Evaluating Fuel Properties of Strained Polycycloalkanes for High-Performance Sustainable Aviation Fuels

Abstract

Sustainable aviation fuel (SAF) is a drop-in alternative to conventional jet fuels, designed to reduce greenhouse gas (GHG) emissions while requiring minimal infrastructure changes and certification under the American Society for Testing and Materials (ASTM) D7566 standard. This study assesses recently identified high-energy-density (HED) strained polycycloalkanes as SAF candidates. Strain energy (E_se) was calculated using density functional theory (DFT), while operational properties such as boiling point (BP) and flash point (FP) were predicted using support vector regression (SVR) models. The models demonstrated strong predictive performance (R² > 0.96) with mean absolute errors of 6.92 K for BP and 9.58 K for FP, with robustness sensitivity analysis. It is found that approximately 65% of these studied polycycloalkanes fall within the Jet A fuel property boundaries. The polycycloalkanes (C₉–C₁₅) with strain energies below approximately 60 kcal/mol achieve an balance between energy density and ignition safety, aligning with the specifications of Jet A. The majority of structures were dominated by five-membered rings, with a few three- or four-membered rings enhancing favorable trade-offs among BP, FP, and HED. This early pre-screening indicates that moderately strained polycycloalkanes are safe, energy-dense candidates for next-generation sustainable jet fuels and provide a framework for designing high-performance SAFs.

1. Introduction

The rapid growth of the aviation industry is expected to drive a significant increase in global aviation fuel demand, which is projected to more than double by 2050 [1], intensifying both demand and environmental concerns. As a result, it is crucial to develop SAF to reduce the aviation industry’s carbon footprint and to mitigate the impact of GHG emissions. These developments are becoming particularly essential as the aviation industry seeks all viable pathways, including electrification techniques and hydrogen propulsion, as other alternatives to achieve sustainability [2,3,4]. Among the available options, SAF offers a more practical solution, providing an intermediate reduction in carbon emissions of up to 80% over its lifecycle compared to conventional jet fuel [5].

SAF must meet the stringent safety and performance standards set by the ASTM D7566 [3,6]. Currently, seven ASTM-approved pathways for the production of SAF [7] include Fischer-Tropsch synthetic paraffinic kerosene (FT-SPK) and FT-SPK with aromatics (FT-SPK/A), bio-based routes such as hydroprocessed esters and fatty acids (HEFA-SPK) and catalytic hydrothermolysis jet fuel (HC-HEFA-SPK), hydroprocessed fermented sugars to synthetic iso-paraffins (HFS-SIP), alcohol-to-jet synthetic paraffinic kerosene (ATJ-SPK), and fuels derived from the hydrothermal conversion of fatty acids (CHJ) [3,8]. Each of these pathways represents the leading technologies for SAF production, each with distinct feedstocks and processing methods, while all adhering to the rigorous performance and safety standards required for aviation use [2,3,4,9,10]. Blending SAFs with conventional jet fuels—up to 50% by volume—can preserve critical properties such as HED and favorable O-ring swelling, both of which are essential for optimal engine performance [3,9,10,11]. However, incomplete combustion of these blends can result in substantial soot and particulate emissions, highlighting a trade-off between achieving performance standards and minimizing environmental impact [2,9].

To address these challenges, strained polycycloalkanes have emerged as promising HED SAF candidates [2,3,12]. Their unique molecular structures and synthetic accessibility contribute to enhanced energy performance, with the potential to increase both volumetric specific energy and energy density [2]. These properties are critical in aviation due to the limited fuel capacity of Jets, where a higher energy content directly translates to greater payload capacity and extended flight range [2,4]. In particular, alkanes (paraffins) are saturated hydrocarbons consisting of carbon–carbon (C–C) and carbon-hydrogen (C–H) single bonds, which confer greater chemical bond energy compared to unsaturated hydrocarbons. Alkanes burn more cleanly than alkenes, producing less soot and particulate emissions [13]. This results in lower engine deposits and reduced maintenance requirements for aviation engines, enhancing overall efficiency and performance [14].

