Emergence of Artificial Intelligence in the Structure-Controllable Synthesis of Carbon Nanotubes

Abstract

Carbon nanotubes (CNTs) exhibit remarkable properties that have spurred extensive exploration across domains such as integrated circuits, aerospace, and energy storage. With application scenarios becoming increasingly specialized, the structure-controllable synthesis of CNTs faces escalating challenges. This review summarizes chemical vapor deposition (CVD) techniques for controlled CNT synthesis and examines the structural-control mechanisms during growth, emphasizing the critical factors influencing CNT diameter, electronic properties, and chirality. Conventional trial-and-error approaches have become inadequate in addressing the demands for precise structural manipulation during synthesis and complex variable optimization in scaled production. Recently, artificial intelligence (AI) has substantially advanced scientific research and technological innovation. In the concluding perspective, we highlight emerging paradigms that incorporate AI into CNT synthesis, where the synergy between data-driven experimentation and physics-informed constraints may enable the development of accurate and efficient digital twins of CNT growth systems. Such approaches offer promise for the inverse design of synthesis routes and deeper insight into structure-control mechanisms. We conclude by identifying promising directions for AI-enhanced CNT synthesis, including multiscale computational simulations, catalyst design, automated experimental platforms, and pilot-scale production, which may collectively advance the Frontier of precision nanomanufacturing.

1. Introduction

Carbon nanotubes (CNTs) are unidimensional nanomaterials, first discovered by Iijima in 1991. , They exhibit excellent electrical, mechanical, and thermal properties and play important roles in various application scenarios. − Notably, the carrier mobility of CNTs can reach up to 1 × 10⁵ cm² V^–1 s^–1. Field-effect transistors fabricated based on single-walled CNTs (SWCNTs) can operate with low energy consumption (because few carriers scatter during transport, suppressing short-channel effects , ), making them promising to replace silicon-based semiconductors and usher in an era of carbon-based chips. The theoretical Young’s modulus of perfect CNTs is 5.5 TPa. Despite the presence of defects in as-synthesized CNTs, the tensile strength of as-produced CNT fibers exceeds 116.8 GPa, substantially higher than those of the three major fibers, namely, carbon, aramid, and ultrahigh-molecular-weight polyethylene fibers. Furthermore, incorporating CNTs into composite materials enhances mechanical strength , and facilitates the development of carbon allotropes with excellent mechanical and interesting electronic properties. The thermal conductivity of SWCNTs exceeds 3500 W m^–1 K^–1 at room temperature, making them a promising material for fabricating high-efficiency thermal management units and addressing heat-dissipation challenges associated with high-power devices.

The structure of a material determines its properties. SWCNTs are formed by curling and closing a single graphene layer into a hollow tubular structure. By contrast, multiwalled CNTs (MWCNTs) comprise multiple SWCNTs with varying diameters arranged in a multilayered tubular configuration. The diameter of SWCNTs is controlled by a finite number of carbon atoms, with the smallest known diameter being 0.33 nm. When the diameter exceeds ∼6 nm, SWCNTs collapse into a ribbon-like graphene structure because of their inability to support their own weight. CNTs exhibit a wide range of lengths, from a few nanometers to several decimeters. Notably, the conventional preparation technique for CNTs does not entail graphene curling but rather the catalyst-assisted direct growth of tubular structures. CNT synthesis methods are introduced in Section , focusing on CVD. The CNT growth mechanism is discussed in detail in Section . The chirality and diameter of CNTs can be quantitatively described based on the curling direction and curling length of the graphene. Because of the hexagonal symmetry of the graphene layer, the range of values for the chiral angle (θ) is 0 ≤ θ < 60°, where θ = 0, 0 < θ < 30°, 30°, and 30° < θ < 60° correspond to zigzag, dextrorotatory, armchair, and levorotatory tubes, respectively. The rolling vector is uniquely determined by the chirality (n, m), resulting in variations in the electronic structures of CNTs. Typically, CNTs are metallic (m-CNTs) when n – m = 3q, where q = 0, 1, 2, 3···; otherwise, they are semiconducting (s-CNTs).

Over the past decade, several notable advancements have been made in controlled CNT synthesis. Through the design of catalysts and precise control of growth conditions, >99% pure SWCNTs have been produced. Ultralong tubes up to half a meter long have also been synthesized. In addition, s-CNTs can be produced with 99% purity by introducing an etchant. SWCNTs with a chirality enrichment abundance of over 90%, such as (6, 5), (12, 6), and (14, 4), can be synthesized via vapor–solid–solid (VSS) growth. − However, the reliable synthesis of CNTs with multiple superior properties, such as a high aspect ratio, high density, and a high proportion of semiconducting-type tubes, remains a significant challenge.

In recent years, a series of studies have emerged on the utilization of AI to facilitate CNT synthesis, − revealing the immense potential for applying AI to CNT growth. Ji et al. devised a high-throughput CNT growth scheme (Figure b) to obtain 1280 data points, which were employed to train a random forest regression (RFR) machine-learning model, achieving a coefficient of determination (R ²) value of 0.88, thereby demonstrating the considerable potential of machine-learning models in CNT synthesis. Lin et al. employed 585 and 16,000 sets of real and virtual experiments, respectively, to train a machine-learning model for optimizing CNT growth conditions, achieving a 48% improvement in CNT growth efficiency (Figure d). Li et al. developed an AI-based intelligent synthesis platform for nanocarbon materials, named Carbon Copilot (CARCO, Figure c). Through modular design, CARCO integrates language models, automated experiments, and data-driven machine learning (ML) to efficiently fabricate and optimize CNT horizontal arrays. CARCO identified a titanium–platinum bimetallic catalyst and fabricated arrays with specified densities, revealing the powerful capabilities of AI and automated systems in exploring complex material systems.

1.

