Machine Learning-Assisted Polymer and Polymer Composite Design for Additive Manufacturing

Abstract

Additive manufacturing (AM) of polymers and polymer composites is changing how customized, lightweight, and complex parts are produced across various industries. However, predicting the final properties of printed parts remains challenging due to variations in material compositions, processing conditions, and microstructural characteristics. This review explores how machine learning (ML) is being used to address these challenges. It examines the application of various ML approaches in polymer and polymer composite design for AM, including supervised, unsupervised, semi-supervised, self-supervised, and reinforcement learning, for predicting key properties such as mechanical strength, thermal stability, and electrical performance. The review also highlights hybrid techniques that combine ML with physics-informed modeling, including the use of digital twins, to enhance AM process control. Challenges and future perspectives, such as data scarcity, model interpretability, and computational demands, are discussed. In summary, ML is showing strong potential to support faster, more reliable, and more sustainable development of advanced polymers and polymer composites for AM.

1. Introduction

Scientists believe that additive manufacturing (AM), or 3D printing, is reshaping material production by constructing complex objects with precision [1,2]. Unlike conventional methods that remove material through machining [3,4,5], AM can minimize material loss by building structures from digital models [6,7]. This approach has gained popularity in fields like aerospace, healthcare, automotive, and consumer goods production [8,9,10].

A major benefit of AM persists in its ability to create intricate shapes [11] that traditional techniques struggle to produce. Features such as hollow structures and lattices can be printed with improved efficiency while reducing material use [12,13,14]. Among all the materials used in AM, including biomaterials, cells, ceramics, composites, electric materials, metals and alloys, and polymers, polymers have been popular due to their lower cost, wide availability, multifunctionality, tailorability, and broad applications.

Polymer composites can be adapted for different uses. Standard polymer-based AM materials often fall short in strength, heat resistance, and functionality, but adding reinforcements like fibers, nanoparticles, or hybrid fillers helps address these weaknesses [15]. In addition, certain functional fillers can introduce features like electrical conductivity, heat resistance, electromagnetic interference shielding, and self-repair, expanding their use in advanced engineering applications [15].

Despite the benefits, predicting the properties of 3D-printed polymer composites remains challenging. Variations in processing conditions, like printing speed, layer thickness, and temperature, can affect mechanical and functional performance [16,17,18,19,20]. Additionally, factors such as fiber alignment, porosity, and interlayer bonding add complexity to understanding structural behavior as well as the polymers’ physical and chemical properties [21,22,23,24,25]. Conventional models often fail to account for these complexities; as such, there is a need for more accurate prediction methods.

Machine learning (ML) has become a useful tool for tackling the challenges of predicting the properties of AM polymers and polymer composites. Instead of relying on traditional methods that require a lot of laborious experimentation or complex simulations, ML uses data-based approaches to find patterns in large sets of data [26,27,28,29]. By processing both experimental and simulation data, ML models can predict material properties with good accuracy based on reliable data entries, helping researchers and manufacturers improve AM processes more effectively. From a mathematical perspective, the prediction involves looking for a suitable function linking the input and output. It is supported by the universal approximation theorem, which has been simplified [28].

In the past decade, because of increasing power in hardware, many investigators began to use a data-based approach to predict material properties. Traditional predictions of material properties often rely on first-principles or physics law-based approaches, such as density functional theory (DFT) [30], molecular dynamics (MD) [31], and finite element analysis (FEA) [32], which demand a deep understanding of the underlying physics and are time-consuming. Although these techniques have been useful, they struggle to capture the unpredictability of AM processes. On the other hand, data-driven models can identify complex connections between factors like printing conditions and material composition and the resulting properties [33,34]. This approach is more adaptable and offers a scalable solution.

ML plays a crucial role in several key domains of AM polymer composites. Process parameter optimization is one of them, where models based on ML are employed to calculate optimal levels of printing parameters for the desired mechanical and functional properties. By training on large datasets, ML models can propose optimal parameters that eliminate defects [35], enhance structural integrity [36], and provide superior material performance [37]. In addition, ML aids in defect detection through advanced image processing and real-time monitoring techniques. Through continuous monitoring of process parameters and material behavior, ML algorithms can detect early warning signs of inconsistency and make automatic adjustments to avoid defects. This predictive ability ensures higher reproducibility and reliability in AM polymer composite production, addressing one of the most enduring problems in the field.

This review analyzes the intersection of ML and AM polymers and polymer composites and how ML methods enhance polymer design and property prediction, process optimization, and efficiency in production. The review captures new advancements in ML algorithms such as supervised and unsupervised learning, hybrid ML models, and technical uses in modeling functional and mechanical properties. In addition, it delves into such challenges as data quality, model interpretability, and computational cost, while working on exploring various solutions and industrial applicability. Lastly, the review discusses some future directions, which include autonomous manufacturing using AI-based approaches and generative AI-based polymer design, and drives towards sustainability. This review employs a narrative synthesis approach. The relevant literature was identified from Scopus, Web of Science, and Google Scholar using keywords such as ‘machine learning’, ‘additive manufacturing’, ‘polymer composites’, and ‘process optimization’. Only peer-reviewed studies published between 2010 and 2025 were included. Papers lacking ML application, insufficient methodological detail, or not related to polymers or polymer composites were excluded.

2. Additive Manufacturing of Polymers and Polymer Composites

Advancements in additive manufacturing (AM) for polymers and polymer composites have expanded beyond basic thermoplastics to include high-performance thermoplastics, thermoset resins, gels, and elastomers enhanced with fiber, nanoparticle, and hybrid reinforcements [38,39,40,41,42]. These innovations have improved mechanical strength, thermal stability, and multifunctionality, making AM materials increasingly viable for aerospace, automotive, biomedical, and electronic applications.

Additive manufacturing (AM) of polymers and polymer composites offers several benefits, but faces challenges in optimizing the printing process [34,43,44], ensuring material property consistency [45], and predicting the final part performance [46]. These limitations affect the final properties of the printed parts.

2.1. AM Techniques for Polymers and Polymer Composites

The production of polymer and polymer composite parts involves various AM techniques with distinct processes, material compatibility, and characteristics. These methods include Fused Deposition Modeling (FDM) [47], Stereolithography (SLA) [47], Digital Light Processing (DLP) [48,49], Selective Laser Sintering (SLS) [50], and Material Jetting [51]. New AM techniques are also being developed to improve both material performance and processing efficiency.

2.1.1. Fused Deposition Modeling (FDM)

Due to its cost-effectiveness and ease of use, the FDM technique is the most commonly employed method in polymer and polymer composite production. It involves extruding a heated thermoplastic filament in layers onto a platform, where each layer bonds before solidifying, often while still semi-liquid, enabling the creation of complex shapes. To enhance their properties, composite materials containing reinforcing fillers such as carbon and glass fibers, as well as nanoparticles, are introduced into the thermoplastic filament before extrusion.

Amin et al. [52] reported that the FDM 3D printing process has the inherent limitation of creating rough surfaces due to visible layer lines. Notably, FDM exhibits poor adhesion and weak mechanical properties [53,54,55,56]. The weak interlayer bonding and inconsistent mechanical properties in FDM printing arise from its layer deposition process. Factors like nozzle temperature, layer thickness, printing speed, and print orientation significantly affect the mechanical performance of the polymer composites printed through this process [57,58,59,60,61,62,63,64,65].

2.1.2. Stereolithography (SLA) and Digital Light Processing (DLP)

SLA and DLP are photopolymerization methods that shape detailed polymer parts by curing liquid resin with ultraviolet (UV) light. While SLA relies on a laser to solidify resin layer by layer, DLP speeds up the process by using a digital projector to cure a full layer at once [66,67]. Also, the hardware setup for DLP offers a higher lateral printing resolution [68].

Table 1. Comparison of SLA and DLP 3D printing technologies.

Feature	SLA	DLP	Refs.
Light Source	Laser beam traces each layer	Digital projector cures entire layer at once	[66]
Print Speed	Slower (layer by layer)	Faster (whole layer cured simultaneously)	[49]
Resolution	High	High	[69,70]
Build Volume	Larger	Smaller	[71]
Surface Finish	Smooth surfaces	Smooth surfaces	[72,73,74]
Cost	High	High	[75]
Postprocessing	Required	Required	[76,77,78]
Mechanical Properties	Good	Better	[79,80,81]

compares SLA and DLP 3D printing technologies. These methods are known for creating detailed, high-resolution components with smooth surfaces. Nevertheless, the availability of suitable resins is limited, and cured resins also tend to be brittle. The strength and toughness of the printed parts can be enhanced by adding nanoparticles [66].

2.1.3. Binder Jetting

Binder jetting (BJT) is a low-temperature additive manufacturing technique that produces parts by applying a liquid binder onto a powdered bed in successive layers. Unlike fusion-based approaches such as SLS or FDM, it eliminates the need for heat during printing, making it more suitable for thermally sensitive materials [82]. The printing process involves spreading a thin powder layer across a build platform. A binder is then selectively applied through an inkjet-style print head to fuse the powder particles, forming a solid layer. As the platform lowers incrementally, the cycle is repeated, and new layers are added to form the part. Once printing is complete, excess powder is removed, and the resulting “green body” is strengthened through additional steps such as curing or sintering to achieve its final properties [83,84]. Water-soluble, biodegradable, and recyclable polymers [84] and polymer powders [85] have been investigated.

A key benefit of this binder jetting approach is that the powder bed naturally provides structural support during fabrication [85]. Despite these advantages, key challenges, such as uneven binder application, inconsistent powder flowability, and deformation during curing, must be addressed to maintain part quality [86].

2.1.4. Selective Laser Sintering (SLS)

Selective Laser Sintering (SLS) is an advanced additive manufacturing technology that uses a laser to fuse powdered material, layer by layer, into solid parts based on 3D CAD models. As a powder-based approach in AM, SLS supports the integration of reinforcing fillers directly into polymer blends, making it useful for producing composites. Materials often used include polyamides combined with carbon or ceramic elements to enhance strength and performance [87,88,89].

According to Häfele et al. [90] its ability to form complex shapes without extra support is a major advantage. This makes it good for the bioprinting of scaffolds with the advancement of biocompatible materials [91]. However, concerns remain about powder density, laser energy use, recyclability, and postprocessing requirements [50,92,93].

2.1.5. Material Jetting

Material Jetting is an AM process that works like an inkjet printer, but instead of ink, it deposits tiny droplets of liquid photopolymer, wax, or other build materials onto a build platform [94]. The droplets are cured by UV light, layer by layer, to form a 3D object. This method provides accurate printing with the ability to use multiple materials, making it suitable for functional polymer composite fabrication. While its high accuracy and multi-material use are advantageous, limitations in available materials restrict its broader use in structural composites [95].