The inherent stability of alkanes supports energy storage and combustion safety. In addition to increasing energy density [2,3,11,12,15,16,17,18], structural strain within polycyclic alkanes can also alter fuel properties that are critical for efficiency and operational safety. For example, JP-10—exo-tetrahydrodicyclopentadiene (C₁₀H₁₆), an HED synthetic strained polycycloalkane fuel primarily used in military applications—exhibits superior combustion characteristics and operational performance compared to conventional Jet A aviation fuel in commercial use [2,3,18]. Optimizing strained hydrocarbons for aviation applications requires a careful design to balance energy output with operational safety. However, despite their potential to be HED fuels, the safety profiles of these polycycloalkane candidates remain underexplored. To ensure their viability as SAF, these compounds must also satisfy ASTM D7566/D4054 certification standards [6,19], which set rigorous thresholds for key physicochemical and operational properties, including specific energy (SE ≥ 42.80 MJ/kg), density (0.76–0.84 g/mL at 288.15 K), freezing point (FrP ≤ 226.15), (FP ≥ 311.15 K), and (BP ~ 433.15–573.15 K) [12]. In practice, consistently meeting the stringent FrP, FP, and density requirements, along with compatibility tests such as O-ring swell, remains particularly challenging for many candidates.

In one of our previous studies, we identified several HED-strained hydrocarbons as promising candidates for aviation fuel applications [2]. In this work, we employed DFT calculations [20] to analyze the molecular structures and E_se of a selected group of polycycloalkane SAF candidates. We then employed machine learning (ML) techniques [21,22] and the group contribution (GC) method [23,24] to predict key fuel properties. Subsequently, we evaluated quantitative structure–property relationships (QSPR), specifically the BP and FP. This computational evaluation supports early-stage pre-screening of SAF candidates and assesses their safety for aviation applications, in accordance with ASTM D7566 [6] and D4045 standards [19]. This comprehensive assessment aims to accelerate SAF development through the optimization of synthetic pathways.

2. Computational and Predictive Methods

2.1. Target Polycycloalkanes and Strain Energy Calculation

Twenty selected strained HED polycycloalkanes, identified as potential SAF candidates through ML [2], are evaluated further for aviation fuel properties. These candidates are listed as entries 1–20 in Table S2 of the Supplementary Materials (SM), with their corresponding structures shown in Figure 1.

E_se in hydrocarbons refers to the extra energy stored in a molecule due to distortions in its bond angles, bond lengths, or torsional angles from their ideal, low-energy conformations. The total E_se of a compound is typically the result of several contributing factors [25], including bond strain, angle strain (also called Baeyer strain), torsional strain, and steric strain (or van der Waals/Prelog strain). Bond strain occurs when chemical bonds are stretched or compressed beyond their optimal lengths. This type of strain is energetically costly and uncommon in stable organic molecules. Molecules generally adopt conformations that minimize bond strain, as the high energy cost associated with it is unfavorable. Angle strain (also called Baeyer strain) is caused by deviations from ideal bond angles (e.g., 109.5° in sp³-hybridized carbon) [26]. While angle strain contributes to the overall energy of a molecule, it is generally less severe than torsional or steric strain. Torsional strain, also known as eclipsing or Pitzer strain, results from eclipsed interactions between adjacent bonds, where atoms or groups are aligned in a way that leads to increased electron repulsion. Torsional strain often contributes significantly to the overall strain energy of a molecule. Steric strain (or van der Waals/Prelog strain) occurs when atoms are forced close together, leading to repulsion between electron clouds [26]. A specific type of steric strain, known as transannular strain, is common in medium-sized rings, where non-bonded atoms across the ring interact unfavorably.