Definition of the fine structure of CNTs and AI methods for CNTs. (a) Definition of diameter, wall number, electrical properties, chirality and defect structure of CNTs. Adapted with permission from ref . Copyright 2020 Elsevier. (b) Batch synthesis and characterization of CNTs using high-throughput synthesis technique for machine learning analysis. Adapted with permission from ref . Copyright 2021 Springer Nature. (c) Carbon Copilot (CARCO) with catalyst prediction workflow (left side) and controlled-density growth workflow (right side) demonstrate CARCO’s capabilities in innovation and precision manufacturing within carbon-based nanomaterials synthesis research. Adapted with permission from ref . Copyright 2025 Cell Press. (d) The trade-off relationship between tube length and crystallinity of CNT simulated by using 585 sets of real experiments and 16,000 sets of virtual experiments to train the machine learning model. Adapted with permission from ref . Copyright 2023 American Chemical Society.

In terms of predictive accuracy, although traditional (e.g., RFR) ML models achieve R ² values of 0.92 for predicting CNT array densities, they are outperformed by the large language model (LLM)-integrated carbon BERT which achieved 93% accuracy in catalyst screening. Regarding data dependency, traditional ML requires hundreds of experimental samples for reliable analysis, limiting its applicability to low-data regimes (e.g., catalyst design). By contrast, LLMs demand pretraining with >50,000 materials science publications and fine-tuning with 1000 experimental data sets to achieve high generalizability. For interpretability mechanisms, traditional ML leverages Shapley additive explanation (SHAP) values to quantify feature contributions. Conversely, LLM-integrated carbon BERT utilizes attention mechanisms to correlate synthesis descriptors (e.g., carbon flux and growth temperature) with growth outcomes. Concerning risk mitigation, LLM-generated synthesis recipes exhibit ∼44% unphysical prediction risk, necessitating human–AI collaboration for experimental validation to ensure controlled CNT growth.

Recent advances have also confirmed that AI is crucial for elucidating fundamental CNT growth mechanisms. Conventional molecular dynamics (MD) simulations face severe limitations in capturing CNT evolution during growth, such as defect-healing dynamics, due to prohibitive computational costs. Machine-learning force fields (MLFFs) overcome this barrier by enabling microsecond-scale simulations with quantum-mechanical accuracy. For example, the DeepCNT-22 MLFF revealed that under optimized conditions (at 1300 K and a growth rate of <0.5 ns^–1), interfacial defects undergo spontaneous self-repair prior to incorporation into CNT walls, thereby explaining the defect-free growth of (6, 5) single-walled CNTs on Fe₅₅ catalysts. Furthermore, active learning frameworks (e.g., the carbon-growth-on-metal machine-learning potential (CGM-MLP)) integrate Gaussian approximation potentials with timestamped force-biased Monte Carlo (tfMC) sampling, facilitating direct simulation of graphene nucleation dynamics on Cu(111) substrates and identification of critical edge-passivation pathways, a process previously inaccessible to conventional density functional theory (DFT)/MD simulations owing to time scale constraints.

In previous works on controllable CNT synthesis, researchers have applied AI methods, including open-source machine-learning models, such as RFR and support vector machine (SVM). By inputting hundreds of original experimental data points, they have been able to obtain analysis results tailored to specific experimental tasks. Through data-driven methods, several tasks can be achieved; for example, specific experimental parameters can be optimized within a particular experimental system and experimental outcomes can be predicted, and synthesis parameters can be recommended (including experimental parameters and catalyst components) by leveraging existing data. This approach has introduced a research paradigm to traditional CNT synthesis. Data-driven parameter optimization has markedly improved the efficiency and accuracy of parameter adjustment. Although recommended experimental parameters offer convenient inspiration, they have several limitations: (1) ML models are typically confined to interpreting experimental results from a single experimental device; therefore, data generated by different experimental devices cannot be effectively harnessed by these models. (2) Data analysis results lack explanations based on scientific principles. (3) Although the generation of experimental data requires considerable time and materials, the data volume has not increased by an order of magnitude and is approaching a bottleneck.

Nonetheless, compared to traditional methods, the use of AI in the controllable synthesis of CNTs offers unparalleled advantages in terms of data quality and reverse prediction capabilities. It offers strong prospects for enabling breakthroughs in controllable CNT synthesis.

In Sections , 3.2, and 3.3, synthesis methods for controlling CNT diameter, electrical properties, and chirality index are respectively reviewed. The final section provides a detailed analysis of the current challenges in the CVD synthesis of CNTs, highlighting that AI will introduce novel methodologies to the field of CNT synthesis; for example, digital-technology-driven high-dimensional fitting models will facilitate parameter optimization and mechanism explanation, aiding fine-structure regulation during SWCNT growth.

2. Carbon Nanotube Synthesis Method and Growth Mechanism

2.1. Carbon Nanotube Synthesis Method

Structure-controllable synthesis is the future of CNT material research and development. The principal techniques employed for synthesizing CNTs encompass arc discharge, laser ablation, and CVD. In arc discharge, a direct current is applied to graphite electrodes containing catalyst precursors in an inert gas environment. Arc generation evaporates carbon, subsequently depositing CNTs on the cathode surface. The high temperature generated by the arc (above 2000 °C) produces highly crystalline, perfectly structured CNTs. Although arc discharge rapidly produces CNTs, it may introduce additional metal impurities, which can subsequently affect the overall purity of the final product. Furthermore, arc discharge is not a continuous production process and requires considerable energy input. Additionally, the system’s instability may generate non-CNT byproducts, such as amorphous carbon, reducing the system purity and impeding further application of arc-produced CNTs. In laser ablation, a high-energy laser beam is focused on a carbon source, evaporating and condensing it to CNTs. Because of its stable and uniform high-temperature environment, laser ablation enables the effective control of the CNT diameter. Furthermore, the number of walls can be precisely controlled in the resulting CNTs, with the overall SWCNT purity typically reaching over 95%. Similar to arc discharge, laser ablation also provides a high temperature, producing highly pure, perfectly structured CNTs. Nevertheless, because of the restricted heating area, the yield obtained using laser ablation is inferior to that obtained using arc discharge.

