Behind this paper lies a long-standing scientific frustration that many materials engineers quietly share: ceramic-reinforced metal matrix composites promise extraordinary mechanical and thermal performance, yet they remain among the most stubborn materials to join reliably. These composites are designed precisely to resist deformation, wear, and heat, which makes them perfect for aerospace, automotive, and defense components but these same qualities turn welding into a delicate balancing act between plastic flow, particle damage, and tool degradation. Friction stir welding emerged as the best available solution because it avoids melting and suppresses brittle intermetallic formation, but even this solid-state process becomes disorderly once hard ceramic particles disrupt the flow of metal under the rotating tool. My motivation for this paper came from repeatedly encountering this mismatch between industrial promise and manufacturing uncertainty, where small changes in rotational speed, traverse rate, plunge depth, or tool geometry could cause catastrophic tunnel defects, particle clustering, or excessive tool wear. The question that shaped this work was simple but demanding: can artificial intelligence learn the hidden, nonlinear rules that govern friction stir welding in such complex, heterogeneous materials, and in doing so help transform composite welding from an empirical craft into a predictive science?
The literature offered a paradox. On one hand, artificial intelligence and machine learning have rapidly advanced in welding science, with neural networks, support vector machines, random forests, and deep learning being used to predict tensile strength, hardness, and defect formation in conventional aluminum and steel alloys. On the other hand, ceramic-reinforced metal matrix composites arguably the most difficult welding systems of all remained at the fringe of AI-driven research. Most models were trained on monolithic alloys where material flow is relatively smooth and thermal fields are easier to interpret. In composites, however, ceramic particles fragment, cluster, block flow streams, and accelerate tool wear in ways that are not just nonlinear but often discontinuous. This review was therefore conceived as a focused corrective to that imbalance. Rather than surveying artificial intelligence in welding at large, the paper deliberately zooms in on friction stir welding of ceramic-reinforced composites, where AI is most needed but least mature. What quickly became clear during the review process is that the primary bottleneck is not the sophistication of machine learning algorithms but the scarcity, fragmentation, and lack of standardization in experimental datasets for composite welding.
From a materials science perspective, ceramic reinforcements introduce an additional hierarchy of length scales and physical mechanisms into friction stir welding. The rotating tool does not simply stir a metallic continuum; it must continuously redistribute brittle, non-deformable particles whose size, volume fraction, chemistry, and interfacial bonding dictate the weld’s fate. Poor dispersion leads to vortex-driven particle bands and weak zones, excessive fragmentation degrades load bearing capacity, and thermal mismatches induce residual stress and microcracking. Traditional process windows defined for aluminum alloys collapse under these composite specific interactions. This is precisely where artificial intelligence becomes more than a convenience, it becomes a necessity. Machine learning models are uniquely suited to mapping interactions that cannot be reduced to simple equations, including the coupled effects of reinforcement morphology, tool geometry, heat input, and strain rate on joint integrity. One of the most important insights that emerged from this review is that composite welding models must move beyond treating reinforcements as passive inclusions and instead encode particle behavior itself as a dynamic feature in AI frameworks.
The paper shows how supervised learning models such as artificial neural networks, support vector regression, and ensemble learners already demonstrate impressive accuracy in predicting strength and hardness in composite welds when trained on carefully designed datasets, while deep learning architectures such as convolutional neural networks enable automated detection of voids, cracks, porosity, and lack of bonding defects directly from images. These capabilities are not merely analytical conveniences; they offer a path toward real-time quality assurance in environments where manual inspection is too slow and too subjective. At the same time, physics informed neural networks begin to bridge the long-standing divide between empirical data and thermomechanical theory by forcing machine learning models to respect conservation laws, constitutive behavior, and heat transfer constraints. This hybridization of physics and data is especially critical in composite welding, where purely data-driven models risk overfitting small datasets and failing when transferred between different reinforcement systems.
Yet the review also makes clear that today’s AI implementations in friction stir welding of composites remain largely offline and diagnostic rather than adaptive and autonomous. Most studies stop at predicting joint properties after the fact or identifying defects in post-process images. True intelligence in manufacturing will only emerge when these predictions feed back into the process itself in real time. That vision points directly toward digital twin architectures, where a virtual welding process evolves in parallel with the physical one, continuously updated by sensor streams and refined by AI models trained on both experimental and simulation data. Such systems would not merely forecast weld quality; they would actively regulate rotational speed, plunge depth, and travel rate to suppress defects as they form. This shift from prediction to autonomy represents one of the most important future directions identified in the paper.
A recurring theme throughout the review is trust. Engineers will not hand over process control to opaque algorithms unless they understand why predictions are made and how sensitive they are to specific inputs. This is why explainable artificial intelligence emerges as a central scientific and industrial requirement. Methods such as SHAP and local interpretability frameworks can reveal how reinforcement fraction, particle size, tool speed, or axial load contribute to predicted strength or defect probability. Without such transparency, even highly accurate models remain scientifically unsatisfying and commercially risky. The review also highlights uncertainty quantification as a missing but essential component in today’s welding AI landscape. Composite manufacturing is inherently variable due to powder processing, particle dispersion, and thermal gradients, and AI models must be able to express not only what they predict but how confident they are in those predictions.
Beyond the algorithms, the paper underscores a deeper systemic challenge: the fragmentation of data across laboratories, industries, and proprietary platforms. Ceramic-reinforced metal matrix composites differ widely in chemistry, particle architecture, processing route, and welding hardware, making it difficult to align datasets into unified training corpora. This is why federated learning and transfer learning appear as promising strategies for the next generation of composite welding intelligence. Instead of centralizing data, models can learn across distributed datasets while preserving confidentiality, gradually building generalized intelligence across multiple composite systems. In this sense, the future of AI in friction stir welding is as much about data governance and collaboration as it is about computing power.
Writing this review also reinforced how deeply intertwined materials science and artificial intelligence have become. Welding is no longer only about heat input and plastic deformation; it is now equally about feature engineering, dataset curation, model interpretability, and algorithmic validation. The traditional separation between manufacturing science and computational intelligence is dissolving, and this creates both an opportunity and a responsibility. The opportunity lies in accelerating discovery and deployment of advanced joining technologies; the responsibility lies in ensuring that models remain physically grounded, interpretable, and robust under industrial conditions. The review therefore does not present AI as a replacement for scientific understanding but as a powerful extension of it an instrument that amplifies what we already know while revealing patterns we could not otherwise detect.
Ultimately, this paper is about more than friction stir welding or artificial intelligence alone. It is about the broader transformation of manufacturing into a data driven, adaptive, and increasingly autonomous discipline. Ceramic-reinforced composites represent one of the most demanding test cases for that transformation, because they push both materials physics and machine learning to their limits. If AI can successfully tame the complex interactions of ceramic particles, metal flow, heat transfer, and tool dynamics in friction stir welding, then the same principles are likely to propagate across additive manufacturing, solid-state joining, and hybrid processing routes that define next-generation manufacturing systems. This is why the results and gaps identified in this review extend well beyond one welding technique. They point toward a future where machines do not simply execute pre-set instructions but continuously learn how to join materials more intelligently than any static process window ever could.