Cyclic hydrocarbons adopt conformations that minimize the total E_se by balancing various types of strain [25]. Since bond strain is particularly destabilizing, molecules typically avoid it by tolerating more manageable forms of strain, namely torsional, steric, and angle strain. Cyclic compounds are classified as small rings with (three- or four-membered), common rings with (five- to seven-membered), medium rings with (eight- to eleven-membered), and large rings with (twelve- or more-membered). Rings of similar sizes tend to have comparable characteristics. Most of the cyclic hydrocarbon fuel candidates exhibit small rings (three- or four-membered), and common rings (five- to seven-membered), which are often grouped into classical rings [25]. Small rings exhibit rigidity and significant strain due to substantial deviations from ideal tetrahedral bond angles, resulting in pronounced angle strain, as well as considerable torsional strain from eclipsed interactions. In contrast, common rings maintain bond angles close to the tetrahedral ideal, with their substituents oriented exclusively outward from the ring structure [25].

Ring strain energy (RSE) refers to the increase in energy or destabilization that occurs when an open-chain (acyclic) molecule forms a ring structure [27]. The E_se of polycycloalkanes can be estimated using two different methods. That is, the ring superposition (RS) method [28,29] and the isodesmic reaction (IR) approach [30]. In the RS method, the E_se of a polycycloalkane is calculated as the sum of the strain energies (E_ring(i)) of the individual rings [28,29].

E_se = ∑n_i E_ring(i)

(1)

where n_i indicates the number of various members of the ring of type i, and E_ring(i) represents the E_se of the ith ring. The E_se of an individual ring (E_ring(i)) is given in Table S1 in the SM. For example, Compound 19 (camphane) appears with two five-membered rings as detailed in Table S2 in the SM. Therefore, the E_se of compound 19 (camphane) is given by

E_se (camphane) = 2 × (6.19) = 12.38 kcal/mol

The IR method [30] to estimate the E_se of a polycycloalkane is determined as

E_se = ΔH_f(C_n+mH_2(n+2m+1)) − ΔH_f(C_nH_{2(n−m + 1)}) − m ΔH_f(_ethane)

(2)

where ΔH_f denotes the standard enthalpies of the compounds involved in the isodesmic reaction. Here, n corresponds to the number of carbon atoms in the polycycloalkane, while m represents the number of ethane units introduced to balance the isodesmic reaction. This IR method compares the target molecule (polycycloalkane) to a set of reference compounds with similar bonding environments [27]. For example,

camphane + 2 ethane → tetradecane

(3)

So that E_se (camphane) = ΔH_{f(tetradecane)} − (ΔH_f(camphane) + 2 ΔH_f(ethane)) = 12.72 kcal/mol.

Details regarding the IR method used are given in Tables S3 and S4 of the SM. For large-strain polycyclic molecules, direct high-accuracy quantum chemical calculations become computationally prohibitive due to the steep scaling with system size [31]. To overcome this limitation, isodesmic reactions exploit systematic error cancelation by comparing molecules to reference compounds with similar bond types, providing a practical and reliable means to compute strain energies. Although the IR method is often described as “bond-conserving,” polycyclic strain energy calculations may involve equations that are formally unbalanced in hydrogen count. This approach assumes that only the differences between similar bonding environments—rather than the absolute energies—are relevant, allowing errors to cancel out between comparable fragments. By carefully matching carbon environments (sp³, strained, bridgehead, etc.), most basis set and correlation errors are minimized, making the IR method a rapid and practical estimation tool.

The enthalpies in Equation (2), used to calculate E_se, were computed using DFT; therefore, the overall approach is referred to as the DFT-based IR method (DFT-IR). These calculations were performed at the B3PW91/aug-cc-pVTZ level of theory, following the geometry optimization. The B3PW91 DFT functional was selected based on its demonstrated reliability and consistency in predicting thermochemical properties across diverse hydrocarbon systems. A recent benchmark study by Dutra and Custodio [32], which evaluated 24 DFT functionals, identified B3PW91 as one of the most robust performers under varied computational conditions. Additionally, the literature reports show that B3PW91 yields enthalpy predictions with a mean absolute error of approximately 2 kcal/mol [33], making it suitable for evaluating strain energies in hydrocarbon fuels. The aug-cc-pVTZ basis set complements this choice by offering triple-zeta quality with diffuse functions, which is essential for accurately describing strained and delocalized electronic environments. As a result, B3PW91/aug-cc-pVTZ offers a well-balanced approach that strikes a balance between computational efficiency and predictive accuracy, particularly for systems involving ring strain and non-classical bonding. The absence of imaginary frequencies in the optimized structures indicates that these structures correspond to true energy minima, indicating they are stable [15,16,18]. All DFT calculations are performed using the Gaussian 16 program [34].