CVD was initially developed for growing semiconductor thin films. In the 1990s, CVD was initially applied for synthesizing CNTs and has since become one of the most widely used CNT synthesis methods. In CVD, tubular furnaces are primarily used as reaction vessels. Carbon sources (hydrocarbons, carbon monoxide, ethanol, etc.) are introduced into the furnace, where they undergo catalytic decomposition under high-temperature or plasma-assisted conditions, forming CNTs. CVD is widely used because of its simple operating conditions and flexible parameter-adjustment range, particularly for large-scale production. CVD systems are mainly classified into three main categories, namely, thermal, plasma-enhanced, and floating-catalyst CVD, each possessing distinctive advantages. In recent years, notable progress has been made in CVD-based CNT synthesis, which is discussed in the Section .

Thermal CVD relies on high temperatures to facilitate the decomposition of carbon precursors, typically yielding high-quality SWCNTs. It is convenient to operate, making it suitable for cultivating CNTs with defined structural characteristics. The CVD growth process is illustrated in Figure a. The carbon source is catalytically decomposed in a high-temperature environment, and the resulting carbon fragments subsequently grow into CNTs through catalytic action. Furthermore, the design of catalysts and optimization of growth conditions (including the carbon source, catalyst, and reaction temperature) enable precise control of the CNT structure to meet specific application requirements.

2.

Three common CVD systems for growing CNTs. (a) Growth of CNTs in the thermal CVD system. The base growth mode (b) and the tip growth mode (c) of CNTs. (d) Schematic diagram of the plasma-enhanced (PECVD) system. (e) Schematic diagram of the plasma formed between the electrodes. Equipment diagram (f) and schematic diagram (g) − of the floating-catalyst CVD (FCCVD). (h) Comparison of thermal CVD, PECVD and FCCVD in different aspects of CNTs preparation. (f) Adapted with permission from ref . Copyright 2021 John Wiley and Sons. (g) Adapted with permission from ref . Copyright 2022 American Chemical Society.

Plasma-enhanced CVD (PECVD), based on plasma-assisted excitation reactions, enables carbon source decomposition at lower temperatures (below 800 °C), providing an effective solution for synthesizing CNTs under low-temperature conditions. Furthermore, PECVD offers substantial benefits in forming vertically aligned CNTs. For instance, Maekawa et al. used PECVD and ligand-modified hollow nanoparticle catalysts to synthesize ultradense CNT forests with uniform inner diameters. Although PECVD circumvents the random agglomeration of catalytic nanoparticles at elevated temperatures, the necessity for a vacuum environment and plasma generation makes PECVD an expensive technique, thereby limiting its scalability for large-scale applications.

In floating-catalyst CVD (FCCVD), the catalyst is in the form of floating nanoparticles. The continuous feeding of catalyst precursors and carbon sources enables the continuous growth of CNTs, producing a high growth rate. To date, numerous studies have utilized FCCVD’s continuous growth to prepare macroscopic structures. For instance, Wang et al. employed layer-by-layer condensation to continuously prepare CNT films while effectively controlling the film size and thickness. Zhang et al. developed a chlorine (Cl)/water (H₂O)-assisted FCCVD method to adjust intertube CNT interactions, mechanically strengthening macroscopic fibers. Zhu et al. utilized biomass-derived tannic acid to continuously generate highly graphitized CNT yarns, addressing the challenge of producing biomass-derived CNT yarns.

Clearly, these CVD methods have distinctly characteristic underlying reaction principles, operational temperatures, and resulting products. The selection of an appropriate CVD method is very important to prepare CNTs tailored for specific application scenarios. As this review is concerned with the controlled preparation of CNTs, the subsequent discussion is primarily based on thermal CVD methods.

2.2. Carbon Nanotube Growth Mechanism

The vapor–liquid–solid (VLS) mechanism is widely used to explain CNT growth. The VLS mechanism was first proposed by Wagner in 1964 to explain the growth of silicon nanowires. Subsequently, Baker et al. extended the VLS mechanism to grow carbon nanowires. According to the model’s description, CNT growth mainly comprises the following continuous stages: chemical reaction between the catalyst and carbon source, migration of carbon atoms within the catalyst, and CNT formation.

In the reaction step, the carbon source is absorbed on the catalyst surface, which atoms facilitate the cleavage of the carbon source and formation of carbon radicals. In the subsequent migration step, the carbon radicals are diffused. As the carbon source dissolves, it reaches supersaturation. During CNT formation, the carbon source precipitates on the catalyst surface, forming a “carbon cap” matching the catalyst particle size and possessing a specific chirality. Carbon cap formation represents CNT nucleation. Each carbon cap is associated with a particular chirality. As the carbon source precipitates, additional carbon atoms follow the carbon cap template, growing CNTs possessing a specific chirality beneath the cap.

During CNT growth, carbon transitions from a gas to a metastable liquid carbide, which subsequently crystallizes to form CNTs. The VLS growth mechanism explains microscopic CNT growth catalyzed by low-melting-point transition metals − (Fe, Co, Ni, etc.). The apparent activation energy for CNT growth is almost identical to that for carbon dissolution in metals, , such as iron and cobalt, providing substantial support for the VLS mechanism. However, the dissolution of the carbon source in the second step is disputed. The initial VLS mechanism proposed that molten carbon migrated within the catalyst particles. However, subsequent studies have revealed that carbon can migrate on the catalyst surface instead of dissolving into the catalyst particles. In 2004, Helveg et al. used time-resolved in situ high-resolution transmission electron microscopy to characterize MWCNT growth on solid Ni catalysts. Because of the limitations of the VLS mechanism in accounting for these observations, the VSS mechanism was proposed as an alternative. The VSS mechanism postulates that catalyst particles retain their solidity during the entire growth process and that the carbon source cannot dissolve into the interior of the catalyst particles and instead migrates via surface diffusion and carbide formation. Subsequent research has shown that CNTs grown using high-melting-point metals via the VSS mechanism possess finer structures and particularly improved chiralities. This is discussed in Section 3.3. Environmental transmission electron microscopy (ETEM) has emerged as a powerful tool for exploring the microscopic CNT growth processes. He and Li used ETEM to observe in situ CNT growth over liquid and solid catalysts, respectively, , further corroborating the VLS–VSS model (Figure a), which provides an effective theoretical framework for CNT growth, substantially enhancing our understanding of CNT growth and providing a robust theoretical basis for using CVD to prepare CNTs.