2.2. Polymer Composite Materials in AM

A polymer matrix forms the base structure in composites produced via additive manufacturing. It holds reinforcing elements and significantly affects the final material’s strength, thermal behavior, flexibility, durability, and overall performance while determining its suitability for various AM methods [96,97,98,99,100]. In polymer composite AM, the matrices fall into three main groups: thermoplastics, thermosets, and elastomers. Each type presents unique benefits and drawbacks, depending on its processing behavior and application. Table 2 summarizes the polymers used in additive manufacturing. Figure 1a represents a schematic for the molecular structures of thermoplastics, elastomers, and thermosets, highlighting the distinct network configurations that contribute to their mechanical and thermal properties.

Currently, thermoplastics are the dominant polymer matrices in AM. Their ability to repeatedly soften with heating and solidify upon cooling, especially in techniques like FDM and SLS, enables reshaping and recycling. Commonly used thermoplastics in AM are acrylonitrile butadiene styrene (ABS), polylactic acid (PLA), polyether ether ketone (PEEK), polycarbonate (PC), and nylon-based polymers such as PA11 and PA12 [101,102,103]. Thermoplastics are valued for their ease of fabrication and ability to integrate with fiber or nanoparticle reinforcements. However, shrinkage, warping, and anisotropic properties can affect their structural performance [89]. Figure 1b illustrates a representative of a configuration of a polymer matrix composite, showing a randomly distributed particle phase embedded in the matrix.

Thermosets, on the other hand, tend to offer better stiffness, better dimensional precision, and smoother surfaces than thermoplastics. Common choices of thermosetting resins include epoxy, acrylate, and polyurethane resins. However, their brittleness, limited ability to be reprocessed, and challenges like short shelf life or the need for post-curing make them less ideal in applications where impact resistance or manufacturing flexibility is critical [104,105,106,107].

Elastomers are flexible, elastic polymers capable of withstanding significant deformation without permanent damage. Their adaptable nature makes them valuable in AM, particularly for producing parts used in soft robotics, medical devices, wearables, and impact-absorbing designs [108]. According to Nethani et al. [109], thermoplastic- and silicone-based elastomers are useful in 3D printing, where methods like Material Jetting and modified FDM systems enable the creation of flexible components with intricate shapes. However, challenges such as managing ink viscosity, minimizing shrinkage, and achieving strong interlayer adhesion remain central to optimizing these processes.

Table 2. Summary of polymer matrix categories in AM and their key properties.

Property	Thermoplastics	Thermosets	Elastomers	References
Thermal behavior	Reversibly soften and harden with heat	Irreversibly cured via crosslinking	Remain elastic across a wide temperature range	[42,90,109]
Reprocessability	Can be reheated and reshaped; recyclable	Cannot be reshaped once cured	Limited and material-dependent	[90,104,109]
Mechanical properties	Moderate strength; prone to warping and shrinkage	High stiffness and precision, but brittle	Stretchable and impact-absorbing	[42,90,109]
Surface quality	May need postprocessing to remove layer artifacts	Naturally smooth finish with high dimensional accuracy	Finish varies by technique; can be rough without tuning	[90,107,109]
Common AM techniques	FDM, SLS	SLA, DLP	Material Jetting, modified FDM	[101,102,103]
Main limitations	Warping, anisotropy, shrinkage	Brittleness, short shelf life, requires post-curing	Viscosity control challenges; weak interlayer bonding	[90,104,109]

Table 2. Summary of polymer matrix categories in AM and their key properties.

Property	Thermoplastics	Thermosets	Elastomers	References
Thermal behavior	Reversibly soften and harden with heat	Irreversibly cured via crosslinking	Remain elastic across a wide temperature range	[42,90,109]
Reprocessability	Can be reheated and reshaped; recyclable	Cannot be reshaped once cured	Limited and material-dependent	[90,104,109]
Mechanical properties	Moderate strength; prone to warping and shrinkage	High stiffness and precision, but brittle	Stretchable and impact-absorbing	[42,90,109]
Surface quality	May need postprocessing to remove layer artifacts	Naturally smooth finish with high dimensional accuracy	Finish varies by technique; can be rough without tuning	[90,107,109]
Common AM techniques	FDM, SLS	SLA, DLP	Material Jetting, modified FDM	[101,102,103]
Main limitations	Warping, anisotropy, shrinkage	Brittleness, short shelf life, requires post-curing	Viscosity control challenges; weak interlayer bonding	[90,104,109]

2.2.1. Reinforcing Materials in Polymer Composite AM

The integration of reinforcing agents in polymer matrices significantly improves the performance of composites used in AM. This allows for targeted adjustments to mechanical, thermal, electrical, and other functional properties, making the materials suitable for a wide range of applications [110,111]. AM incorporates reinforcement strategies that generally fall into three categories: fibrous materials, nanomaterials, and hybrid fillers [111]. Each contributes distinct benefits to the material’s structural integrity and functionality.

As described by Armstrong et al. [112], recent advancements in fiber encapsulation AM and modified FDM systems allow for the direct integration of continuous fibers into thermoplastic matrices. This development enhances the strength-to-weight ratio of composites, making them ideal for use in industries such as aerospace, automotives, and defense. With polymeric nanoparticles (Figure 1c), nanostructures tailored for biomedical and multifunctional uses can be manufactured. Integrating small amounts of nanomaterials into polymers can significantly improve their performance. Additives like carbon nanotubes (CNTs), graphene, fullerene, graphite, and metal or metal oxide nanoparticles help boost properties such as mechanical properties, thermal conductivity, electrical behavior, electromagnetic interference shielding, and flame resistance, and enable self-sensing features [113]. A schematic of CFRP is shown in Figure 1d.

SEM micrographs confirm the strong interfacial bonding and dispersion of CNTs in the polymer matrix, which is critical for enhancing composite mechanical properties [114]. These micrographs in Figure 2 reveal limited agglomeration and good particle distribution, which support efficient load transfer and heat conduction across the material.

Hybrid composites can outperform single-filler systems in energy absorption, thermal resistance, and bonding between layers. Prasanthi et al. [114] revealed that adding CNTs and graphene platelet nanoparticles improved the mechanical properties compared to a pure-carbon fiber-reinforced plastic composite. Hybrid reinforcement can also create functionally graded structures [115].

On the other hand, incorporating reinforcements into additive manufacturing poses several key difficulties. Non-uniform dispersion of fillers, clogging in extrusion nozzles, and maintaining print resolution and interlayer adhesion remain significant concerns. Adjusting the rheological properties of the material is often necessary to optimize feedstock properties, particularly when dealing with viscous inks or high filler loading. Furthermore, the bond between the polymer matrix and filler must be carefully engineered to prevent debonding or delamination under cyclic or high-stress loading. To enhance this interfacial adhesion, reinforcement surfaces are often modified, and coupling agents are used to ensure better compatibility between the materials [116].

Figure 1. (a) Structure of thermoplastic, elastomer, and thermoset, (b) configuration of a polymer matrix composite, (c) schematic representation of polymeric nanoparticles, including nano-capsules and nanospheres, with core–shell structures (reproduced from [117]), (d) typical structure of CFRP (reproduced from [118]).

Figure 1. (a) Structure of thermoplastic, elastomer, and thermoset, (b) configuration of a polymer matrix composite, (c) schematic representation of polymeric nanoparticles, including nano-capsules and nanospheres, with core–shell structures (reproduced from [117]), (d) typical structure of CFRP (reproduced from [118]).

2.2.2. Functional Fillers in Polymer Composite AM

Functional fillers are being used increasingly, not just for mechanical property enhancement but to introduce specialized features like electrical conductivity, thermal regulation, magnetic behavior, optical effects, or self-healing abilities. Integrating these materials into AM expands the potential for producing smart systems, multifunctional structures, and components engineered for specific operational environments.

One popular application of functional fillers is in the development of electrically conductive polymer composites. Materials like carbon nanotubes (CNTs), graphene, carbon black, silver nanoparticles, and metal-coated fibers are added to thermoplastics or thermosets to support uses such as sensors, flexible circuits, electromagnetic interference (EMI) shielding, and electrostatic-discharge protection [119,120,121,122,123].

Figure 3a shows the microstructure of 3D-printed carbon black-filled PLA (CB-PLA). It exhibits the formation of continuous conductive paths once a percolation threshold is achieved. This conductive network, formed by carbon black particles bridging across PLA filaments, enables electron mobility through the material. The effectiveness of AM in integrating such fillers is further demonstrated in the surface morphology observed through scanning electron microscopy (SEM). Figure 3b (left) compares the printed surfaces of neat PLA and CB-PLA, highlighting the textural changes due to filler incorporation. The high-magnification SEM images on the right confirm that the filler maintains its dispersion before and after the printing process.

Achieving this conductivity hinges on reaching the percolation threshold, which is influenced by factors such as filler aspect ratio, dispersion quality, and concentration. Yan et al. [119] have demonstrated low percolation thresholds for CNT/polymer composites, with 0.23 vol% and 0.18 vol% reported for CNT/PLA and CNT/HDPE, respectively. Works of Sit et al. [124] proved that functionalized multi-walled carbon nanotubes (FMWCNTs) in ethylene methyl acrylate (EMA) polymer have shown a superior EMI shielding effect and mechanical properties compared to non-functionalized MWCNTs, with 10 wt% FMWCNT loading achieving 25.1 dB of EMI shielding effect.

Thermally conductive fillers serve a vital role in enhancing thermal performance in polymer components. Fillers like graphene, boron nitride, or silicon carbide are chosen for their high intrinsic thermal conductivity, enhancing heat dissipation in applications where regulating temperature, such as in battery casings or heat sinks, is critical [125,126,127,128]. These fillers improve heat transfer in polymers, which helps extend the lifespan and performance of parts made by AM. Moser et al. [129] noted that because filler distribution and orientation during printing affect the material’s thermal conductivity, careful control of the print parameters is essential to address the anisotropy introduced by the layer-wise process. Min et al. [130] also suggested filler orientation can be controlled through thermophoresis, which induces particle rotation in a temperature gradient. However, the anisotropic nature of 3D-printed composites can be advantageous, particularly for vertical heat dissipation from a source, as stated by Schleifer and Regev [126]. Likewise, Melchert et al. [131] acoustically patterned fillers during 3D printing to align and compact particles, enabling highly efficient, anisotropic thermal pathways with up to 300% improved conductivity over unpatterned composites. Lopes et al. [123] studied topographical maps of PA12 surfaces embedded with different concentrations of MWCNTs (0.5 wt%, 1.75 wt%, and 3.00 wt%), showing increased surface irregularity with higher filler content, Figure 3c.

