop validation remain important challenges. The most promising option is a hybrid paradigm that combines AI and physical models, digital twins, multimodal sensing, and human expertise. These can improve data efficiency, enhance physical credibility, and enable real-time decision-making. Over the next 10 years, optimizing a few parameters is expected to become more self-monitoring, adaptive, and autonomous in composite manufacturing. The fundamental research challenge is therefore to build reliable AI systems that understand the process, not just correlate the data. With benchmark datasets, rigorous validation methods, physics-informed models, and manufacturing-level KPIs, AI/ML can provide a feasible foundation for generating higher-quality, lower-waste, and more responsive composites ^12,17.

Table 4 summarizes the quantitative performance of AI/ML applications used for composite production (process modeling, non-destructive evaluation (NDE), and additive manufacturing). The results show that data-driven methodologies can improve analysis and characterization of manufacturing problems. The PixelRNN surrogate model achieved a maximum accuracy of 97.35% when trained on 50% of the training data, and it took about half the computation time of exhaustive simulations for resin transfer molding (RTM) ¹⁹. The machine learning model obtained the best performance among the models for thermographic impact-damage detection, with balanced accuracies of 99.8% for the machine learning model, 98.8% for DenseNet121, and 98.6% for VGG19, respectively ^65,66. The deep learning and infrared thermography methods of the authors give a mean average precision (mAP) of 99.3% and an error in defect area within 6% under the experimental circumstances in composite additive manufacturing ⁵⁵. These quantitative results demonstrate AI/ML’s potential to support surrogate modeling, defect detection, and process-quality assessment across a broad spectrum of composite manufacturing methods. However, comparisons based on reported performance values are limited because datasets, sensor types, defect types, and performance indicators differ. Future research needs to be directed toward generating consistent datasets, machine-to-machine validation, physics-informed learning, and metrics for performance at the manufacturing level, such as defect reduction, computational efficiency, cycle time, and energy consumption, to validate the industrial reliability and value of the AI/ML assisted composite manufacturing systems.

Future research should move from proof of concept to validated manufacturing environments. First, gather standardized datasets that include all metadata, such as material batch, fiber architecture and architecture parameters, sensor configuration, geometry, machine type, and quality outcomes. Second, validate models using external validation rather than only random train/test splits. Third, develop physics-informed and multimodal models to improve transferability. Fourth, include AI in digital twins and restricted control systems, and demonstrate measurable savings in defects, cycle time, energy, and/or prices. The medium-term objective is to construct a manufacturing cell that can sense process conditions, anticipate quality, determine the underlying cause of any deviation, and recommend and/or implement corrective action. Longer-term systems could involve generative design and autonomous process planning, foundation-style multimodal models, federated learning, and ongoing digital-twin updates. Work has already begun on a unified approach to customizable CFCs, in which AI explores combinations of architecture and process that are difficult to evaluate by hand.