3.

Phenomena in CNTs growth and the corresponding mechanism. (a) Liquid catalysts grow CNTs follow the VLS mechanism, resulting in a variety of chiral CNTs. While solid catalysts grow CNTs follow the VSS mechanism and grow symmetry matching chiral CNTs. Adapted with permission from ref . Copyright 2021 American Chemical Society. (b) H₂ selective etching small tubular s-CNTs for the enrichment of m-CNTs. Adapted with permission from ref . Copyright 2014 American Chemical Society. (c) Multiple modes of CNTs growth termination in Inappropriate chemical environment. Adapted with permission from ref . Copyright 2017 American Chemical Society. (d) Different contact modes between CNTs and liquid catalyst lead to different growth phenomena. Adapted with permission from ref . Copyright 2018 Royal Society of Chemistry. (e) The selectivity in ultralong CNTs growth. Adapted with permission from ref . Copyright 2022 John Wiley and Sons.

As CNT research has progressed, several interesting phenomena have been discovered, for which specific CNT growth mechanisms have been proposed to elucidate the underlying causes. During the in situ transmission electron microscopy characterization of CNT growth, tangential and vertical growth contact modes were observed. The varying solubility degrees of the carbon sources in the catalyst produced different wetting abilities of the liquid catalyst on the CNTs. Wetting alters the catalyst morphology and forms different contact modes between the catalyst and CNTs (Figure b). For low-solubility carbon sources (e.g., CH₄), the catalyst wets the CNTs, forming a tangential growth mode. In contrast, for high-solubility carbon sources (e.g., CO), the catalyst maintains a spherical cluster morphology, forming a vertical growth mode with CNTs. H₂ is a mild etchant for selectively etching small-diameter s-SWCNTs, which are more chemically active, synthesizing an 88% yield of large-diameter m-SWCNTs. The length distribution of ultralong CNTs is correlated with the fine CNT structure (Figure e). As the s-CNT decay rate is substantially lower than those of m-CNTs and defect CNTs (d-CNTs), although the decay rate of double-walled CNTs (DWCNTs) is lower than those of SWCNTs and triple-walled CNTs (TWCNTs), the chirality of ultralong CNTs is enriched around (2n, n) and (n, n – 1). CNT growth termination is associated with catalyst deactivation, such as by carbon deposition on the catalyst surface. Zhang et al. posited that CNT growth terminated because of a discrepancy between the supply and consumption rates of the carbon source (Figure c). When insufficient carbon source is supplied for CNT growth, necking occurs at the contact point between the CNT and catalyst, which, in turn, seals CNTs. Upon resumption of the carbon source supply, multiple nucleation events occur, forming a pea pod shape or new CNT growth on the same catalyst surface. Conversely, when excess carbon source is suppled, the CNT diameter increases, growing a DWCNT inner tube or forming dead carbon caps inside SWCNTs, thereby terminating CNT growth. Furthermore, the authors highlighted that increased stress inside CNTs can reduce the contact degree between the CNTs and catalyst, ultimately terminating growth.

3. Controlled Synthesis of Fine-Structured Carbon Nanotubes

3.1. Diameter

In controlled CNT growth, the CNT diameter can be regulated by adjusting the catalyst size and type. Liu et al. (Figure b) employed monodispersed ferritin as a precursor to prepare catalysts, effectively restricting the CNT diameter to a narrow range of 0.8–1.2 nm. Further, Cheung et al. decomposed iron pentacarbonyl to produce iron nanocluster catalysts of varying sizes. Using iron nanoparticles with average diameters of 3, 9, and 13 nm, they synthesized CNTs with average diameters of 3, 7, and 12 nm, respectively. The CNT diameter exhibited a strong correlation and uniformity with those of iron nanoparticles. Moreover, Li et al. found that the state of the catalyst influences SWCNT tube diameter (Figure c). For SWCNTs grown on solid catalysts, e.g., Co₇W₆, the SWCNT diameter is controlled within an extremely narrow range, irrespective of the catalyst particle size. For SWCNTs grown in liquid Co, the catalyst particle size (d _NP) is nearly equal to the CNT tube diameter (d _NT), while for SWCNTs grown on solid Co, d _NP is greater than d _NT. Thus, during CNT synthesis, the CNT diameter can also be controlled by regulating the state of the catalyst.

4.

Methods for controlling CNT diameter. (a) The effect of temperature on CNT diameter and wall numbers. Reproduced from ref . Available under a CC-BY license. Copyright Jun Gao et al. (b) Schematic diagram of the effect of catalyst particle size on CNT diameter. (c) Effect of reduced temperature on SWCNTs diameter. Adapted with permission from ref . Copyright 2009 American Chemical Society. (d) The correlation between the SWCNTs and catalysts size. Adapted with permission from ref . Copyright 2022 American Chemical Society. (e) The influence of the type of carbon precursor on CNT diameter. Adapted with permission from ref . Copyright 2002 American Chemical Society. (f) A chart showing the distribution of SWCNTs diameters for different carbon precursor concentrations: red, 4200 ppm; green, 14,400 ppm. Adapted with permission from ref . Copyright 2006 American Chemical Society. (g) SWCNT diameter distribution on different substrates. Adapted with permission from ref . Copyright 2005 Elsevier. (h) The effect of catalyst type on SWCNTs. Adapted with permission from ref . Copyright 2019 Elsevier. (i) The effect of the carbon-to-hydrogen ratio on the diameter of SWCNTs. Adapted with permission from ref . Copyright 2022 Beijing University Press.

Alternatively, CNT diameter can also be tuned by adjusting CVD growth conditions. − Wen et al. found that with increasing temperature, both the CNT diameter and number of walls decreased (Figure a). Additionally, Zhang et al. found that the SWCNT diameter decreased with increasing temperature, enabling precise diameter control in small-diameter SWCNT systems through temperature variation (Figure d). Furthermore, the carbon source type influences CNT diameter. Cheung et al. discovered that when C₂H₄ was employed as the carbon source, CNTs exhibited defects, and a considerable amount of amorphous carbon was generated. With CH₄ as the carbon source, relatively clean SWCNTs and DWCNTs can be synthesized (Figure e) because different types of carbon exhibit distinct chemical activities. Carbon–carbon double/triple bonds supply additional energy for CNT growth, resulting in accelerated reaction rates and the growth of larger-diameter CNTs. Meanwhile, the high reactivity of the carbon source promotes spontaneous reactions and other side reactions. By altering the carbon source and optimizing growth conditions, the SWCNT diameter can be controlled more precisely.