Figure 2. SEM images of hybrid carbon fiber- and carbon nanotube-reinforced epoxy composite (reproduced from [114]) (a) CNT aggregates around carbon fiber. (b) Pullout of carbon fibers from the matrix. (c) Clean carbon fiber surface without pullout. (d) CNT dispersion and network formation at higher magnification.

Figure 2. SEM images of hybrid carbon fiber- and carbon nanotube-reinforced epoxy composite (reproduced from [114]) (a) CNT aggregates around carbon fiber. (b) Pullout of carbon fibers from the matrix. (c) Clean carbon fiber surface without pullout. (d) CNT dispersion and network formation at higher magnification.

Recent discoveries highlight efforts to create self-healing composites in additive manufacturing applications. These materials often integrate microcapsules, vascular networks, or shape-memory components within a polymer matrix [108,132,133]. Upon the occurrence of damage, such as cracking, the repair mechanism is triggered, either by the release or thermal activation of healing agents to restore the material’s mechanical properties [134]. Integrating reversible dynamic bonds, such as Diels–Alder-based compounds and covalent adaptable networks (CANs) into AM workflows has enabled the creation of materials that can respond to external conditions, like temperature, light, or mechanical force, for damage healing [135]. Gyabaah et al. [136] 3D-printed shape-memory vitrimer (SMV)-regolith composite, which exhibited a good thermomechanical property with a compressive and tensile strength of 139.16 MPa and 13.99 MPa, respectively, and a good shape-memory effect. They suggested it is promising material for lunar base applications.

Figure 3. (a) Scheme of the microstructural composition of 3D-printed CB-PLA showing conductive paths formed by carbon black particles (reproduced from [137]) and (b) SEM images showing surface morphology comparison between printed PLA and conductive CB-PLA. Insets: high-magnification SEM of CB-PLA before and after 3D printing (reproduced from [137]). (c) Surface topography maps of PA12 samples containing 0.50 wt% (top), 1.75 wt% (middle), and 3.00 wt% (bottom) MWCNTs, illustrating increasing surface irregularity with higher filler concentrations (reproduced from [123], with permission).

Figure 3. (a) Scheme of the microstructural composition of 3D-printed CB-PLA showing conductive paths formed by carbon black particles (reproduced from [137]) and (b) SEM images showing surface morphology comparison between printed PLA and conductive CB-PLA. Insets: high-magnification SEM of CB-PLA before and after 3D printing (reproduced from [137]). (c) Surface topography maps of PA12 samples containing 0.50 wt% (top), 1.75 wt% (middle), and 3.00 wt% (bottom) MWCNTs, illustrating increasing surface irregularity with higher filler concentrations (reproduced from [123], with permission).

Despite their benefits, functional fillers present certain challenges during their incorporation into AM processes. High filler loading, especially during material preparation and printing, can greatly affect the flow behavior of inks or filaments, leading to the clogging of nozzles, uneven extrusion, or internal voids [138,139]. As such, uniform dispersion becomes an important factor in ensuring consistent performance. This often requires techniques like surface modification or the use of dispersing agents to prevent agglomeration, especially when working with nanoscale fillers [140]. Interfacial adhesion between the filler and the polymer matrix is also a critical factor. To promote effective stress transfer and reduce the risk of debonding in the final part, the interface must be optimized for both mechanical and functional integration. Surface treatments are often employed to achieve this.

2.2.3. Factors Affecting Material Properties in AM

The performance of AM composites depends on multiple factors, including material formulation [96,141], processing conditions [11], microstructure development [1], and post-production steps [120]. These factors collectively shape the thermal, mechanical, and functional behavior of the end product.

Process parameters are one of the most influential categories that contribute to material performance during the AM process. Process conditions like printing temperature, layer height, print speed, infill density, build orientation, and cooling rate directly affect the materials’ interlayer adhesion, porosity, and internal stresses [142,143,144,145]. As shown in Figure 4a, Moser et al. [129] illustrates a 3D-printed heat sink structure designed to assess thermal performance under varying processing configurations and reinforcement orientations. The geometry facilitates directional heat transfer and reveals how anisotropic material behavior arises from the layer-by-layer deposition process. The thermal response of these structures is mapped in Figure 4b, which presents infrared thermography data showing surface temperature distributions across PA12 specimens reinforced with MWCNTs, under different orientations (horizontal vs. vertical) and printing configurations (with/without top cover).

The properties of both the polymer matrix and the reinforcements used also significantly affect the final properties of AM parts. The viscosity, thermal conductivity, and the matrix’s flow behavior affect the material flow during deposition, and the rate of solidification, all of which determine the quality of the printed part [129,146]. The shape, type, and placement of reinforcements, such as fibers or nanoparticles, directly influence the efficiency of filler loads and the strength of the material [114].

Another factor of great concern is the microstructural features that arise during the printing process, such as pore formation, fiber alignment, and layer waviness. Due to the directional nature of material deposition in the AM process, anisotropic structures are introduced into the material which leads to differences in performance depending on direction. Polyzos et al. [147] have shown that the contour of carbon fibers affects the elastic response of composites. Rimašauskas et al. [148] proved that printing parameters like layer height and line width impact void volume (18.5–27.5%) and fiber content (13.4–19.1%) in printed composites, with layer height having the most significant effect on tensile properties. Polyzos et al. [147] also identified that thermal stresses during printing can lead to warpage and layer deformation, as demonstrated by their observations of distorted Technomelt PA 6910 box walls. In the box wall, glass fibers aligned along the printing direction, highlighting the influence of deposition direction and residual stress on dimensional stability. Fiber alignment, void content, and distribution affect both mechanical and thermal efficiency [149,150]. Tools like scanning electron microscopy (SEM) [151], micro-computed tomography (micro-CT) [152], and thermal imaging [96] are commonly used to analyze these internal structures and improve production methods. Other factors like postprocessing, such as post-curing, can affect the properties of the printed parts [153,154,155,156]. Another influencing factor is the environmental conditions during printing, including humidity, temperature, and oxygen levels, which can significantly affect material performance [157,158,159,160,161]. Works of Ramian et al. [162] have shown typical warpage and cracks in printed parts as a result of thermal deformation (Figure 5). To ensure material consistency and efficient performance, it is necessary to maintain controlled printing environments, typically using enclosed heated chambers or inert gas atmospheres.

2.2.4. Industrial Applications of Polymer AM Composites

One use of AM composites is in the aerospace industry for producing lightweight, high-strength components such as fairings, ducts, and brackets that improve fuel efficiency and payload capacity [163]. Often reinforced with carbon or glass fibers, thermoplastics like PEEK are employed due to its high-temperature stability. Khan and Riccio [164] argued that lattice structures, produced through AM, offer excellent strength-to-weight ratios and are increasingly utilized in aerospace applications.

The use of AM composites is being leveraged in the automotive sector for the rapid production of durable custom tools, as well as end-use parts such as dashboard panels, air intake systems, and structural reinforcements [165]. Functional fillers further extend their use in managing heat and shielding electronics in battery systems [125,126,127,128].

In the biomedical field, AM is enabling the creation of implants, prosthetics, and devices tailored to individual patients. Using materials like biocompatible thermoplastics and resorbable polymers, manufacturers can enhance mechanical performance and support tissue repair by incorporating fillers or reinforcing fibers. This approach has been applied in areas such as cranial reconstruction, spinal support, dental care, and targeted drug delivery, with added potential for integrating features like antimicrobial coatings or embedded sensors [166,167,168]. One example of this is bioprinting, where cells are isolated, proliferated, and printed into functional scaffolds for use in tissue repair, disease modeling, or drug screening [169].

Works of Lam et al. [170] as shown in Figure 6a–d, compares four major bioprinting techniques: inkjet bioprinting, which uses thermal or piezoelectric actuation to deposit cell droplets; laser-assisted bioprinting, which relies on laser pulses to propel bioink from a donor layer; extrusion bioprinting, which continuously dispenses bioink using pneumatic or mechanical pressure; and Stereolithography, which employs light to selectively crosslink photosensitive bioinks. Each method offers trade-offs in resolution, speed, and cell viability, indicating the versatility of AM technologies in biomedical applications, as described by Mirzaali et al. [171]. Figure 6e–g illustrates the bioprinting process, which involves three main stages: preprinting (cell culture preparation), bioprinting (bioink formulation and 3D printing of tissues), and postprinting (host–transplant integration and functional evaluation). It visually represents the progression from cell cultivation to in vivo tissue testing.

The consumer products and electronics industries also produce custom enclosures, wearables, smart textiles, and aesthetically refined components with embedded electronic pathways, soft-touch surfaces, and enhanced durability [172,173,174]. Ruckdashel et al. [175] have shown that the components of e-textiles, including base fabrics, interconnects, sensors, actuators, and power storage, can be integrated at multiple scales through AM, from fibers to coatings, without requiring major changes in production systems.

3. ML for Material Property Prediction in Additive Manufacturing

AM material property prediction used to rely on empirical and theoretical models that demanded extensive experimental data and detailed simulations. However, these methods often fail to deliver precise, quick predictions due to the complexity of AM processes and the complex behavior of materials. This limitation has driven the search for more effective approaches [176].

3.1. Overview of Data-Driven vs. Physics-Based Modeling

Traditionally, property prediction models were used to forecast the mechanical and thermal behavior of polymer composites. Although these methods are grounded in core principles, they demand significant computational resources and are influenced by assumptions made at the input stage [177,178,179]. They may also struggle to account for variations and defects arising in AM.

In contrast, data-driven ML models provide an alternative approach by learning patterns from existing datasets without relying on explicit physical equations. These models leverage experimental, simulation-based, and real-time sensor data to identify complex relationships between processing parameters, material composition, and performance characteristics. Unlike traditional methods, ML can uncover hidden correlations that are not easily captured through conventional modeling, as stated by Udu et al. [180]. This makes ML highly effective for predictive analytics in AM polymer composites.

During the AM process, several types of defects can be created. To link between processing conditions, material choices, and part performance, ML has been used [176]. An example is shown in the experimental results by Udu et al. [180], which highlight the effect of fabrication temperature on carbon fiber-reinforced polyamide (CF-PA) composites. Increasing the fabrication temperature led to reduced porosity and improved mechanical properties. According to Wang et al. [181], these data-driven approaches help establish process–structure-property relationships, supporting both offline optimization and online feedback control.

