Machine Learning-Driven Analysis of Nanostructures for Enhanced Structural Characterization

INTRODUCTION
Machine learning approaches offer a new lens through which to view the analysis of nanostructures for enhanced characterization. Such advanced characterization is essential for progress in the field of nanotechnology. Nanotechnology is an increasingly promising field with potential applications in sustainability and electronic materials; however, it also requires the ability to define and understand structures at the smallest scale. The surface structure modeled in this work is previously simulated, as few molecular-level characterization methods could adequately model the surface structure. The ability to accurately model the entire metal surface is important for correctly representing behavior in an electrochemical system [1]. In nanotechnology, as industry and the engineering solutions to societal problems grow, characterization at the nanoscale (accuracy in the structures that vary from approximately 100 nm to 1 nm) is still an essential roadblock that must be addressed. “Nanostructures” are molecularly defined objects with features between 1 nm and 100 nm. These structures can confer new behavior and properties for the underlying material and have been the focus of much attention in recent years for their potential applications in sustainability, including the development of more efficient photovoltaic and electronic materials. The meeting of the two fields between scanning technology and machine learning opens possibilities for one of the remaining grand challenges in nanoscience [2, 3]. 
By the close of the last century, a considerable amount of attention had been focused on nanostructured materials and nano-enabled instrumentation, including biological, physical, and chemical sensor arrays, multiferroics, photonic materials, and engineered composites. Enhancements in mechanical properties, catalysis, tribological coatings, and molecular recognition species also fall into this domain. Consequently, micro/nano-electromechanical systems, microsystems technologies, and nano-electromechanical systems have had significant applications in science as well as industry. The SEM and TEM spectroscopic observation methods are commonly used to characterize experimental results [4]. Subsequently discovered methods like X-ray photoelectron spectroscopy, Auger electron spectroscopy, and scanning tunneling microscopy methods are also currently employed, although they require expensive equipment and are often complicated to operate. 
Now, the use of machine learning algorithms together with manufacturing methods has – and continues to show – significant promise. Machine learning continues to bring about transformative developments across a variety of domains. Despite this, the desire to utilize machine learning as a tool for the fine-scale manipulation and optimization of materials remains comparatively underdeveloped in the domain of nanotechnology, where it could revolutionize experimental methodologies for the analysis of nanostructure properties, helping to bridge the gap between theoretical assumptions and practical applications, and providing advanced tools for the fabrication of new types of materials with superior properties [5, 6]. Furthermore, recent studies emphasize the relevance of tailoring nanoparticle properties through advanced techniques such as focused ion irradiation and bio-synthesis, providing a foundation for synergistic integration with ML frameworks [7, 8]. Recent progress in quantumdot cellular automata (QCA) show how nanoscale logic devices can benefit from precise structural design, further motivating MLdriven characterization for defect detection and performance prediction [9]. A growing body of work demonstrates how artificialintelligence pipelines can directly accelerate nanotechnology breakthroughs across medicine, energy, and materials science [10]. 
This study pursues three objectives: to assess the capacity of machine learning methods to describe key nanostructural characteristics including adsorption energy, pore diameter, and surface area; to use cutting-edge analytical and machine learning approaches to create connections between structural factors and functional results, supporting a predictive framework; and finally, to illustrate the accuracy and automation gains by contrasting the suggested methodology with current ML-driven nano-characterization techniques and provide a strong, flexible machine learning framework for automated nano-analysis that improves scalability for a range of applications while lowering manual labor. The overarching objective of this study is to provide practical advice on how to incorporate machine learning into processes for computational and experimental nanotechnology. 