2.2. Machine Learning for Fuel Property Prediction

ML models were developed to evaluate two key thermophysical properties essential for the fuel candidates: BP and FP. The BP is defined as the temperature at which a pure liquid transitions to the gas phase under 1 atm pressure, while FP is the lowest temperature at which a liquid emits sufficient vapor to ignite in the presence of a flame or spark [35]. Although BP and FP are experimentally measurable, predictive modeling offers significant value by enabling rapid screening of strained polycyclic molecules, where evaluation of the ASTM standard can be costly and time-consuming. Establishing these structure–property relationships, therefore, supports early-stage SAF design by prioritizing candidates with favorable volatility and safety profiles. Rücker and Rücker [36] refined structure–BP correlations for saturated hydrocarbons using high-quality data and topological indices, noting errors in earlier cycloalkane values. However, their study did not examine strained polycycloalkanes, which will be addressed in the present work. The property dataset consisted of 85 cycloalkanes for BP and 80 cycloalkanes for FP, where the latter included a combination of experimental and computational values. Data were collected exclusively for cycloalkane structures from peer-reviewed literature [11,37] and authoritative chemical databases, including the CRC Handbook [38], the NIST Chemistry WebBook [39], and the Chemical Properties Handbook [40].

Molecular descriptors were generated from Simplified Molecular Input Line Entry System (SMILES) representations of the compounds using RDKit [41] and Mordred [42], resulting in more than 900 initial descriptors. A systematic preprocessing workflow was applied to ensure descriptor quality and reduce redundancy following the same steps as in recent work in fuel [43]. First, non-numeric and constant (low-variance) descriptors were removed. Next, descriptors exhibiting strong pairwise correlations (Pearson’s |r| > 0.90) were excluded to avoid multicollinearity. Moreover, feature relevance was assessed using Spearman correlation [44] and Recursive Feature Elimination (RFE) with a Random Forest Regressor (RFR) [43]. The final 10 most informative descriptors for each property (Table S6, SM) were identified and standardized using Z-score normalization [45].

Model development was carried out using SVR, which had been previously benchmarked in our earlier work [2] and consistently outperformed alternatives such as Random Forest (RF) and k-Nearest Neighbors (kNN) for related fuel property predictions. The preference for SVR is also supported by recent reviews that identify it as one of the most widely applied algorithms in QSPR modeling of fuel ignition properties [46]. Its ability to model both linear and nonlinear relationships, while maintaining good generalization on small datasets, makes it particularly suitable for this application. Datasets were partitioned into 80:10:10 splits for training, validation, and testing, respectively. Hyperparameters (regularization constant C, epsilon tolerance ε, and kernel type) were optimized through GridSearchCV [47]. The linear kernel consistently emerged as the most effective choice for both BP and FP prediction, as detailed in optimized parameters (Table S7, SM). Final models were retrained on the combined training and validation sets using the optimal hyperparameters, and predictive performance was assessed on the independent test sets.

Model validation was performed using three repeated 5-fold cross-validations in training data [48], while model sensitivity was assessed through bootstrap analysis [49]. To determine the applicability of the model to a new molecule, the classical Joback and Reid GC models [23,24] were also implemented and evaluated for the candidate molecules (Table S10, SM). All modeling and visualization were performed in Python 3 (Jupyter) [22,50], utilizing a standard data science stack: Pandas (1.5.3) and NumPy (1.24.4) for data handling, scikit-learn (1.3.2) for feature selection, model training, and validation, SciPy (1.10.1) for statistical analysis, and Matplotlib (3.7.1) and Seaborn (0.12.2) [50] for visualization.

