Figure 1. Yearly total publications of seven composite journals since 2010, AI/ML-enabled publications as a percentage of total publications. Source (All Figures) | “Artificial intelligence and machine learning in composite materials: A comprehensive literature review and bibliometric analysis”
Artificial intelligence (AI) and machine learning (ML) applications in composites have grown significantly in recent years. AI/ML-enabled publications now account for more than 7% of the annual output in major composites journals, with applications extending across material design, processing and manufacturing, mechanics and performance, structural health monitoring and sustainability (Fig. 1). While part of this growth reflects the broader expansion of AI/ML across engineering disciplines, the composites community has also adapted these methods to address challenges that are specific to composite materials and structures.
To understand the current landscape, the University of Washington recently conducted a comprehensive review of AI/ML applications in composite materials1. The study included a bibliometric analysis of publications from seven major composites journals between 2010 and 2025. From a total of 50,361 publications, a targeted search identified 1,474 AI/ML-enabled papers. This was followed by an in-depth technical review of more than 200 representative studies covering different composites applications.
The analysis shows a clear change in publication activity around 2019. Before this point, AI/ML applications represented about 1% of annual publications in the selected journals. This share increased rapidly over the following years, reaching approximately 7.1% in 2025, corresponding to more than 300 papers in that year alone.
Why AI/ML is being used in composites
Despite the wide adoption of composites, significant challenges remain in material development, manufacturing, design, optimization and certification. Major progress has been made in high-fidelity, multiscale and multiphysics simulation tools. However, complex model calibration and high computational cost often limit their applicability for design-space exploration and optimization. The inherently stochastic nature of composites — stemming from material and process uncertainty, anisotropic behavior and nonlinear, multimode damage and failure — further compounds these challenges.
AI/ML is increasingly used to augment experimental testing and physics-based simulation with fast data-driven models. In many applications, these models can provide near real-time predictions after being trained using experimental or simulation data. This makes it possible to explore large parameter and design spaces that would otherwise require a large number of high-fidelity simulations or experiments.
Beyond accelerating forward analysis, AI/ML is also changing how inverse problems are addressed. In composites manufacturing and structural applications, important quantities such as process parameters, internal material states, damage conditions or spatially varying boundary conditions may not be directly measurable. AI/ML methods can use sparse sensor data or targeted experimental information to infer these quantities. Similar approaches are increasingly used for inverse design, where the objective is to identify material architectures, layups or processing conditions that satisfy a desired response.
Another important area is the analysis of image- and time series-based data. Thermography, X-ray computed tomography (CT), microscopy, digital image correlation, acoustic emission and guided waves generate large datasets that are difficult to analyze manually. AI/ML methods are increasingly used for image segmentation, defect classification, feature extraction and pattern recognition, enabling automated analysis and, in some cases, in-process monitoring and control.
Composite materials also present a particularly challenging problem for AI/ML. Material and process variability exist across multiple length and time scales. Variability at the constituent and microstructural scales can influence laminate properties and structural response. At the same time, material storage, manufacturing, testing and service introduce different time scales that can influence material state and performance.
Composites manufacturing also involves sequential processes. Variability in raw material state and processing conditions can affect microstructure and defect formation. Uncertainty introduced during an early processing step may therefore propagate through subsequent stages and ultimately contribute to panel-to-panel and batch-to-batch scatter in the final part. These characteristics become particularly important when AI/ML is used beyond prediction and optimization to support testing, qualification and certification.
AI across the composites life cycle
The university’s review shows that AI/ML has now been applied across the composites life cycle. We organized these studies into four main categories: Material Design & Properties, Processing & Manufacturing, Mechanics & Performance and Structural Health Monitoring & Sustainability. Under these four categories, 17 subcategories were identified, covering applications ranging from material design and process science to impact, fatigue, inspection and recycling (Fig. 2).
Figure 2. Taxonomy of composites research topics for AI/ML-enabled studies. Segment sizes are proportional to the number of AI/ML-related composite publications from 2010-2025.
In Material Design & Properties, AI/ML is increasingly used to establish structure-property relationships, predict properties from compositional and microstructural features, reconstruct microstructures from imaging data and support inverse design. Recent studies have moved beyond simple property prediction toward microstructure-aware models that use micro-CT, microscopy and related imaging methods to quantify features such as porosity, fiber orientation, tow geometry and phase distribution.
In Processing & Manufacturing, AI/ML is being used for process characterization, process simulation and optimization, in situ inspection, defect prediction and process control. Applications include automated fiber placement (AFP), additive manufacturing, resin infusion, autoclave processing, cure cycle optimization and process-induced deformation. The literature is also moving from deterministic black box surrogate models toward more physics-informed and uncertainty-aware approaches.
Mechanics & Performance has the largest AI/ML footprint. Applications include progressive damage and failure prediction, multiscale modeling, impact and crash response, fatigue, design allowables and inverse identification of material parameters. Much of this work uses finite element simulations to train fast surrogate models for damage prediction, failure envelope reconstruction and performance assessment.
The literature is increasingly moving from purely data-driven black box prediction toward approaches that incorporate available physics and uncertainty.
Structural Health Monitoring has a smaller publication footprint, but AI/ML has become an important tool for interpreting acoustic emission, guided-wave, ultrasonic, thermographic and embedded sensor data. Recent studies are moving from damage detection and classification toward damage severity, residual strength and remaining useful life. Sustainability, recycling and reuse currently contain fewer dedicated studies, although there is increasing interest in property screening and process-structure property relationships for recycled and bio-based composites.