Based on dimensionality, materials can either be classified as bulk or nano; the latter description possesses unique physical properties distinct from bulk materials. In the present analysis of nanostructures, they are classified as such based on sizes that fall in the range of a few nanometers, where quantum effects begin to play a significant role. Owing to this, properties like catalytic activity, conductance, magnetization, physical strength of a material, and its melting point can be distinctly different from their corresponding bulk material. Such quantum effects endow nanostructures with distinct physical and chemical properties, most notably a high surface-area-to-volume ratio, strong size-dependent behavior, and enhanced quantum confinement effects, having no direct analogue in bulk materials. Practically, different types of nanostructures are used, illustrated in Fig. 1, such as nanocomposites, quantum dots, fullerenes, nanorods, and nanowires. Zero-dimensional (0D), one-dimensional (1D), and two-dimensional (2D) are the characteristic dimensions of nanostructures. The 0D nanostructures include nanoparticles, quantum dots, and fullerenes, where all the atoms are confined in three dimensions. The 1D and 2D nanostructures include nanorods, nanowires, and nanosheets/nanoplates, as some atoms are free to move in one or two dimensions, respectively. The nanostructured materials, in general, exhibit significantly enhanced properties relative to their bulk counterparts, primarily due to their large surface-area-to-volume ratio. In order to increase the surface area, the dimension should be as small as possible, typically below only a few nanometers. Furthermore, the increase in surface area also leads to a redshift in their properties, such as a smaller bandgap, as better described in the following sections. While various properties can be modeled, the present review focused predominantly on the catalytic properties of adsorption and reaction [11, 12]. The high surface-to-volume ratio underpins a range of functionalities relevant in vaccine delivery systems, biofilm inhibition, and electrochemical applications. Among these systems, graphitic carbon nitride (gC3N4) photocatalysts exemplify twodimensional nanostructures whose visiblelightactivated behavior depends strongly on lattice defects and surface terminations; machinelearning pipelines can expedite the discovery of such structure–activity relationships [13]. These nanostructures also exhibit distinct antimicrobial, catalytic, and morphological behaviors that are governed by structural tuning and environmental conditions, as demonstrated in biosynthesized and conjugated nanoparticle systems [7, 14, 15].  
A wide range of machine learning algorithms is available for use in nanostructure characterization (Fig. 2). On a high level, these algorithms fall into two broad categories: supervised machine learning, which models the relationship between input samples with known outcomes, and unsupervised machine learning, which identifies patterns among samples without labels. Supervised machine learning is further subdivided into linear, non-linear, ensemble, and particularly, neural network models. This type of machine learning has attracted considerable attention from the research community because of its strong performance across both linear and non-linear analysis tasks. Unsupervised methods, by contrast, consist of clustering and dimensionality reduction algorithms, varying in suitability depending on the structure and scale of the dataset. Regardless of the algorithm selected, computational efficiency, running time, achievable accuracy, and consistent outcomes are key evaluation criteria that must be considered during model selection [16].
Generally, the use of machine learning algorithms in nanostructure analysis has led to significant returns; however, their application does not always produce genuinely novel contributions beyond what prior works have already established. In what follows, each algorithm is discussed in terms of its superiority and suitability for specific characterization tasks. For instance, Linear Discriminant Analysis (LDA), as a linear classifier, is able to achieve efficient classification of nanostructures, but it is constrained by its assumption of a linear decision boundary. Random Projection (RP) and Independent Component Analysis (ICA), while computationally demanding, can uncover latent chemical and physical features in select nanostructure areas. Neural networks (NN) are better suited to modelling composite nanostructures and offer improved predictive accuracy on the effects of the properties. In practice, various combinations of these machine learning techniques are employed to explain the chemical and physical performance of nanostructures in detail. Major considerations for the selection of the algorithm are the level of complexity and size of datasets, and computational efficiency. For instance, 2D Finite Automaton (FA) is faster than its 3D, multinomial, and partial least squares regression FA counterparts, which can be used in the correlation analysis of power epoxy floor with desired properties, despite correlation by 3D FA generally providing better predictive properties [17, 18].