Moreover, Parsazadeh et al. [182] stressed on the hybrid approach that integrates data-driven methods with underlying physical principles to enhance model reliability and interpretability. They can dynamically adjust print settings in response to observed deviations, reducing waste and ensuring greater consistency across builds. Furthermore, ML accelerates the discovery and development of new polymer and polymer composite formulations by forecasting material performance and suitability using data on structure, thermomechanical properties, and printability indices [176,183]. This approach reduces the need for repeated testing, making the development cycle faster and more efficient.

3.2. Data Sources for ML Models

The reliability of ML predictions depends largely on the variety and quality of the training data. Major challenges with ML are data scarcity and lower quality [184]. To address these issues, three categories of data are typically employed: experimental data, simulation-based data, and real-time sensor data [185]. These categories are summarized in Figure 7, which outlines the diverse experimental and simulation techniques that serve as critical data inputs for training ML models in additive manufacturing and materials science.

According to Loisel et al. [186], experimental data remains central due to its critical role in capturing real-world material behavior. However, Karande et al. [187] argued that generating high-quality experimental data requires significant investment and careful consideration of data collection strategies. Experimental data, according to Zhu et al. [185], grapples with issues of the five Vs—namely, volume, velocity, variety, veracity, and value—and three Ms—namely, multicomponent, multiscale, and multistage challenges. Figure 8 shows an example of how data quantity and heterogeneity affect ML performance. Here, deep learning models demonstrate improved accuracy with more data and greater diversity, compared to traditional random forest methods [187].

Experimental data are derived from various characterization techniques, including mechanical testing, thermal analysis, and microscopy [180]. Li et al. [188] reported that measurements from tensile, compression, and impact testing help determine the strength and toughness of printed composites. Thermal methods like differential scanning calorimetry (DSC) and thermogravimetric analysis (TGA) are often used to evaluate thermal stability, decomposition temperature, and changes in phase [189,190,191]. Chakraborty et al. [192] stated that techniques such as scanning electron microscopy (SEM) and atomic force microscopy (AFM) provide microstructural data such as porosity, surface morphology, and fiber arrangement, which affect the material’s overall performance.

To complement empirical tests, computational simulations also serve as a source of data for training ML models. Finite element analysis (FEA) is used to simulate how materials respond to stress and strain, helping to create reliable training datasets [193,194]. On a smaller scale, molecular dynamics simulations provide insight into nanoscale behaviors like segmental motion, thermal transport, and how fillers interact with matrices [195]. Integrating ML with such physics-based simulations [196] enhances prediction accuracy across different materials and process conditions [192]. However, challenges remain, including data requirements, chemical transferability limitations, and speed–accuracy trade-offs in simulations [195].

A third important source of data is real-time sensor data generated through in situ monitoring. With improvements in in situ monitoring, dynamic feedback from the printing process can now be used to train models and adjust parameters as printing occurs [197]. Optical methods, including high-speed cameras and laser scanners, provide real-time assessment of layer alignment and surface anomalies. Thermal imaging, particularly infrared sensing, monitors temperature gradients to detect defects and capture spatiotemporal temperature distributions across the printed layers [198]. This is useful for spotting residual stress and heat-affected areas. In the works of Chen et al. [197], acoustic, thermal, and optical signals were collected, filtered, and processed to extract thermal gradients, melt pool characteristics, and surface geometry, which were then used as inputs for ML models to predict defects, microstructural features, and mechanical performance. This integration of diverse data improves accuracy, consistency, and quality monitoring in additive manufacturing.

3.3. ML Approaches Used in Property Prediction

ML models can be grouped into supervised learning, unsupervised learning, semi-supervised, self-supervised, and reinforcement learning [199]. These models, their suitable data types, and the commonly used implementation tools in AM workflows are visualized in Figure 9. These models are applied to various data types, including numerical, categorical, time-series, and text data [200]. Amiri and Soleimani [201] argued that their different advantages depend on the nature of the dataset, the complexity of the relationships being modeled, and the intended application. Model training typically involves dividing the dataset into training, testing, and validation sets, with cross-validation used to assess performance [202]. ML algorithms are implemented with Python-based libraries like Scikit-learn, TensorFlow, and PyTorch; or commercial platforms like MATLAB, and Minitab, integrated with AM simulation tools [203,204]. For example, Kuehne et al. [205] used Scikit-learn to implement ML models for predicting melt pool dimensions and part density in laser powder bed fusion.

3.3.1. Supervised Learning Models

Supervised learning models remain a widely used method for predicting material properties in polymers and polymer composites manufactured via AM. These models are trained on labeled datasets, where known input parameters connect with measured output properties [206,207,208]. Regression-based algorithms are one of the simplest supervised models. They include linear regression (LR) and support vector regression (SVR), and are used to predict continuously valued properties. LR assumes linear dependencies and is simple and interpretable, while SVR can handle nonlinear relationships using kernel functions [209,210]. These models offer insights into how individual parameters influence material behavior [207]. Nasrin et al. [211] used active learning to predict the tensile strength of extruded polymer composite. Approaches like random forest, gradient boosting, and XGBoost are often preferred over basic regression models when addressing complex, nonlinear relationships in data [212,213,214,215]. Jayasudha et al. [216] in their comparative analysis, concluded that in 3D printing applications, XGBoost regression proves most effective for estimating tensile strength, followed by gradient boosting, AdaBoost, random forest, and linear regression. By combining multiple decision trees and aggregating predictions, these methods can effectively identify subtle interactions in mixed datasets that include both numerical and categorical variables, such as print speed, temperature, and filler material.

A detailed ML pipeline for supervised regression in AM polymer composites is shown in Figure 10, illustrating the stages of data preprocessing, hyperparameter tuning, and model evaluation across multiple algorithms. The process begins with feature scaling and encoding of eight printing-related variables, followed by model training using randomized hyperparameter tuning and repeated k-fold cross-validation. The regression algorithms, including Bayesian Ridge Regression (BAY), CatBoost Regressor (CAT), K-Nearest Neighbors Regression (KNN), Lasso Regression (LAS), Ridge Regression (RDG), Support Vector Regression (SVR), and Random Forest Regression (RFR), are then evaluated using both accuracy and computational performance metrics to identify the best-performing predictive model.

Similarly, Challapalli and Li [210] after exploring several models settled on Gaussian Process Regression (GPR) to optimize lattice unit cell structures for AM, resulting in up to 57% higher buckling strength and 35% improvement in flexural strength over the conventional octet structure in sandwich composites. In a related study, similar ML models were used to mimic natural structures for reinforcement. The authors designed 3D-printable biomimetic rods, achieving a 150% improvement in buckling resistance compared to naturally inspired shapes [215].

Artificial neural networks (ANNs) can perform supervised learning by handling complex, high-dimensional data and uncovering nonlinear relationships. Their layered structure, inspired by biological neurons, allows them to detect fine connections between a material’s internal structure and its mechanical behavior. Another advantage is that ANNs do not assume a fixed function form; they learn a flexible, data-driven mapping between inputs and outputs, and adjust their structure through training to minimize prediction error. For polymer composites made through additive manufacturing, ANNs have been applied to estimate properties such as bonding strength between layers, toughness, and conductivity, using data such as print design and image-based details of porosity and fiber alignment [207,217,218]. Table 3 summarizes supervised ML models for predicting AM polymer composite properties.

3.3.2. Unsupervised Learning

Unsupervised learning reveals hidden patterns and relationships in data without labeled outputs [219]. This is especially useful for large datasets involving process or material characteristics where exact property values are not always available. Among the techniques used, clustering is common for grouping data based on shared features. Clustering algorithms such as K-means clustering and hierarchical clustering have been used to categorize polymer composites based on print parameters, reinforcement content, or thermal behavior [219,220,221]. These methods offer insight into the design of optimal materials by identifying trends and distinguishing between high-performing and underperforming formulations.

Table 3. Summary of supervised ML models for predicting AM polymer composite properties.

Model Type	Common Algorithms	Applications in AM	Key Strengths	References
Basic Regression	Linear Regression (LR), Ridge, SVR	Predicting tensile strength, modulus, thermal conductivity	Simple, interpretable; SVR handles nonlinearity with kernel functions	[199,200,201,202,222,223]
Ensemble Methods	Random Forest, Gradient Boosting, XGBoost, AdaBoost	Estimating tensile strength, capturing interactions in mixed feature sets (e.g., print speed, filler type)	High accuracy; robust to noise; captures complex feature interactions	[203,204,205,206,215]
Neural Networks	Artificial Neural Networks (ANNs)	Predicting bonding strength, toughness, conductivity; uses print data, porosity, fiber alignment	Excellent with high-dimensional, image-rich, or complex data; learns nonlinearities	[207,208,209,223]

Table 3. Summary of supervised ML models for predicting AM polymer composite properties.

Model Type	Common Algorithms	Applications in AM	Key Strengths	References
Basic Regression	Linear Regression (LR), Ridge, SVR	Predicting tensile strength, modulus, thermal conductivity	Simple, interpretable; SVR handles nonlinearity with kernel functions	[199,200,201,202,222,223]
Ensemble Methods	Random Forest, Gradient Boosting, XGBoost, AdaBoost	Estimating tensile strength, capturing interactions in mixed feature sets (e.g., print speed, filler type)	High accuracy; robust to noise; captures complex feature interactions	[203,204,205,206,215]
Neural Networks	Artificial Neural Networks (ANNs)	Predicting bonding strength, toughness, conductivity; uses print data, porosity, fiber alignment	Excellent with high-dimensional, image-rich, or complex data; learns nonlinearities	[207,208,209,223]

Feature extraction for establishing microstructure property correlations [224] is one application of unsupervised learning in polymer composite AM. The works of Alsenan et al. [225] in Figure 11 provides a conceptual overview of feature selection and feature extraction strategies, which are crucial steps in unsupervised learning workflows for polymer composite characterization in additive manufacturing. Through this process, researchers can better understand the complex relationships governing AM behavior to guide future data collection, model training, and process optimization. As data complexity continues to grow in AM research, unsupervised learning will be paramount in exploratory analysis and knowledge discovery.

3.3.3. Semi-Supervised Learning

Semi-supervised learning (semi-SL) bridges the gap between supervised and unsupervised learning by leveraging a small amount of labeled data along with a larger pool of unlabeled data to improve model performance [226]. This approach is valuable where the acquisition of labeled data, such as experimentally measured mechanical properties, can be costly and time-consuming, while unlabeled process or material data is often abundant. Semi-SL is also useful when dealing with high-dimensional data generated during in situ monitoring, such as thermal imaging or acoustic signals, where only a few samples are linked with failure modes or performance metrics [227,228,229]. Models can be initially trained on a small subset of fully characterized samples and then refined using unannotated data such as raw sensor outputs, process logs, or microstructural images [230,231].