Additionally, the carbon source precursor concentration, carbon-to-hydrogen ratio, catalyst type, and substrate material influence the SWCNT diameter. Liu et al. observed that increasing the carbon source precursor concentration from 4200 to 12,400 ppm increased the SWCNT diameter, accompanied by a broader Gaussian distribution peak (Figure f). Li et al. found that increasing the hydrogen flow rate weakened and intensified Raman radial breathing mode peaks at 100–180 and 200 cm^–1, respectively (ω = 227.0/d _t – 0.3 , ), indicating that lower carbon-to-hydrogen ratios reduced SWCNT diameters (Figure i). Kauppinen et al. discovered that replacing a Co catalyst with Fe reduced the SWCNT diameter (Figure h). Moreover, different substrate materials also influence diameter-controlled synthesis. Tsuji et al. demonstrated that switching from an a-plane sapphire substrate to a c-plane sapphire substrate increased the CNT diameter distribution range and average diameter (Figure g).

SWCNTs collapse when their diameter exceeds a critical threshold (∼5 nm). Factors such as chirality can further reduce the SWCNT collapse diameter; for example, when n ≥ 30, the collapse diameter is only 2.06 nm. Therefore, optimizing catalysts and growth conditions is essential for precise SWCNT diameter control.

3.2. Electrical Conductivity

The electrical conductivity of CNTs can be categorized as metallic (m-CNT) or semiconducting (s-CNT) depending on CNT bandgap widths. Compared to m-CNTs, s-CNTs possess a wider bandgap, high carrier mobility, and quantum confinement effects, enabling them to exhibit unique advantages in electronic and optoelectronic devices. Compared to individual s-CNTs or random networks, aligned s-CNTs can provide higher drive currents and lower contact resistances, making them particularly favorable for application in field-effect transistors and integrated circuits. Currently, high-purity s-CNTs are mainly prepared through chirality control, which is divided into two categories: direct synthesis via CVD and postpurification. Compared to postpurification methods, such as density gradient ultracentrifugation, , gel chromatography, , and deoxyribonucleic acid (DNA) sorting, direct s-CNT synthesis facilitates easier maintenance of a higher aspect ratio, produces highly aligned arrays, and reduces reagent contamination. Direct preparation methods mainly involve the precise control of growth parameters, such as catalysts, − substrates, carbon sources, , etchants, − and growth conditions, − or ensuring consistency in conductive properties through methods such as molecular cloning. Wang et al. transformed m-CNTs to s-CNTs via electric field disturbance during growth (Figure a). Specifically, an electric field was applied at time t ₀ to positively charge m-CNTs, while s-CNTs remained uncharged because of their wider bandgap. At time t ₁, the electric field polarity was reversed to negatively charge the catalyst. This polarity reversal disrupted the catalyst, forming chiral junctions, while the external electric field reduced the barrier between m and s transitions compared to that between m and m transitions, thus producing high-purity s-SWCNTs. Because the best template for CNT growth is the CNT itself, Yao et al. used molecular cloning to regrow CNTs on originally cut CNTs, ensuring chirality and thus the consistency of conductive properties; however, this method has a low growth efficiency (Figure b). Moreover, Kenichiro Itamias synthesized carbon nanorings and nanobelts (CNRs and CNBs, respectively) with well-defined structures and segmented them into CNT molecules, revealing the potential of CNRs and CNBs as seed molecules for precisely synthesizing CNTs.

5.

The controllable synthesis of metal or semiconducting SWCNTs. (a) The conversion of CNTs conductivity from metallic to semiconducting type by converting external electric fields. Adapted with permission from ref . Copyright 2018 Springer Nature. (b) Molecular cloning for growth of single chiral SWCNT (blue represents open-ended SWCNTs seeds, red represents repeatedly grown SWCNTs). Adapted with permission from ref . Copyright 2009 American Chemical Society. (c) Schematic diagram of the process for growing s-SWCNTs using SiC as a catalyst and H₂ as an etchant. Adapted with permission from ref . Copyright 2018 Elsevier. (d) Schematic diagrams of the SWCNTs growth process under conditions with and without the introduction of O₂. Adapted with permission from ref . Copyright 2022 Elsevier.

Although metals and their compounds are usually chosen as catalysts for CNT growth, the presence of metal components inevitably introduces impurities that interfere with the conductive properties of the final product. Therefore, Cheng et al. selected SiC as a catalyst to effectively avoid the introduction of a metal catalyst and chose H₂ as an etchant to remove m-SWCNTs, thus obtaining high-purity s-SWCNTs (Figure c). Li et al. found that Fe catalyst clusters containing a trace of O₂ were substantially smaller than O₂-free Fe catalyst clusters because of Fe–O bonds and longer carbon chains, contributing to isolated SWCNTs. Additionally, O₂ can selectively etch m-SWCNTs, facilitating contact between isolated SWCNTs and O₂ and thus m-SWCNT etching (Figure d).