Machine learning has found practical applications in the development of improved resolution or capabilities for the characterization of nanostructures, as well as driving improvements in the analytical resolution and interpretive capability, as summarized in (Fig. 3). This section examines how machine learning addresses longstanding challenges in nanostructure characterization techniques, focusing particularly on the enhancement in both resolution and accuracy. Machine learning is not recipe-driven, but generally needs to be trained on as many samples as possible to be as generalizable as the data would allow. In this regard, machine learning can serve to address the lack of reliable angle information in large volume-oriented techniques regardless of its specific case, i.e., interference, emission, or scattering-based. Documented case studies focusing on successful applications and analysis of their broader implications could provide a sound basis for further development in nanostructure studies, highlighting future directions in respective academic or industry fields. Beyond optimizing existing techniques, machine learning opens pathways to new modes of nanostructure analysis. 
Machine learning has been actively explored and used in a variety of fields, from materials science to biological fields, to overcome complex characterization tool responses. It is understood that nanostructured materials, rationally designed and fabricated, offer valuable functional characteristics across a wide range of research fields including optoelectronics, energy research, bioengineering, and environmental science. The prediction of graphene nanofluid viscosity through combined responsesurface methodology and ML regression underscores the role of hybrid statistical–ML frameworks in predicting transport properties critical for thermofluidic nanodevice design [19]. Combined computational fluid dynamics and ML studies on nanofluidenhanced heat exchangers demonstrate how datadriven surrogate models accelerate thermal optimization with the simultaneous variation of both nanoparticle concentration and baffle geometry [20]. ML-assisted frameworks have also been successfully applied in studies involving nanoparticle synthesis for biomedical and agricultural domains. For example, ML can help optimize surface acoustic wave applications in tumor modeling [21, 22] or predict plant behavior in agricultural systems [23] or assist in pollution issues [24, 25]. The rapid pace of advances in nanofabrication technology has necessitated increasingly precise and versatile characterization tools [6, 26, 27]. Beyond laboratory settings, ML models have been deployed for large scale environmental monitoring predicting pollution sources in the Euphrates and Tigris rivers [28] and performing real time biodiversity assessments via deep learning image analysis on the Tigris River [29] highlighting the versatility of datadriven frameworks developed in this study.

MATERIALS AND METHODS
This study applies established machine learning algorithms to the problem of nanostructure characterization rather than developing new ML methods. The contribution lies in the integration of heterogenous data sources and the comparative benchmarking of algorithms, not in algorithm design (Fig. 4). For the structure editing, the widely known, versatile, freely available molecular dynamics package LAMMPS, was used for the simulation of particle ensembles. In this study, the input file for the single aluminum nanoparticle consists of a single atom with a defined crystallographic orientation, from which the full FCC lattice was generated to the target diameter. Generated metal nanoparticle generally fell within the conventional 1-100nm size range characteristic of metallic nanostructures. Recent implementations demonstrate how Taguchi design-of-experiments coupled with supervised learning can optimize tribological responses of nanoparticle‑reinforced polymers [32], offering a transferable template for parameter‑tuning in our nano‑characterization pipeline.
Fourteen machine learning algorithms were initially considered, spanning a variety of methodological families: linear models (Linear Regression, Ridge Regression, Lasso, Logistic Regression, LDA), instance-based methods (k-Nearest Neighbours), tree-based and ensemble methods (Decision Tree, Random Forest, Gradient Boosting, AdaBoost), kernel methods (Linear SVM, RBF-kernel SVM), probabilistic models (Naïve  Bayes), and neural networks (Multilayer Perceptron). Selection of algorithms for the detailed reporting was guided by the criteria: low complexity, predictive accuracy, training time efficiency, learning ability, etc. The four algorithms presented in this study (Random Forest, SVM, MLP, and Gradient Boosting) emerged as the top performers under these criteria. All algorithms were evaluated under a unified protocol: the dataset was partitioned into an 80% training set and a 20% held-out test set, with a 5-fold stratified cross-validation applied within the training partition for hyperparameter selection. Performance was quantified using accuracy, precision, recall, F1-score, and ROC-AUC, enabling direct comparison across algorithmic families and against prior benchmarks reported in the literature. Microfluidic-based experimental setups, previously validated in chemotaxis and biofilm studies, were referenced to refine our nano-environmental controls [33].