Algorithms like semi-supervised support vector machines, self-training, and co-training techniques [232,233,234] allow these models to iteratively label and learn from the unlabeled data, thereby improving generalization and robustness. Figure 12 shows the flowchart of the Co-ANN algorithm. Liang et al. [235] developed a Co-training-style semi-supervised ANN (Co-ANN) model, which demonstrates significantly better agreement with experimental thermal conductivity values for BN-filled polymer composites, as compared to traditional theoretical models. The Co-ANN shows superior accuracy compared to Maxwell, Bruggeman, and effective medium models [235]. This highlights the capability of semi-SL approaches to leverage limited labeled data alongside abundant unlabeled data, leading to more accurate material property predictions. Figure 13 shows the pseudocode of the Co-ANN model.

3.3.4. Self-Supervised Learning

Self-supervised learning (SSL) is an emerging paradigm in ML that constructs supervisory signals directly from the data itself, eliminating the need for manually labeled datasets [236,237,238,239]. Unlike semi-supervised learning, which still relies on a small portion of labeled data, self-supervised approaches learn data representations by solving pretext tasks that require no external annotation. Recent research by Seneviratne et al. [240] has demonstrated the effectiveness of SSL for malware detection, with one study achieving 97% classification accuracy on a binary malware/benign dataset, outperforming existing baselines.

In AM, SSL can be applied to extract meaningful features from unlabeled data sources, which can then be fine-tuned for downstream tasks such as defect classification, mechanical property prediction, or anomaly detection. In image-based datasets, contrastive learning methods like SimCLR or autoencoder-based models can be trained to learn robust embeddings of microstructural images, identifying variations in fiber alignment, porosity distribution, or interlayer bonding [241,242,243]. These representations can significantly improve performance on supervised tasks even when labeled examples are scarce. As illustrated by Wei and Chen [244] in Figure 14, a self-supervised inference model with an interaction-based material network (IMN) can accurately infer stress–strain responses across previously unseen composite microstructures, demonstrating its robustness in generalizing across diverse configurations and loading conditions. The predicted curves align with the response of the direct numerical simulation (DNS), validating the model.

Moreover, self-supervised pretraining provides a foundational model that can be transferred across different AM systems or composite materials with minimal retraining [245,246]. This learning transfer capability reduces the need to build models from scratch for every new setup, thus accelerating deployment and reducing computational and experimental costs.

3.3.5. Reinforcement Learning

In reinforcement learning (RL) an agent learns to make optimal decisions by interacting with its environment and receiving rewards or penalties based on its actions. RL is designed to learn policies that maximize long-term rewards through trial-and-error [247,248]. This makes it well-suited for sequential decision-making tasks inherent in the AM of polymer composites, such as real-time process control and adaptive manufacturing [249]. Chung et al. [250] discussed RL-based methods that have been developed for online defect mitigation during printing. This model utilizes both offline and online knowledge to reduce required training samples for closed-loop control of print parameters in polymer composite AM.

Additionally, deep reinforcement learning (DRL) [251], which combines deep neural networks with RL, has been applied to discover novel material architectures and reinforcement patterns that achieve desired mechanical properties with minimal material usage. Another emerging application involves multi-agent reinforcement learning (MARL) [252], where multiple agents collaboratively optimize different aspects of the AM process such as temperature control, deposition strategy, and defect mitigation in real time. These approaches enable greater adaptability and resilience in AM systems, especially in uncertain or changing production environments.

3.3.6. Hybrid AI Models Combining ML with Physics-Based Approaches

Hybrid AI models combine ML techniques with traditional physics-based simulations to enhance the predictive accuracy and interpretability of computational tools. Hybrid models address the limitations of standard ML approaches, which are their reliance on large datasets and their tendency to make predictions that may conflict with fundamental scientific laws [253]. One such approach, known as Physics-Informed ML (PIML), embeds equations of heat transfer, mechanical stress, thermodynamics, or fluid flow into the ML models [254,255]. The framework in Figure 15 illustrates how the PIML pipeline embeds physical constraints throughout the learning process via data, model structure, and optimization [254]. PIML ensures that predictions are consistent, even when extrapolating beyond the bounds of the training data.

In addition to PIML, hybrid models often integrate ML with simulation tools such as finite element analysis (FEA) [256] and molecular dynamics (MD) simulation [257]. Combining ML models with simulation methods supports analysis of complex, multiscale behaviors that are difficult to capture through experiments or simulations alone. Continued development of such hybrid AI models may help link predictions with physical insights in designing composites.

Das et al. [258] presented a generative deep learning approach with a conditional variational autoencoder (CVAE) for the synthesis of thermoset shape-memory polymers (TSMPs) with targeted thermal and mechanical properties for additive manufacturing. Their framework, trained on a graph-encoded description of polymer structure and conditioned on functional group constraints (e.g., epoxy, amine), synthesized 22 novel TSMPs with low glass transition temperature and high recovery stress. A CNN regression model also predicted these properties from the learned latent space. Their work demonstrates the ability of hybrid generative–predictive models to enable domain-specific material discovery for 3D printing applications.

Overall, the performance of different ML approaches in polymer additive manufacturing depends strongly on data size, feature quality, and target-property complexity. Tree-based models such as random forest and XGBoost deliver robust and accurate predictions for small to medium datasets and provide useful feature-importance insights, but they struggle with high-dimensional data and extrapolation. Support Vector Regression achieves reliable accuracy on limited data but is sensitive to kernel selection and parameter tuning. Neural network-based models (ANNs, DNNs, CNNs, and GNNs) excel at capturing complex nonlinear relationships and microstructural information from images, yet they require large datasets, extensive training, and risk overfitting. Unsupervised and clustering algorithms help uncover hidden data structures but lack quantitative predictive power. Reinforcement-learning frameworks show promise for adaptive process control and defect mitigation but remain underexplored due to the scarcity of real-time reward data. These comparative insights highlight that model selection should balance accuracy, interpretability, computational demand, and data availability for each specific polymer AM application.

4. Advances in ML-Driven Property Prediction

The application of ML has brought notable progress to AM by improving material property predictions and optimization. Unlike the conventional trial-and-error approach, ML offers data-driven insights that speed up material development, refine process control, and boost polymer composite performance. Studies by Deshmankar et al. [259], Jyeniskhan et al. [260], Wang et al. [181], and other researchers show that ML models effectively predict mechanical, thermal, and electrical properties, as well as accurately model process–property relationships. Furthermore, digital twins powered by ML enable real-time monitoring, defect detection, and predictive maintenance, enhancing the reliability and efficiency of AM [261]. While generative models like GANs are commonly used, Challapalli et al. [262] have shown that inverse ML frameworks based on regression and statistical correlation can achieve faster convergence and better results.

This section explores recent developments in property prediction by ML, covering mechanical, thermal, and electrical properties, as well as process–property relationship modeling. It also discusses the role of AI-driven digital twins in AM quality control, highlighting their impact on defect detection and process optimization.

4.1. Mechanical Property Prediction

When evaluating AM polymer composites, mechanical properties like tensile strength, toughness, impact resistance, and elastic modulus are key factors to consider. However, the layering process of AM leads to anisotropic behavior, making it difficult to maintain consistent mechanical properties.

ML methods have been used to establish the correlation between printing parameters like temperature and layer thickness to the mechanical behavior of polymer composites in AM. Various supervised learning algorithms, including decision trees, random forests, and support vector regression, have been used to correlate printing parameters with mechanical behavior [263]. Prada Parra et al. [206] found that K-Nearest Neighbors and CatBoost models provided the most accurate predictions for elastic modulus and tensile strength in fiber-reinforced composites. Moreover, Thomas et al. [264] proposed an ML-accelerated methodology to simultaneously determine elastic properties and fiber orientation state in short fiber-reinforced polymers, potentially reducing the need for extensive material characterization. A comparative performance study by Deb et al. [263] in Figure 16, involving ensemble models, Extremely Randomized Trees (ERTs), and traditional ML models (KNN, SVM), further demonstrates the superiority of ensemble approaches. Using an 80–20 train–test data split, the models were trained to predict both surface roughness and tensile strength in FDM-printed PLA specimens. The random forest model achieved the best results for surface roughness prediction, with a root mean squared error (RMSE) of 0.408, a mean absolute error (MAE) of 0.31, and a mean absolute percentage error (MAPE) of 9.28%. For tensile strength prediction, the Extremely Randomized Tree model outperformed the others, yielding an RMSE of 1.03, MAE of 0.82, and a low MAPE of 2.20%. These metrics show the enhanced capability of ensemble learning models to accurately forecast mechanical performance by capturing intricate dependencies within the dataset, whilst also reducing experimental burden and accelerating material optimization workflows.

Also, deep learning techniques such as artificial neural networks (ANNs), deep neural networks (DNNs) [265] and convolutional neural networks (CNNs) [266] have also been trained on microstructure images to detect features such as void distribution and interlayer adhesion. Malley et al. [176] developed a neural network model to forecast mechanical behavior of AM particulate composites, showing excellent agreement with experimental data. These tools facilitate the optimization of printing conditions.

4.2. Thermal and Electrical Property Prediction

Factors such as matrix interactions, filler content, and how the material is printed influence the thermal behavior of polymer composites. ML can be used to estimate properties like thermal conductivity using data from tools like thermogravimetric analysis (TGA) and differential scanning calorimetry (DSC). Yan et al. [267] were able to predict the glass transition temperature (Tg) of polymers using hybrid deep learning models with an error margin of less than 3% when their model was validated experimentally.

Various ML algorithms, including random forest, gradient boosting, and neural networks, have been successfully applied to estimate thermal conductivity in composites with different fillers like graphene and carbon nanotubes [268,269,270]. A study by Liu et al. [270] proves the use of Shapley Additive Explanations (SHAP) analysis to interpret ML predictions of electrical conductivity in CNT/polymer nanocomposites. They used SHAP dependence plots to visualize how key input features such as volume fraction, energy barrier height, and CNT conductivity interact with each other to affect the model output. These insights help in identifying the most influential parameters and optimizing nanocomposite design.

ML algorithms are widely used to model how conductive polymer composites behave in electronic systems, especially in terms of electrical conductivity. By training algorithms on lab data, they can now predict how parameters like filler dispersion, percolation threshold, and processing conditions affect performance [207]. To account for the effects of filler agglomeration and segregation, micromechanics-based models have been developed to predict electrical conductivity in these composites. The works of Yoon et al. [271] show ML techniques enabled high-throughput optimization of electrical conductivity of doped conjugated polymers using absorbance spectra data, improving measurement efficiency by 89%. ML has made it easier to quickly test and develop high-performance conductive materials for sensors, electromagnetic interference shielding, and flexible electronics [119,272].