In practical application, besides the energy band structure, the macroscopic CNT conductivity is important. As shown in Figure a, low-conductivity CNTs can be applied to gas sensors; moderate-conductivity CNTs suit transparent conductive films, flexible electronic devices, etc.; while high-conductivity CNTs are applied in lithium-ion batteries, supercapacitors, and other areas. The factors affecting the electrical conductivity of CNTs are primarily categorized as external and internal. External factors include mechanical deformation (Figure b), or doping with different elements (Figure c). Figure b shows the variation in the bandgaps of CNTs possessing different chiralities under tensile stress. Among these CNTs, the bandgap of metallic (5, 5) SWCNTs consistently remained at zero, maintaining excellent conductivity, while the bandgaps of (9, 0) and (9, 6) SWCNTs increased with increasing tensile strength, indicating that m-CNTs may transform to s-CNTs at high tensile strengths. Although the bandgaps of semiconducting (6, 4) and (10, 0) CNTs changed differently with increasing tensile strength, these CNTs remained the semiconducting type. Figure c shows that with increasing AuCl₃ concentration, the sheet resistance of SWCNT films substantially decreases. At lower AuCl₃ concentrations, sheet resistance rapidly drops by approximately 60%, and increasing the AuCl₃ concentration further reduced the sheet resistance, ultimately by approximately 90%, because the Fermi level moved deeper into the valence band with increasing AuCl₃ concentration, reducing the Schottky barrier’s height between the metal and semiconductor nanotubes and, thus, lowering the contact resistance. When the Fermi level was in deeper parts of the valence band and the AuCl₃ concentration was further increased, the Schottky barrier no longer existed, further reducing the sheet resistance.

6.

Influencing factors of conductivity of CNTs and its application. (a) Application of CNTs with different electrical conductivity. (b) Relationship between bandgap and tensile strain for CNTs with different chirality. Adapted with permission from ref . Copyright 2004 Elsevier. (c) Work function and sheet resistance of CNTs doped with various AuCl₃ concentrations. Adapted with permission from ref . Copyright 2008 American Chemical Society. (d) Relationship between wall number and conductivity. Adapted with permission from ref . Copyright 2014 Elsevier. (e) Relationship between average bundle length and conductivity in SWCNTs networks, following the power function σ ∼ L _av ^1.46. Adapted with permission from ref . Copyright 2006 AIP Publishing. (f) Relationship between effective grain size L _α and resistance. Adapted with permission from ref . Copyright 2021 Elsevier.

Moreover, intrinsic CNT properties, such as the wall number, length, and crystallinity, also affect CNT electrical conductivities. As shown in Figure d, the electrical conductivity of CNT fibers initially increased and then decreased with increasing average wall number, reaching a maximum value of 169 S cm^–1 at an average wall number of 2.7. Figure e reveals the relationship between the average CNT bundle length (L _av) and conductivity (σ), showing that σ was proportional to the 1.46th power of L _av and gradually increased with increasing L _av. The CNT grain size also affects σ (Figure f), with a very sharp threshold observed at L _α ≈ 11 nm, where the resistance of the low-crystallinity CNTs is nearly 50 times higher than that of high-crystallinity CNTs. This substantial difference originates from the single-grain size within the tube walls, akin to confinement in graphene nanoribbons.

3.3. Chirality

Chirality reflects the SWCNT structure and band structure and is vital for their applications. ,− The direct synthesis of single-chirality SWCNTs has long been pursued, and considerable progress has been made.

The thermodynamic model for the chirality-controllable growth of SWCNTs (Figure a), proposed by Christophe et al., categorizes carbon atoms at CNT edges as armchair- and zigzag-edged C atoms. The interfacial energy between the zigzag-edged C and the catalyst is higher than that between the armchair-edged C and the catalyst. The combination of armchair- and zigzag-edged C configurations enables the formation of diverse chiral CNT edge structures, facilitating computational elucidation of the correlations among the CNT–catalyst interfacial energy, temperature, and chirality, which suggests that SWCNT chirality originates from the entropy-driven nanotube boundary configuration.

7.

Thermodynamic model for chiral controlled growth of SWCNTs: (a) the symmetry matching strategy; (b) and the structure matching strategy; (c) the configurational entropy drive theory. Kinetic model for chiral control: (d) the helical dislocation theory to achieve single chiral enrichment; (e) and chiral switching (f). A continuum model (g) indicating a synergistic regulatory strategy for obtaining a narrow chiral distribution (h). (a) Adapted with permission from ref . Copyright 2018 The American Association for the Advancement of Science. (b) Adapted with permission from ref . Copyright 2017 Springer Nature. (c) Adapted with permission from ref . Copyright 2022 American Chemical Society. (d) Adapted with permission from ref . Copyright 2009 National Academy of Sciences of the USA and ref . Copyright 2011 American Physical Society. (e) Adapted with permission from ref . Copyright 2021 American Chemical Society. (f) Adapted with permission from ref . Copyright 2019 Elsevier. (g) Adapted with permission from ref . Copyright 2014 Springer Nature.

On the basis of the “entropy-driven” theory and inspired by enzyme-catalyzed structure matching, Li et al. hypothesized a relationship between catalysts and CNT structures and developed high-melting-point, low-symmetry Co₇W₆ alloy catalysts for growing (12, 6), (16, 0), and (14, 4) chiral-enriched SWCNTs on (0 0 12), (1 1 6), and (1 0 10) catalyst facets, respectively. ,, This discovery points to a new direction for the structurally controlled growth of CNTs.

Considering the epitaxial relationship between the growth structure of CNTs and the surface crystal structure of solid catalysts, Zhang et al. proposed a crystal-floating growth model and a more specific symmetry-matching strategy (Figure b). This strategy posits that when the catalyst crystal surface and CNT chiralities are satisfied, CNTs of a specific chirality exhibit reduced nucleation barriers and accelerated growth rates on the catalyst surface. By designing the lattice planes of solid WC and Mo₂C catalysts, (12, 6) and (8, 4) CNTs were enriched with specific symmetries. In addition, Zhao et al. employed in situ ETEM (Figure c) to capture the nucleation and growth of SWCNTs on Co₇W₆ catalysts via the VSS pathway, revealing the formation of a (16, 0) CNT matching the crystalline structure of the Co₇W₆ catalyst grown vertically on the (1 1 6) facet. This elucidated the matching relationship between the atomic scale of CNTs and catalysts at a deeper level.

Based on the correlation between the crystal structures, researchers have proposed a kinetic model for chirality-controllable growth (Figure d), namely, helical dislocation theory. , In this model, theoretical calculations have revealed that C atom insertion is the pivotal step for SWCNT growth, which is analogous to Frank dislocation-assisted crystal growth. ZZ tubes were postulated as the slowest-growing “perfect” SWCNTs, whereas large-hand-corner SWCNTs exhibited accelerated growth.