These applications show that AI/ML is moving beyond accelerated simulation toward broader integration with manufacturing data, inspection and structural performance. Potential applications include reducing simulation and testing effort, supporting process optimization and control, and enabling earlier detection of manufacturing defects.
Where research activity is concentrated
Although AI/ML now spans the composites life cycle, the distribution of research activity is highly uneven. Mechanics & Failure has the largest footprint, comprising approximately half of all AI/ML-enabled composites publications since 2010. In 2025, this category accounted for approximately 54% of AI/ML publications. In contrast, topics such as Bonding & Joining, Thermo-Mechanical Properties, Thermal Protection and Sustainability have much smaller publication footprints (Fig. 3).
The strong concentration in mechanics and failure is not surprising, given the complexity of composite damage and failure and the computational cost associated with high-fidelity simulations. However, several other areas show growing activity and may be particularly important for industrial applications.
Figure 3. Annual publication counts of AI/ML-enabled composites studies by topic.
Defects & Quality is one example. Vision-based methods are increasingly used for in-process inspection and classification of defects such as gaps, overlaps, wrinkles and twists in AFP. More recent studies are moving from postprocess defect detection toward forecasting defect formation before deposition is complete, creating a pathway to proactive process correction.
Process Science is another area where AI/ML may have a significant impact. High-fidelity thermo-chemical and thermo-mechanical process simulations remain essential, but characterization and computational cost can limit their use for optimization. Recent studies have applied AI/ML to cure simulation, resin flow, process-induced deformation, pyrolysis and kinetics characterization. Physics-informed models, theory-guided approaches and probabilistic methods are increasingly being explored in these areas.
Thermo-chemical properties, thermo-mechanical behavior and thermal protection currently represent smaller but emerging areas of AI/ML-enabled composites research, with increasing publication activity reflecting growing interest in these applications.
What is still missing
Despite the rapid growth in AI/ML applications, a significant virtual-to-real gap remains. A model can achieve excellent accuracy on simulation data or within a limited experimental dataset and still fail to represent the variability encountered in the next material batch, panel or manufactured part.
Much of the literature relies on data generated from deterministic finite element or representative volume element simulations, or on relatively small experimental datasets. Models are then often evaluated within the same data space, with performance reported using metrics such as R² or RMSE. While these results are useful, they do not necessarily establish reliability when material systems, processing conditions, geometries, loading conditions or service histories change.
This limitation is particularly important for composite materials because variability is inherent to the material and manufacturing system. Material state, batch-to-batch variability, panel-to-panel variability, fiber orientation, porosity, thermal history and process-induced defects can all contribute to scatter in final properties and performance.
Composite materials present a particularly challenging problem for AI/ML [because] material and process variability exist across multiple length and time scales.
Another important limitation is that most studies focus on a single process step, material system, geometry or processing window. Composite manufacturing, however, involves sequential processes in which material and process uncertainties can accumulate and propagate across the process chain. End-to-end models capable of tracking this evolution remain limited. This is particularly important for qualification and certification, where the quantity of interest is not only the expected response, but also its variability and the conservative bounds required for engineering decisions.
Validation also remains an important gap. Random coupon-level splitting can overestimate model performance when coupons from the same panel or batch share process history and defect statistics. Depending on the application, validation should instead consider held-out panels, batches, geometries, loading paths, processing windows or sensor configurations.
Data traceability is equally important. For deployable AI/ML, datasets should preserve information on material batch, processing history, test conditions, inspection methods and service history. Composite response depends strongly on this life cycle history, and models intended for engineering use need to retain that context.
Where the field is moving
The literature is increasingly moving from purely data-driven black box prediction toward approaches that incorporate available physics and uncertainty. Physics integration is particularly important for composites because experimental datasets are often sparse, while considerable knowledge already exists through analytical models, process models, finite element simulations and constitutive relationships. Incorporating this information can reduce data requirements and constrain models in regions where data are limited. However, physics integration does not remove the need for independent validation outside the training domain.
A second major direction is uncertainty-aware AI/ML. For many engineering applications, particularly qualification and certification, a point prediction is not sufficient. Models need calibrated prediction intervals and, where possible, separation of uncertainty associated with limited knowledge from the inherent variability of the material and manufacturing process.
Recent studies are also increasingly combining probabilistic models with sparse in situ sensing, targeted testing and Bayesian updating. These approaches can reduce experimental burden while providing a practical pathway for updating models when new materials, processes or operating conditions are introduced.
Overall, the literature shows that AI/ML is becoming more widely used across the composites life cycle, while composites-specific requirements are also influencing the methods being developed and adopted. The field is moving toward models that incorporate physics, material and process variability, uncertainty and information from real manufacturing and testing.
This is especially relevant when uncertainty accumulates across multiple processing stages, length scales and time scales before appearing as variability in final material properties and structural performance. Accounting for this evolution will be necessary for reliable use of AI/ML in material development, manufacturing, testing, qualification and certification.
References
1H. Fu, A. Eskandariyun, P. Portales Picazo, K.A. Johnson, A. Morton, M. Wynn and N. Zobeiry, “Artificial intelligence and machine learning in composite materials: A comprehensive literature review and bibliometric analysis,” Composites Part A: Applied Science and Manufacturing, 210 (2026) 110092. DOI: 10.1016/j.compositesa.2026.110092.
About the Author
Navid Zobeiry
Navid Zobeiry is an associate professor of materials science and engineering at the University of Washington, with an adjunct appointment in aeronautics and astronautics. His research focuses on developing physics-informed and uncertainty-aware machine learning methods for composite materials and structures, with applications in advanced manufacturing, testing and characterization, process optimization and accelerated qualification and certification.