4.3. Process–Property Relationship Prediction

Understanding how processing affects material properties is vital for improving additive manufacturing. Changes like porosity, interlayer bonding, and residual stress can alter the material’s properties; hence, modeling these factors helps to maintain consistent material quality. ML-based approaches have been used in this area to optimize processing conditions for enhanced composite performance.

Studies show that porosity and interlayer bonding are key factors in the mechanical performance of 3D-printed polymer composites [273,274]. Since pore formation and adhesion defects are often unpredictable, ML models have been trained on X-ray computed tomography (CT) scans and microscopy images to predict porosity patterns and bonding strength. With an accuracy of 80 to 99%, Udu et al. [180] developed ML models that can effectively relate porosity to other mechanical properties, with ensemble tree learners and K-NN regressors. Similarly, Mohammed et al. [275] employed artificial neural networks trained on X-ray CT images to predict density/porosity of AM parts with high accuracy (R² = 0.9981). Some approaches also include finite element simulations to assess stress build-up during printing, enabling the design of low residual stress AM processes. Factors influencing residual stress formation, such as scan patterns, preheating, energy density, part geometry, and substrate constraints, can be analyzed using thermomechanical FEM models [276].

The integration of ML in AM spans various stages, including preprocessing design, parameter optimization, anomaly detection, in situ monitoring, and postprocessing [277]. To boost material performance and minimize defects, ML methods like genetic algorithms, Bayesian optimization, and reinforcement learning methods are being used to optimize print settings such as layer thickness, extrusion speed, and cooling rates [276]. When paired with live monitoring, these tools help streamline the production of polymer composites in additive manufacturing, leading to less waste and high-quality materials. Notably, Teimouri et al. [265] used Taguchi-based design and DNN models to evaluate key geometry parameters before optimizing structure mass using an imperialist competitive algorithm.

4.4. ML-Based Digital Twins for AM Quality Control

Digital twin technology, powered by ML, has emerged as a game-changer in AM quality control. These virtual models mirror physical systems that update continuously in real time based on sensor input and predictive models. In AM, digital twins enable real-time monitoring, defect detection, and predictive maintenance, improving process reliability and efficiency.

Real-time monitoring tools leverage sensor data from thermal, acoustic, and optical imaging to detect printing defects during printing. By analyzing these signals, ML models can recognize problems such as layer separation, overheating, and print deviation. When connected to live feedback systems, these tools can adjust printing parameters dynamically, preventing defects before they compromise part quality.

Acoustic emission sensors can detect warpage by analyzing vibration changes during printing, with ML models trained on frequency and time domain features [222]. Optical and infrared cameras are used to monitor the printing process and build area, detecting surface and subsurface defects [198]. CNNs applied to images captured during printing can identify geometrical anomalies in infill patterns, enabling real-time defect detection [278]. Infrared thermography has been integrated into printers to collect layer-wise thermal data, allowing for the detection of defects by comparing spatiotemporal information between non-defective and defective parts [279]. This approach significantly reduces the need for postprocessing inspections, minimizing production costs and time.

Combining ML with sensor data allows digital twins to mimic AM processes and predict equipment performance. By identifying potential failures, predictive maintenance algorithms extend the lifespan of AM equipment and minimize unplanned downtime. According to Sicard et al. [280], by integrating ML algorithms, digital twins can improve data processing, modeling, and trend prediction for AM equipment. They also facilitate process optimization by simulating different printing scenarios [261] so that adjustments can be made without wasting resources on physical trials. Figure 17 shows a schematic representation of a digital twin framework in additive manufacturing. Real-time sensor data from the physical printing system is integrated with predictive models to create a dynamic virtual replica, enabling monitoring, defect detection, and process optimization.

4.5. Characterization Tools in ML-Enhanced Additive Manufacturing Research

Characterization tools are essential to advancing ML methods in AM of polymer composites. They supply microstructural and empirical data needed to train and refine models for material property prediction, process quality monitoring, and structural optimization. Material behavior from the molecular to the macroscopic level can be better understood using a combination of thermal, mechanical, and microscopic characterizations. These approaches provide the basis for generating reliable, data-driven insights. Mechanical tests like tensile, compression, flexural, and impact tests are essential for assessing properties such as strength and toughness. These measurements provide baseline data used to train supervised learning models to predict the performance attributes of materials based on printing parameters, reinforcement content, and microstructure.

Thermal analysis methods such as DSC, TGA, and dynamic mechanical analysis (DMA) [191,281] are often used to evaluate key polymer behaviors, including transition temperatures, decomposition behavior, and viscoelastic properties. This data serves as input for ML models that predict heat resistance, dimensional stability, and performance under thermal cycling. Microscopy techniques supply critical microstructural information for ML tasks like predicting material properties from images or identifying defects. SEM and TEM are often used to observe features such as filler dispersion, fiber alignment, and interfacial adhesion, while AFM helps analyze surface morphology and nanoscale interactions [282,283]. These imaging datasets are essential for training CNNs, which extract features from micrographs to predict failure mechanisms, detect defects, or correlate structure with macroscopic performance. X-ray computed tomography (XCT) further enables non-destructive, 3D characterization of porosity, internal geometry, and layer bonding [284,285]. The scan data can be segmented to build mesh models for finite element analysis or serve directly in ML-based simulations.

AM systems now integrate in situ and real-time monitoring tools, which provide constant feedback to support adaptive control and predictive maintenance. Techniques such as infrared thermography, acoustic emission monitoring, and optical coherence tomography help track features like thermal gradients, interlayer bonding, and surface integrity. These real-time measurements can be used to train ML models to identify deviations, trigger corrective actions, and update digital twins to match actual print conditions. When combined with post-process characterization, it helps track the entire lifecycle of additively manufactured parts, from material deposition to final performance.

5. Challenges in ML-Based Material Property Prediction

The broader use of ML is limited by challenges such as data availability and quality, model generalization, computational cost, interpretability of AI models, and standardized industrial frameworks. This section outlines key obstacles to implementation and explores the solutions to overcome these barriers for ML applications in this field.

5.1. Data Availability and Quality Issues

The availability of training datasets in AM polymer composites is limited due to several reasons. First, experimental data collection is time-consuming and resource-intensive [37,176,184,185]. Mechanical testing, thermal analysis, and microstructural characterization require expensive equipment and skilled labor, restricting the volume of available data. Additionally, inconsistencies in experimental procedures across different research groups and manufacturing environments lead to dataset heterogeneity [36,180,184,286], making it difficult to develop universally applicable ML models. Second, the absence of open-access datasets restricts progress in building generalized AI-driven solutions for AM [11,287]. Companies involved in AM research and production often consider material property data confidential, limiting data sharing and collaborative ML model development.

To mitigate these data limitations, researchers are exploring data augmentation techniques, such as synthetic dataset generation through computational simulations and generative AI [234,288,289]. Simulations like FEA, DFT, and MD provide valuable synthetic data that can supplement experimental findings. Additionally, techniques such as generative adversarial networks (GANs) [290] can create synthetic microstructural images, expanding training datasets for deep learning applications [291]. Recently, ML models have advanced to use less data. Challapalli et al. [262] designed a framework that achieved global optima with fewer iterations by using targeted structural features and correlation analysis, reducing reliance on large datasets. Despite having only 55 SMP samples, Yan et al. [292] supplemented their dataset using physics-based equations, demonstrating how hybrid models can overcome data scarcity.

5.2. Model Generalization and Transferability

Another challenge in ML-based material property prediction is the difficulty of applying trained models across different AM systems and material types [293,294]. They require frequent retraining, which reduces their effectiveness in broader applications. For instance, ML models trained using FDM may not perform well when applied to SLS or SLA due to differences in each technique. Additionally, variations in machine calibration [295], environmental conditions [157,296], and raw material quality [297] further limit model transferability.

Researchers are increasingly exploring advanced ML techniques that address the limitations associated with data scarcity, model transferability, and generalization in AM. One potential solution is transfer learning [298,299,300], in which pre-trained ML models developed with large, generic datasets are fine-tuned with smaller, task-specific datasets. This reduces the need for extensive retraining, speeds up the development cycle, and reduces computational costs when applying ML models to new AM systems. Teimouri and Li [301] used transfer learning to augment small datasets in shape-memory polymer (SMP) discovery, leading to accurate prediction and generation of polymers with Tg > 190 °C. Closely related to this is domain adaptation. Domain adaptation addresses the challenge of applying ML models to different but related datasets [302,303,304]. This approach strengthens the reliability of models when working with data from various additive manufacturing platforms, materials, or processing conditions, making them more adaptable across related tasks.

Another technique is meta-learning, often described as “learning to learn” [305]. This method trains ML algorithms across different tasks or datasets to understand how to adapt quickly to new challenges with limited data. It helps algorithms perform well across varying materials and processes. These advanced learning strategies are central to developing ML systems in AM that are both efficient and adaptable.

5.3. Computational Cost and Scalability

The intensive computational demands for deep learning limit their practical use in predicting material properties in AM. Deep learning models typically require extensive training time and access to powerful GPUs and cloud computing infrastructure [306,307]. Yu et al. [308] noted that for some applications, simpler models such as SVR or random forests may offer faster predictions, but sometimes at lower accuracy for some other applications. Recent studies have focused on reducing the high computational costs associated with applying ML to AM [287]. By using targeted optimization methods, researchers aim to streamline resource use without compromising prediction quality, supporting more practical use of ML in AM processes. One such notable technique is model compression, which is essential for managing the computational demands of ML in AM. Methods like pruning eliminate redundant connections in neural networks, while quantization lowers the precision of computations, both of which speed up inference times [309].

Another strategy is knowledge distillation, which enables lightweight models to mimic the behavior of larger ones, making them suitable for real-time monitoring and quality assurance in environments with limited computational capacity. Edge AI, for example, enables AM systems to perform on-site analysis without the need to transmit large datasets to centralized servers [310,311]. Sani et al. [312] found that Edge AI facilitates closed-loop feedback for monitoring, automation, and optimization of printing parameters. This approach allows for rapid, localized decisions based on real-time data [313]. By running lightweight models directly on embedded devices integrated into 3D printers, it becomes possible to make rapid, localized decisions. This reduces latency and supports autonomous or semi-autonomous manufacturing setups, which are especially valuable in remote or high-throughput environments.

Cloud-Based Training offers a practical solution for the initial development and refinement of ML models for AM. Cloud platforms and high-performance computing (HPC) resources enable scalable training workflows, allowing researchers to iterate faster and deploy models more effectively across multiple machines or manufacturing lines [314].