However, helical dislocation theory has limitations, specifically, the inadequacy of the growth model for solid catalysts. , Ding et al. proposed a more sophisticated kinetic model, , which hypothesizes that when SWCNTs are grown vertically on the surfaces of solid catalyst particles, both AC and ZZ-SWCNTs exhibit the slowest growth rate, while (2n, n) SWCNTs with the maximum number of active sites demonstrate the fastest growth rate. The SWCNT growth environment can be divided into two categories: etchant-sufficient and etchant-insufficient. Under etchant-sufficient conditions, the growth and predominant chirality of CNTs follow the helical dislocation theory. Conversely, under etchant-insufficient conditions, the kinetic model suggests that a high C/H ratio is conducive to the enrichment of (2n, n)-chirality tubes. Additionally, (12, 6) single-chirality tubes have been successfully grown by increasing the CO flow rate (Figure e). Moreover, this model has further guided the synthesis of highly pure near-zigzag CNTs.

Studies on growth mechanisms have consistently demonstrated that the catalytic interface governs CNT chirality evolution. The morphology and atomic structure of the catalyst surface have been shown to directly affect the edge structure as well as the final chiral structure of CNTs. Consequently, possible CNT-enriched species can be predicted by analyzing the catalyst state and conformation of the crystalline surface. Based on this, Zhang et al. predicted the possibility of growing enriched chiral CNTs for solid Co catalysts (Figure f) and suggested that a catalyst structure favorable for growing ZZ-SWCNTs would afford the selective growth of (n, n – 1)-group SWCNTs by controlling the catalyst size distribution and growth rate. Their prediction was subsequently corroborated experimentally, thereby verifying the thermodynamic engineering of symmetry-matched nucleation and kinetic regulation of the SWCNT growth rate to enrich single-chirality CNTs.

At the CNT–catalyst interface, the synergy between thermodynamics and growth kinetics enhances the chirality of CNTs. Yakobson et al. integrated thermodynamic and kinetic theories pertinent to CNT growth and proposed a nanotube–catalyst continuum model (Figure g), which explains the distribution of chiral abundance observed in CNTs, as postulated by the VSS model. In their model, growth kinetics are facilitated by kinks at the tube edges, promoting the growth of chiral CNTs, where the growth rate is proportional to the chiral angle. By contrast, the inability to perfectly align the liquid catalyst surface with the CNT edge structure introduces a thermodynamic nucleation barrier, favoring the growth of nonchiral kink-free tubes. The functional image revealed that for opposite thermodynamic and kinetic trends, the total product abundance was maximized for (n, n – 1) CNTs.

As shown in Figure h, the chirality-controlled SWCNT growth strategy is divided into three steps: (1) in the early stage, SWCNT end-cap nucleation on the catalyst surface is determined by the thermodynamics of the matching between the CNT lattice and catalyst crystal facets. (2) SWCNT growth is determined by the CNT growth reaction kinetics, which is influenced by both the C/H ratio and etchant. Third, during CNT growth, regulation of the conditions, including the catalyst particle size and reaction temperature, controls the number of walls as well as the tube diameter, ensuring the effectiveness of the chirality-controlled SWCNT preparation.

4. Conclusion and Outlook

This paper comprehensively reviewed recent advances in the controlled CVD preparation of finely structured CNTs. We summarized commonly used CVD methods and provided an in-depth comparison of their CNT growth rates, costs, qualities, purities, and yields. Subsequently, we presented an overview of the fundamental principles governing CNT growth, focusing on the VSS mechanism operating on a solid catalyst surface and the VLS mechanism within a liquid catalyst system. Additionally, we examined chirality transition, tube length selection, and growth termination modes during CNT growth.

Structure-controllable synthesis conditions determine the fine structures of CNTs and ultimately their applications. Since the discovery of CNTs in 1991, researchers have made significant strides in the controlled preparation of CNTs with a fine structure, including tuning and optimizing CNT wall numbers, tube diameters, conductive properties, and chiralities, by designing catalyst particle sizes and structures, modulating the growth atmosphere, adding etching agents, and applying external electric fields. Collectively, advances in catalysts, reframing the CNT growth mechanism, and breakthroughs in characterization technology have collectively contributed to continuous breakthroughs in structure-controlled CNT preparation. Table summarizes the recent achievements for various CVDs.

1. Progress and Highest Purity of CNT Fine Structures.

CNT fine structure		method(s)	highest purity
number of walls	single wall	designing the catalyst particle size and structure	99.9%
	double wall		90%
length	ultralong	“furnace-moving” method	550 mm
		substrate interception and direction strategy
conductivity type	metallic (m-)	etchant (SO3)	99.9999% ,
		quasi-static CVD
		defining the catalyst structure
	semiconducting (s-)	chemical etching methods (O2, H2O, etc.) ,	99.9999% ,
irradiation methods (plasma, ultraviolet light, long-arc xenon lamp, microwave, etc.) −		irradiation methods (plasma, ultraviolet light, long-arc xenon lamp, microwave, etc.) −
chirality	(n, m)	combining nanotube/catalyst interfacial thermodynamics with kinetic growth theory	97% (14, 4)
			95% (6, 5)
			92% (12, 6)

Despite all these advancements, SWCNTs still face formidable challenges in their industrial application, which can be broadly categorized as follows: (1) the thermodynamic and kinetic parameters influencing SWCNT growth are numerous and complex, making it challenging to develop a reliable synthesis model using traditional mathematical modeling techniques. (2) Conventional CVD is highly opaque, lacking the capability for the real-time monitoring of CNT growth and collection of experimental data, thereby directly limiting the accuracy of constructed controlled-growth models. (3) The existing CVD equipment exhibits low CNT growth efficiency and poor parameter-control precision, generating inconsistent product quality and increasing production costs. To achieve the controlled preparation and large-scale production of fine-structure SWCNTs, four outlooks are proposed in addition to the integration of research, development, and application of AI in materials science (Figure ). In controllable CNT fabrication, the core AI mechanisms encompass generative design (optimizing virtual growth processes through AI-driven digital twins and multiscale simulations), closed-loop control (dynamically adjusting processing parameters based on in situ monitoring feedback), and inverse performance design (establishing quantitative process–microstructure–property mapping relationships).