Experimental Setup
The dataset used in this study was assembled from a combination of experimental and simulated sources. Scanning Electron Microscopy (SEM) imaging was performed using a TESCAN system to capture high-resolution surface morphology, while Atomic Force Microscopy (AFM) measurements were conducted using an AraResearch instrument to obtain topographical information and pore-related features. Simulated nanostructures were additionally incorporated to introduce controlled variations in geometric and structural properties that are difficult to isolate experimentally. These combined data sources were then integrated into a unified analytical dataset for subsequent machine learning analysis. To enable joint modelling across modalities each of the 900 records was treated as a distinct specimen described by a common feature schema.

Data collection
This integrated dataset comprised 500 SEM images, 300 AFM images, and 100 simulated structures, yielding 900 records in total (Table 1). SEM images mainly supported surface area analysis, while AFM images were used for surface roughness and pore related morphology, and simulated structures contributed structural features as part of the machine learning input. The final feature set comprised surface area, pore diameter, aspect ratio, conductivity, adsorption energy, and reconstruction descriptor. Surface area and pore diameter are attributed to SEM and AFM derived data, while computational models derived the aspect ratio, conductivity, adsorption energy, and reconstruction descriptor (Table 2). The reconstruction descriptor is a binary structural feature indicating whether a given surface exhibits atomic-level reconstruction relative to the ideal bulk-terminated geometry. Surfaces showing measurable atomic displacement from the bulk positions were assigned a value of 1, while the unreconstructed surfaces were assigned 0. The surface area, pore diameter, and adsorption energy values were scaled using Min-Max normalization. Aspect ratio used Z-score normalization, while conductivity was left unscaled. The reconstruction descriptor was encoded as 0/1, suitable for direct use in tree-based and gradient-boosted models without further transformation. Following feature extraction, normalization, and standardization, the resulting dataset was used as an input for the model training and performance comparison (Table 3).

RESULTS AND DISCUSSION
The results show that the investigated nanostructure descriptors display moderate variation across the dataset, with surface area, pore diameter, aspect ratio, and conductivity all exhibiting relatively close mean and median values, indicating a generally balanced distribution of observations. As summarized in Table 4, surface area had a mean of 25.3 nm²/g and ranged from 10 to 40 nm²/g, while pore diameter averaged 12.7 nm with a range of 5–20 nm. Aspect ratio showed the smallest variability among the measured parameters, with a mean of 1.25 and a standard deviation of 0.15, whereas conductivity had a mean value of 5.8 S/m and ranged from 3 to 7 S/m. The binary reconstruction descriptor was distributed as 288 samples with the descriptor, and 612 samples without. Collectively, these statistics confirm that the dataset captures meaningful structural and functional variability without the extreme dispersion that can destabilise model training.
A strong positive association was observed between surface area and adsorption energy as shown in Table 5 and Fig. 5, which show that adsorption energy increased steadily from 55 to 95 kJ/mol as surface area rose from 15 to 35 nm²/g, with the binned averages tracing an approximately linear trend. In addition, pore diameter displayed a clear positive relationship with conductivity. As shown in Table 8 and Fig. 6, conductivity increased from 3.5 to 7.1 S/m as pore diameter rose from 5 to 20 nm. This pattern is consistent with the physical mechanism by which increased surface area exposes a greater density of active binding sites per unit mass, enhancing the cumulative interaction energy between adsorbate molecules and the substrate. The correlation observed here should be interpreted as evidence of association rather than direct causation, since latent factors may co-vary with surface area and contribute to the observed adsorption behavior.