5.4. Explainability and Trust in ML Models

The interpretability of ML models remains a critical challenge in materials science and manufacturing, with many accurate models functioning as “black boxes” [315,316]. ML models may predict a certain tensile strength value, but without explainability tools, it is challenging to determine which factors (whether fiber alignment, porosity distribution, or printing temperature) contributed most significantly to that prediction. This limits trust in AI-driven recommendations in industries. However, Explainable Artificial Intelligence (XAI) [317] has emerged as a solution to address this lack of transparency and interpretability. XAI techniques aim to provide insights into the decision-making process of ML models, enhancing trust and scientific credibility.

One widely adopted approach is Shapley Additive Explanations (SHAP), which assigns importance scores to each feature based on cooperative game theory. SHAP helps quantify the contribution of individual parameters to predicted material properties like tensile strength or thermal conductivity. Cooper et al. [318] has applied SHAP to predict tensile strength in directed energy deposition, identifying key thermal features influencing material properties. Similarly, Kharate et al. [319] found that for 3D-printed biocomposites, SHAP showed infill density as the most influential factor for tensile and flexural strengths, while biochar content primarily affected impact strength. This method allows engineers and scientists to prioritize the most influential factors during process optimization and materials design.

Another technique is Local Interpretable Model-Agnostic Explanations (LIME), which builds interpretable surrogate models around individual predictions. LIME approximates how small perturbations in input data affect the outcome, offering a clearer view of the reasoning behind specific decisions made by complex ML models. However, Zafar and Khan [320] suggested that LIME’s random perturbation method can lead to instability and shifts in data distributions, which may undermine the reliability of its explanations. To address this, they proposed deterministic versions using clustering and nearest-neighbor sampling to improve consistency. Venkatsubramaniam and Baruah [321] also propose formal concept lattices to generate more structured, rule-based explanations, enhancing both interpretability and reproducibility.

Additionally, Partial Dependence Plots (PDPs) [322] are used to visualize the marginal effect of one or two features on the predicted outcome while holding other variables constant. However, PDPs have limitations in high-dimensional settings, leading to the development of extensions like TOTALVIS (a model package) for analyzing grouped covariate effects [323]. Other approaches include a conceptual framework for explorable and steerable partial dependence analysis, which allows for incremental refinement of results and user-guided exploration of feature combinations [322]. Inglis et al. [324] proposed new visualization techniques to jointly display variable importance and interaction effects, such as heatmap- and graph-based displays, as well as matrix-type layouts for single and bivariate PDP.

Furthermore, in deep learning applications, attention mechanisms [325] have emerged as a powerful XAI tool. These mechanisms allow models to focus on crucial input elements, improving tasks like machine translation and sentiment analysis [326]. When incorporated into convolutional neural networks or transformers, attention layers can highlight critical regions in input data, such as areas of porosity or fiber misalignment in microstructural images. This not only improves model performance but also enhances interpretability by revealing the visual or structural features driving predictions.

5.5. Industrial Integration and Standardization

For ML-based material property prediction to gain widespread adoption in industry, standardized data-sharing frameworks and ML-based manufacturing protocols must be established. Currently, the lack of standardized data formats, inconsistent measurement methodologies, and proprietary restrictions hinder the seamless integration of ML into industrial AM workflows [287,293].

A key challenge in industrial integration is bridging the gap between AI-driven predictions and real-world applications. Many ML models are trained on controlled laboratory datasets that may not fully reflect the variability encountered in industrial production environments. Factors such as equipment inconsistencies, environmental fluctuations, and heterogeneous materials input introduce uncertainties that current models may not fully account for, limiting their reliability in real-world deployment [327]. To facilitate industrial adoption, collaborative efforts are needed.

First, the development of open-source AM databases, drawing from both academic research and industrial case studies, can provide diverse, high-quality datasets necessary for robust model training and validation. Tang et al. [328] demonstrated the potential of transfer learning in AM modeling, recommending source domains with larger qualitative similarity and specific target-to-source data ratios. Similarly, Casukhela et al. [329] proposed a multimodal community-based AM database to address challenges in part qualification and material discovery.

Second, the establishment of AI-integrated manufacturing standards by international organizations such as ASTM International and ISO is critical. These standards should outline protocols for ML-based quality control, real-time defect detection, and process optimization, ensuring consistency and interoperability across AM platforms. The ISO/IEC 22989:2022(E) [330] standard on AI concepts and terminology is a significant step towards global AI regulation, acting as a para-regulatory framework for future developments [331].

Third, a shift toward enhanced AI–human collaboration is needed, where ML acts as a decision-support system rather than a fully autonomous operator. This approach empowers domain experts to interpret ML outputs, validate model predictions, and fine-tune process parameters. This synergy allows ML to process vast amounts of data while humans contribute creativity and emotional intelligence [332]. However, challenges persist in optimizing this collaboration, including understanding complementary conditions, assessing human mental models of ML, and determining effective design choices for human–AI interaction [333]. Inkpen et al. [334] stated that the success of human–AI teams depends on factors such as user expertise, algorithmic tuning, and users’ perception of ML performance. To maximize the benefits of human–AI collaboration, it is crucial to develop frameworks that emphasize complementary roles and address ethical concerns related to bias and transparency.

6. Future Directions and Emerging Trends

While current ML approaches have already demonstrated improvements in material property prediction, defect detection, and process optimization, future advancements will further integrate ML into autonomous manufacturing, generative material design, multiscale modeling, and sustainability initiatives.

This section explores the future directions and emerging trends in ML for AM, including self-learning manufacturing systems, generative ML for material design, multiscale property prediction, and ML’s role in sustainable manufacturing.

6.1. ML-Based Autonomous Manufacturing

One of the most promising trends in AM is the development of ML-driven autonomous manufacturing systems, where self-learning algorithms continuously monitor and adapt printing processes in real time. These systems aim to reduce human intervention, optimize production efficiency, and minimize defects by making data-driven decisions during fabrication. The transition from Intelligent AM to Autonomous AM involves a hierarchical framework with integrated layers, enabling machines to independently observe, analyze, and execute operations [335].

6.1.1. Self-Learning ML Models for Real-Time Process Adaptation

Traditional AM workflows rely on pre-programmed parameters that may not account for variations in environmental conditions, material inconsistencies, or machine wear. Self-learning ML models, powered by reinforcement learning and adaptive neural networks, can analyze real-time sensor data and dynamically adjust printing conditions to ensure optimal performance. These models leverage feedback loops to continuously refine their decision-making, improving print quality and reducing failure rates. An example is acoustic emission combined with ML that enables differentiation of process conditions and workpiece quality in laser additive manufacturing [336]. Closed-loop ML-augmented AM systems can monitor, automate, and optimize printing parameters, improving defect detection and prevention, especially in Fused Deposition Modeling printers [312].

Another example is that ML algorithms trained on thermal imaging and acoustic emission data can detect anomalies such as overheating, warping, or interlayer delamination during printing. Upon identifying these issues, the ML system can automatically adjust parameters such as extrusion speed, layer height, or cooling rates to correct deviations in real time. This level of adaptability will be crucial for achieving highly reliable and reproducible AM processes, particularly for industries such as aerospace and biomedical manufacturing.

6.1.2. ML-Enhanced Robotics for Automated AM Production

The integration of ML with robotics is another emerging trend that will drive the automation of AM production lines. ML-powered robotic arms equipped with computer vision and deep learning models can perform complex tasks such as multi-material deposition, in situ quality inspection, and automated part removal [337]. These smart robotic systems will enhance AM scalability, allowing for the high-throughput manufacturing of customized polymer composite parts.

Moreover, ML-based robotic platforms will facilitate hybrid manufacturing, where AM is combined with subtractive and postprocessing techniques. For instance, robotic arms can be programmed to seamlessly switch between 3D printing, milling, and surface finishing processes, ensuring that final components meet stringent quality standards.

6.2. Generative ML for Material Design

ML-driven material design is rapidly transforming how new polymers and polymer composites are developed within the domain of AM. Traditional trial-and-error experimentation is increasingly being replaced by computational approaches that enable the discovery of novel materials with customized properties. Deep generative models, such as variational autoencoders and generative adversarial networks (GANs) [338], have emerged as powerful tools for accelerating the inverse design of materials with targeted properties. These approaches, combined with high-throughput virtual screening and Bayesian optimization [339], enable rapid exploration of vast chemical spaces and optimization of material properties. These ML techniques reduce time and cost associated with materials development while promoting innovation in structure–property relationships.

GANs, a class of deep learning models originally developed for image synthesis, are now being repurposed for advanced material design and inverse material discovery. These deep learning models can generate, translate, and improve microstructural images, producing realistic and indistinguishable results from real samples [290]. A typical GAN architecture comprises two competing neural networks: a generator and a discriminator [340]. GAN is especially advantageous in exploring compositional and structural spaces that are either too complex or cost-prohibitive to investigate experimentally. These deep learning models capture key characteristics such as alignment of fibers, distribution of fillers, porosity gradients, and resin-rich pockets [290,341,342,343,344]. Works of Teimouri et al. [345] emphasized a hybrid inverse-design platform based on a deep neural network and conditional GANs to discover 3D-printable plate–lattice structures with 170% high specific recovery force (SRF). Using a unique fingerprinting approach and GAN, Challapalli et al. [342] discovered quite a few lattice structures and 3D-printed them.

While VAEs and GANs are fundamental generative modeling techniques, advancements are also brought about by transformer-based models, diffusion models, and multimodal models. Transformer models, embodied by the ChatGPT family, offer better sequence modeling and are being explored for Simplified Molecular Input Line Entry System (SMILES)-based polymer design, property prediction, and text-to-material design pipelines [346,347,348,349]. Diffusion models that iteratively refine noise into structured outputs have been competitive in 3D structure creation and molecular design tasks and provide tighter feature distribution control compared with GANs [350]. Muroga et al. [351] have shown a multimodal deep learning framework that can predict diverse properties of advanced polymer composite materials.

Figure 18 demonstrates the TransPolymer architecture and workflow, a transformer model designed for polymer property prediction [349]. Figure 18a shows polymer sequences made up of SMILES representations and molecular descriptors tokenized with chemical context to allow the model to understand complex molecular structures. Figure 18b shows the TransPolymer pipeline, which adopts a pretrain–finetune approach. Figure 18c depicts the core transformer–encoder architecture, pointing to the multi-head attention mechanism that captures fine-grained inter-token relationships. Figure 18d depicts the two-step learning procedure: pretraining includes masked language modeling for filling in the missing tokens, encouraging strong latent representations; finetuning includes the special token (‘<s>’) vector of the final hidden layer as input for subsequent property prediction. The visualization of attention, signified by line thickness and color, maps to the varying importance of token relationships throughout the input sequence. This framework is the new application of deep learning models to polymer informatics and generative design.