8.

AI technology in the controlled CVD synthesis of SWCNTs. Adapted with permission from ref . Copyright 2025 Cell Press. Reproduced from ref . Available under a CC-BY license. Copyright Daniel Hedman et al. Adapted with permission from ref . Copyright 2021 John Wiley and Sons. Adapted with permission from ref . Copyright 2021 The American Association for the Advancement of Science. Reproduced from ref . Available under a CC-BY-NC-SA license. Copyright Fengrui Yao et al. Reproduced from ref . Available under a CC-BY-NC-SA license. Copyright Rufan Zhang et al.

4.1. Development of High-Throughput, Multiscale Material Computational Simulations

CNT growth is influenced by a multitude of thermodynamic and kinetic parameters, including chemical potentials and interfacial, surface defect, binding, diffusion, and reaction activation energies, among others. Conventional computational simulation methods are inadequate for decoupling high-dimensional relationships and coupling complex information. ML-empowered high-throughput multiscale material computations offer a solution to this problem.

This technique fuses high-throughput computing and multiscale simulations, employing ML algorithms to optimize and accelerate materials research. High-throughput computing employs extensive resources and parallel processing to facilitate the rapid calculation and prediction of material properties and characteristics. In ML, the analysis and mining of a substantial corpus of experimental data and computational results can assist researchers in identifying underlying mechanisms and trends, accelerating materials design and optimization, and should contribute to a more comprehensive understanding of the CNT growth mechanism and enhance the global understanding of CNT materials. Furthermore, this will provide a direction for future experimental design and theoretical development, particularly in the selective growth of CNTs possessing specific fine structures.

4.2. Establishment of Efficient Methods for Designing and Preparing Catalysts

To prepare high yields of high-purity, high-quality SWCNTs via structurally controlled growth, the ideal catalyst must exhibit high catalytic activity, ,, high-temperature thermal stability, ,,,, suitable carbon solubility, and carbon diffusivity. − ,

To overcome the limitations of traditional catalysts, interpretable ML should be used to predict catalytic activities by extracting physical meanings. Moreover, an LLM should be used to recommend catalyst components tailored to the requirements of different application scenarios. Therefore, AI will enable researchers to transcend the constraints of traditional metal and conventional alloy catalysts, expand the scope of catalyst compositions, consider the integration of metal/alloy/nonmetal oxide/carbide and other catalyst systems, and achieve multiobjective optimization and on-demand customization for controllably preparing CNTs.

4.3. Construction of Intelligent, Automated Experimental Platforms

Standardization is the foundation upon which data are shared and reproduced and scientific knowledge is iterated. However, material preparation data obtained using different equipment in different environments by different operators often vary considerably, leading to nonuniform data acquisition. Furthermore, data, including images, spectra, and structural data, generated using different characterization equipment types are inconsistently formatted. Despite the extensive literature on structurally controllable CNT growth, previous studies have typically focused on only a few experimental parameters. To date, most studies have been based on trial-and-error experiments, producing a range of outcomes because of differences in growth methods, experimental environments, and data acquisition and generating uncertainty in the accuracy and portability of relevant experimental parameters.

The AI-based intelligent automatic experimental platform enables the precise and efficient execution of experiments, thereby mitigating human errors associated with manual operations, data recording, and subjective cognitive biases. Notably, however, different AI models exhibit unique capabilities and limitations. Therefore, the selection of a suitable model should be tailored to the specific characteristics of the experimental system, guiding the customized development of intelligent experimental platforms. To this end, we systematically compared the performance metrics and interpretability mechanisms of traditional ML models and LLMs for optimizing CNT synthesis, as summarized in Table .

2. Comparative Analysis of Traditional ML Models and LLMs in Optimizing CNT Synthesis.

aspect	traditional ML models	LLMs
predictive accuracy	achieve an R 2 value of 0.92 for predicting CNT array densities	attain 93% accuracy in catalyst screening (e.g., carbon BERT)
data dependency	require 300–500 experimental samples for reliable analysis; limited applicability in low-data scenarios	require pretraining and fine-tuning with >50,000 publications and 1000 experimental data sets to achieve high generalizability
interpretability mechanisms	leverage SHAP values to quantify feature contributions (e.g., d-band center shifts account for 95% the semiconductor purity weight)	utilize attention mechanisms to correlate synthesis parameters (e.g., carbon flux and growth temperature) with outcomes
risk control	prone to data overfitting, producing poor model generalizability	exhibit a hallucination risk of ∼43.75% in generated recipes, necessitating human–AI collaboration for validation

The intelligent automatic experimental platform offers unparalleled advantages in experimental scientific research, and its construction for CNT growth should markedly enhance experimental repeatability and productivity. In addition, automated characterization equipment should be integrated into the experimental platform (for instance, in situ Raman scattering), for automatically generating high-quality characterization data, thus accelerating process optimization and establishing a data foundation for developing a standardized material database and training controllable CNT growth models.

4.4. Setup of Pilot-Scale Chemical Vapor Deposition Equipment

Materials that progress from the laboratory to industrialization always encounter numerous discrepancies. The goal of laboratory research is to develop theories, methods, and indicators, whereas industrial applications are primarily concerned with yield, consistency, stability, cost, and economic benefits. Consequently, the mass production of SWCNTs encounters severe obstacles in equipment, lengthy optimization, and elevated costs.

The development of pilot-scale intelligent equipment integrating research and batch production will bridge the gap between industry and research. Laboratory-scale CNT synthesis should be transferred to the pilot line, and SWCNT production can be further optimized using ML to improve the structural purity and yield. Furthermore, an online monitoring and feedback system should be introduced to detect potential deviations in preparation in real time and implement timely adjustments to enhance production efficiency, reduce unit preparation costs, and ensure the stability and consistency of prepared SWCNT batches. The integration of intelligence and automation, which empower AI, into each aspect of SWCNT production, research, and development will bridge SWCNTs’ lab-to-enterprise divide.

⊥.

Q.H. and Y.L. contributed equally to the work.

The authors declare no competing financial interest.