The correlation analysis quantitatively confirms that adsorption energy is mainly influenced by morphology-related parameters, with the Pearson correlation coefficient r used as the measure of linear correlation. As presented in Table 6, adsorption energy had its strongest correlation with surface area (r = 0.87), followed by pore diameter (r = 0.78). These findings indicate that both the extent of exposed surface and pore geometry are important factors governing adsorption behavior. In comparison, conductivity showed a more moderate relationship with adsorption energy (r = 0.54), while aspect ratio demonstrated the weakest association among the examined variables (r = 0.49). The matrix also revealed a high correlation between surface area and pore diameter (r = 0.85), reflecting the fact that porous nanostructures tend to gain both surface area and pore volume simultaneously as porosity increases.
The machine learning evaluation demonstrated strong predictive performance across all tested algorithms. As shown in Table 7, the Multilayer Perceptron (MLP) achieved the best overall results, with an accuracy of 94%, precision of 93%, recall of 92%, F1-score of 93%, and the highest ROC-AUC value of 0.97. Random Forest also performed comparably, yielding the highest accuracy at 94.2% and a ROC-AUC of 0.95. Gradient Boosting produced intermediate performance, whereas the Support Vector Machine showed the lowest metrics among the four models, though all four models comfortably exceeded the performance reported for traditional linear baselines. Overall, these results indicate that non-linear and ensemble models are well suited for capturing the structure-property relationships in this dataset. Their advantage is consistent with the collinearity and partial non-linearity observed in the correlation analysis. The closeness of the MLP and Random Forest results further suggests that the predictive ceiling on this dataset is being approached, and that future gains will require either richer feature spaces or larger sample sizes rather than further algorithmic tuning.
The results presented in Section 7 confirm the efficacy of machine learning algorithms as a tool for characterization of nanostructures, while also clarifying the structural drivers behind the observed adsorption behavior. In particular, the strong positive correlations observed between adsorption energy and both surface area (r = 0.87) and pore diameter (r = 0.78) demonstrate the significance of the relationship between physical morphology and thermodynamic behavior. This relationship indicates that the optimization of physical dimensions is a primary driver in the functional design of nanomaterials, although it must be acknowledged that the observed correlations reflect statistical associations rather than direct causation. Latent factors such as local electronic structure, defect density, and surface termination chemistry are known to co-vary with surface area and pore geometry, and likely contribute alongside the morphological descriptors quantified here. From an algorithmic perspective, the MLP demonstrated superior performance, achieving the highest predictive accuracy (94.0%) with an exceptional ROC-AUC score of 0.97 demonstrating the capacity of non-linear neural architectures to capture the complex nature of nanoscale phenomena more effectively compared to traditional linear models. Furthermore, we show that predictions were very sensitive to reconstruction descriptors, underscoring that atomic-level surface configuration carries information that is not redundant with bulk morphological features, which ML models of nanostructures should incorporate where available. To contextualize these advancements within the broader landscape of automated nano-characterization, Table 9 compares the proposed framework against recent data-driven approaches in the literature.
The proposed framework aligns well with recent advancements in automated nanostructure imaging. The automated feature extraction pipeline utilized in this study is conceptually parallel to the YOLO-based keypoint detection deployed by [35] for STM image analysis. Likewise, Nagy et al. [26] successfully employed ML to extract crystallite size and dislocation density from XRD data; the present work extends this principle by demonstrating that foundational structural metrics, most notably surface area and pore diameter, can be reliably captured and correlated with thermodynamic outcomes.
The novelty of the present framework lies not in algorithmic invention, but in four integrative choices that distinguish it from prior ML-based nanostructure work. First, it combines three heterogenous data sources: SEM imaging, AFM topography, and atomistic simulation, into a single feature schema, whereas comparable studies typically rely on a single modality (for example, XRD in [26], STM in [35], Raman in [36]). Secondly, it incorporates the reconstruction descriptor as a model input, capturing the atomic-level surface variation that pure morphological pipelines omit. Thirdly, it benchmarks a variety of algorithms on a common dataset rather than tuning a single model, allowing the inter-algorithm comparison reported in Table 7. Fourthly, it maps structural features directly to a meaningful thermodynamic target, rather than to an intermediate spectroscopic or geometric measure. Collectively, these choices reposition the role of ML in nano-characterization from a per-modality acceleration tool towards a unified structure-property prediction framework.