6.3. ML-Based Inverse Design Approaches

Inverse design methods, powered by ML, are revolutionizing the development of polymers and polymer composites by reversing the traditional material discovery process. Instead of starting with material inputs and predicting the resulting properties, inverse design begins by specifying the desired performance outcomes, such as mechanical strength, thermal stability, or electrical conductivity, and then uses advanced algorithms to determine the optimal material composition and process parameters to achieve those goals [352,353]. Different from the traditional viewpoint, the current inverse design does not directly explore the inverse function. On the contrary, most inverse models are composed of two parts: forward prediction and inverse mining [267]. In forward prediction, the model aims to obtain a forward prediction, which will be used to look for a better input by combining with generative models. One way to conduct the inverse design process is to use the forward design model to predict the properties by inputting randomly generated SMILES until the predicted properties achieve or exceed the design level.

Song et al. [354] exemplifies such an approach through a generative inverse-design framework originally developed for optimizing electrocatalyst surfaces for CO₂ reduction. It combines a pre-trained generative model, which creates atomic surface structures, with a graph neural network (GNN) that predicts adsorption energy, and a bird swarm algorithm (BSA) for property optimization. While this example focuses on catalytic surfaces, the underlying methodology, coupling structure generation, property prediction, and iterative optimization, is directly applicable to the design of polymers and polymer composites for additive manufacturing. In this context, the same principles can be harnessed to generate microstructures that enhance performance characteristics.

An ML-based inverse-design system might take a target specification like “Please design a thermoset shape-memory polymer with glass transition temperature around 150 °C and rubbery modulus about 100 MPa” and generate a series of candidate formulations involving varied polymer matrices [355]. These proposed designs can then be evaluated using physics-based simulations or targeted experimental validation, significantly accelerating the pace of innovation. By enabling goal-oriented material engineering, inverse design not only reduces development time and cost but also fosters the discovery of high-performance composites tailored for the specific demands of additive manufacturing processes.

6.4. ML for Multiscale Property Prediction

ML can integrate hierarchical modeling techniques, linking information from nano-, micro-, and macroscale datasets to provide a more holistic understanding of material behavior. This is accelerating the prediction of material properties across scales, from quantum chemistry to continuum modeling [356]. Deep learning models trained on MD simulations can predict polymer–filler interactions at the nanoscale, while CNNs analyzing SEM images can assess fiber alignment at the microscale [357,358,359]. Mianroodi et al. [357], illustrated in Figure 19, employed a convolutional neural network (CNN) to capture the size-dependent behavior of the bulk modulus in nanoporous materials. By leveraging molecular dynamics (MD) simulation data, the model effectively predicted how nanoscale pore geometry impacts material elasticity. These insights can then be fed into an FEA-driven ML model to predict macroscale mechanical performance [360], ensuring consistency across all length scales.

Hierarchical ML models can streamline the multiscale optimization of AM processes, ensuring that properties optimized at one scale do not negatively impact another [361]. For example, improving fiber alignment for enhanced strength at the microscale should not compromise thermal performance at the macroscale. ML-based models can balance these trade-offs efficiently, leading to multifunctional polymer composites with superior performance across all length scales.

6.5. ML in Sustainable and Smart Manufacturing

Sustainability is becoming a critical focus in the field of AM, prompting a shift toward environmentally conscious material development and production practices. Researchers are increasingly leveraging ML-based solutions to address challenges associated with eco-friendly polymer composites, waste management, and energy efficiency. These innovations aim to reduce the environmental footprint of AM technologies while maintaining or even enhancing material performance and production scalability.

ML is playing a transformative role in the design and optimization of bio-based and recyclable polymer composites, which provide a sustainable alternative to conventional petroleum-derived materials. ML models trained on databases of biopolymers and natural fillers can accurately predict key performance metrics such as tensile strength, thermal stability, and biodegradability. This allows for the strategic selection of renewable components like cellulose fibers, lignin nanoparticles, starch-based matrices, and tailoring material formulations for targeted applications [362,363]. In Figure 20a, Wilson et al. [364] employed the PolyID tool, a graph neural network, to successfully predict the glass transition temperatures of bio-based polymers. Their findings demonstrate how specific functional groups and monomer pairings influence thermal performance, offering potential PET alternatives. Similarly, Tao et al. [365] showed a comprehensive benchmark study of seventy-nine ML models for predicting polymer glass transition temperatures, highlighting the importance of proper model selection and feature representation.

Additionally, in the study by Hernández et al. [366] a neural network regression model was trained using a dataset generated via a Design-of-Experiments (DoE) approach to predict the mechanical behavior of agar-based biopolymer films. As shown in Figure 20b, the model used concentrations of agar, glycerin, and water as input features to predict Young’s modulus, ultimate tensile strength (UTS), and elongation at break. The contour plots reveal the influence of component ratios across low-, medium-, and high-concentration regimes, capturing complex nonlinear relationships and aiding in the formulation of biopolymers with targeted mechanical properties.

Moreover, AI-enhanced life-cycle assessment (LCA) models are being used to evaluate the environmental impact of various AM materials and processing methods. These models guide manufacturers in choosing low-carbon-footprint materials and energy-efficient print strategies, aligning product development with circular economy principles. A major challenge in AM for polymer composites is the accumulation of material waste stemming from support structure removal, failed prints, and postprocessing steps. ML offers powerful tools to tackle these issues and enable sustainable material usage and recycling. ML algorithms can predict the recyclability of composite materials by analyzing their chemical structure, degradation behavior, and thermal history, enabling the selection of recycling-compatible formulations [367,368]. ML can also optimize the reuse of polymer composites by monitoring process parameters and material consistency during reprinting, thus ensuring high-quality outputs in closed-loop manufacturing systems. According to Besigomwe [369], these technologies have demonstrated the ability to reduce raw material use by up to 12%, increase material recovery by 25%, and lower operational costs by 15%.

Additionally, ML-based energy management systems can optimize process scheduling, power consumption, and machine usage to reduce the energy intensity of AM operations. However, standardization in energy consumption measurement and reporting remains a challenge, hindering accurate comparisons between different AM systems. Future research should focus on developing standardized methodologies for energy efficiency assessment, exploring the integration of renewable energy sources, and investigating the complex interactions between process parameters to further optimize energy consumption in AM processes [370].

As AM evolves toward a greener and more sustainable manufacturing paradigm, the integration of ML in material selection, process optimization, and waste management will be essential. These ML-enhanced strategies not only improve the environmental sustainability of polymer and polymer composite production but also support regulatory compliance, cost reduction, and long-term innovation in AM. One direction to maintain sustainability is to use ML to discover recyclable and self-healable polymers, such as vitrimers, for additive manufacturing. Quite a few works have been conducted towards this emerging direction [371,372,373].

7. Conclusions

The integration of ML and AI into AM for polymers and polymer composites has significantly advanced the field by enabling precise material property prediction, process optimization, and real-time defect detection. This review explored the role of ML in AM, highlighting the various predictive modeling techniques used to forecast mechanical, thermal, and electrical properties, as well as the impact of ML on quality control and autonomous manufacturing. Despite these advancements, several challenges remain, including data availability, model generalization, computational costs, and the explainability of ML models. Addressing these challenges will be crucial for the widespread adoption of ML-based solutions in AM.

7.1. Key Findings

• Diverse AM Techniques and Material Systems: We detailed the major AM modalities, Fused Deposition Modeling, Stereolithography, Selective Laser Sintering, material and binder jetting, and emerging directed-energy and multi-material methods and their compatibility with thermoplastics, thermosets, elastomers, and a spectrum of fiber, nanoparticle, and functional fillers. Each technique offers unique advantages such as resolution, mechanical strength, and multifunctionality, but also introduces process-induced variability in layer adhesion, porosity, and anisotropy that complicates property prediction.
• ML Paradigms and Applications: ML approaches from classical regression and tree-based ensemble methods to deep neural networks, clustering, and reinforcement learning have been successfully deployed to predict mechanical, thermal, and electrical properties, optimize process parameters, detect defects in real time, and even steer generative and inverse design of novel composites. Self- and semi-supervised learning, as well as hybrid physics-informed models, have emerged to overcome data scarcity and embed fundamental physical laws into data-driven workflows.
• Explainability, Industrial Integration, and Sustainability: Explainable ML techniques, including SHAP, LIME, PDP, attention mechanisms, and formal concept lattices, are mitigating “black box” concerns, fostering trust in safety-critical applications. Yet industrial deployment still lags, impeded by proprietary data silos, a lack of standardized ML protocols, and the gap between lab-scale datasets and real-world variability. Simultaneously, ML is enabling more sustainable AM practices, the design of bio-based composites, self-healing polymers, ML-assisted life-cycle assessment, and closed-loop recycling strategies that reduce waste and energy consumption.

7.2. Future Outlook

Looking forward, the field must address both technical and collaborative barriers to fully realize ML-based AM:

• Federated and Privacy-Preserving Learning: Developing federated ML frameworks will allow multiple stakeholders to train shared models on confidential industrial datasets without exposing proprietary information, broadening the data landscape while safeguarding intellectual property (IP).
• Standardization and Open Data Ecosystems: Establishing AI-integrated standards through ASTM, ISO, and industry consortia, coupled with curated, open-source AM datasets, will enable robust benchmarking, reproducibility, and accelerated model validation across laboratories and production sites.
• Autonomous and Digital-Twin Manufacturing: The integration of ML-powered digital twins, closed-loop control, and multi-agent reinforcement learning will move AM toward self-optimizing “lights-out” factories, where real-time sensor fusion and AI decision-making assure quality and adaptability at scale.
• Next-Generation ML-Based Design: Advances in generative models (GANs, VAEs) and inverse-design algorithms promise to unlock unprecedented composite architectures and functionally graded materials, tailoring microstructure and composition to application-specific performance targets with minimal human intervention.
• Sustainability by Design: Continued growth in ML-enabled life-cycle assessment, eco-design of polymer matrices and fillers, and process optimization for energy efficiency will be essential to align AM with circular economy principles and global decarbonization goals.
• Keep Pace with AM: AM is evolving quickly. ML must keep pace with the new developments in AM. In polymer and polymer composite AM, new developments such as volumetric 3D printing [374] and polymer curing by various electromagnetic waves (for example, visible light) and mechanical waves (for example, ultrasound) [375] deserve attention.

By bringing together data openness, standardized protocols, advanced ML methodologies, and sustainability imperatives, the field is poised to move from exploratory research to a mature, reliable, and environmentally responsible stage of additive manufacturing for polymers and polymer composites.