Looking forward, this machine learning framework presents several promising extensions which merit future investigation. Incorporating Convolutional Neural Networks (CNNs) for direct, aberration-corrected image processing, as demonstrated by [34, 41-44], could further eliminate manual pre-processing bottlenecks that currently constrains throughput. Extending these predictive models to additive-manufactured bio-nanocomposites [38] could accelerate data-driven material design, paving the way for in-situ analytics during 3D printing. Finally, as proprietary computational datasets for advanced nanomaterials grow in scale and value, ensuring the integrity of these frameworks through secure data pipelines will become increasingly critical, particularly when integrating AI with sensitive biomedical and nanotechnological applications [39].
The experimental results from training the machine learning algorithms to accomplish the nanostructure analysis highlighted that the score of the described error depends on the prediction produced per specific machine learning algorithm. At the same time, the results in Tables 6 and 7 are evidence of the clear ability to separate factor scores from the predicted class, especially when machine learning tree algorithms were implemented. The conducted preliminary analysis of the results shows that using machine learning algorithms for further nanostructure characterization is promising, and the next step will include the evaluation of the physical interpretation of the outcome. According to the conducted analysis of the developed data mining application results, we have observed a pattern that seems to confirm the original research hypotheses. Using the methods resulted in the separation of the specific factor values for each class of nanostructures, which supports the fact that using advanced tools for multidimensional arrays analysis for nanostructure recognition could be an effective method [45].
Finally, the direct validation of the experimental results is quite problematic since the development of the collection of benchmarks for analytical applications is a highly nontrivial task. Apart from that, a deeper understanding of nanostructure behavior would be required, including a clearly stated collection of relevant phenomena and disparate theoretical and experimental behavior points. Since we do not have this information available, and the tailored algorithm has no behavior limitations, a straightforward comparison is impossible. A broader survey confirms that MLcentered reviews on nanofluid heattransfer enhancement position data analytics as a linchpin for sustainable energy applications [40, 46], further validating the crossdomain relevance of our study.

CONCLUSION
This study presents an integrated machine learning framework for nanostructure characterization that unifies SEM imaging, AFM imaging, and atomistic simulation into a single 900-specimen dataset. By integrating heterogeneous data sources into a unified framework, the proposed methodology enables accurate prediction of adsorption energy based on structural descriptors. The framework was benchmarked across four models on the basis of predictive accuracy, training stability and computational efficiency. Among these, the Multilayer Perceptron achieved the highest ROC-AUC value of 0.97 at an accuracy of 94.0%, while Random Forest reached a comparable 94.2% accuracy. Correlation analysis identified surface area (r = 0.87) and pore diameter (r = 0.78) as the dominant structural drivers of adsorption energy, with conductivity (r = 0.54) and aspect ratio (r = 0.49) playing secondary roles. The inclusion of a binary reconstruction descriptor materially improved predictive performance, confirming that atomic-level surface configuration carries information not redundant with bulk morphology.
Overall, these findings highlight the potential of machine learning to bridge experimental characterization and computational analysis in nanotechnology, providing a scalable and efficient framework for automated nano-characterization. In addition, it was found that non-linear and ensemble models outperform linear baselines, consistent with the collinearity and partial non-linearity observed in the structural descriptors. Furthermore, the framework’s modality-agnostic feature schema positions it as a transferable template for adjacent characterization problems. Future work should focus on incorporating more complex datasets, expanding to additional material families, including higher-dimensional structural representations and real-time experimental data, to further enhance predictive accuracy and applicability across a wider range of nanomaterials.

CONFLICT OF INTEREST
The authors declare that there is no conflict of interests regarding the publication of this manuscript.

ACKNOWLEDGEMENT
Thanks to the technicians and staff at the University of Misan, College of Science, for their efforts in the collection of data.

CONFLICT OF INTEREST
The authors declare that there is no conflict of interests regarding the publication of this manuscript.