3. Results and Discussion

3.1. Strain Energy in the Polycycloalkanes

E_se plays a crucial role in determining the stability and energy output of polycycloalkanes, which are an important structural characteristic in many HED fuels. The E_se of these target polycycloalkanes is estimated using the RS and the DFT-based IR method (DFT-IR), as expressed in Equations (1) and (2).

Table 1. Validation of calculated E_se against reference values using RS and DFT-IR methods (Kcal/mol).

Polycycloalkanes	RS	DFT-IR	Ref	∆ (RS—Ref) %	∆ (DFT-IR—Ref) %
7 (quadricyclane)	93.48	78.59	78.70 *	18.78	0.14
10 (prismane)	134.11	126.69	128.00 #	4.77	1.02

* Experimental value from Hall et al. using the heat of combustion method [51]. ^# Estimated using ab initio SCF at 3-21G level by Politzer et al. [52].

compares the RS and DFT-IR results with the available reference E_se for compounds 7 (quadricyclane) and 10 (prismane), as reliable data for the studied compounds are limited.

The RS method estimates the total E_se of a molecule using a simple additive approximation; hereby, the overall strain is considered as the sum of the strain energies of its individual ring components (e.g., cyclopropane, cyclobutane) [29,53]. While this approach represents a proper first-order estimate, it overlooks several essential factors that influence strain distribution in polycycloalkanes and therefore often fails to provide accurate predictions. First, the RS method assumes that the strain contributions from each ring are independent. However, in fused and bridged ring systems, the rings are structurally interconnected, and interactions between them can redistribute E_se—often leading to strain relaxation. These inter-ring interactions are inherently neglected in the RS framework. Second, the RS method assumes rigid molecular geometries and does not account for electronic delocalization or steric effects, both of which can alleviate strain. In reality, these factors can significantly reduce the total E_se but are not captured in the additive RS model. Third, the RS method lacks the capacity to accommodate geometric flexibility. It treats the rings as fixed units, whereas real molecules can adopt more favorable conformations through bond angle and torsional relaxation. This geometric adaptability is crucial for determining the true minimum-energy structure, especially in complex polycyclic systems.

In contrast, DFT calculations inherently account for these effects. DFT enables complete molecular geometry optimizations, allowing the system to relax into its lowest-energy conformation while accounting for both electronic structure and steric interactions. This quantum mechanical approach is essential in multi-ring systems, where E_se does not scale linearly with the number of rings due to strain delocalization across the molecular framework.

The advantage of the DFT-IR method over RS becomes clear when comparing calculated E_se to experimental or high-level computational benchmarks. For example, the DFT-IR calculated E_se for compound 7 (quadricyclane, C₇H₈, tetracyclo [3.2.0.0²,⁷.0⁴,⁶]heptane) deviates from the experimental value by only 0.14%, while for compound 10 (prismane, C₆;H₆, tetracyclo[2.2.0.0²,⁶.0³,⁵]hexane), the deviation is just 1.02% as detailed in Table 1. In contrast, the RS method exhibits significantly large and inconsistent deviations: 18.78% for 7 (quadricyclane) and 4.77% for 10 (prismane). These discrepancies emphasize the limitations of the RS method and highlight the accuracy and reliability of quantum mechanical approaches such as DFT-IR. Furthermore, comparisons with prior studies also reveal substantial mismatches between RS-derived and experimentally determined strain energies, reinforcing the need for more sophisticated methods in E_se evaluation [28,29].

For HED fuel applications, it is ideal that the polycycloalkanes have moderate to high positive E_se. High E_se indicates the molecule stores extra potential energy in its bonds due to ring strain. If the E_se is too high, issues such as synthesis, stability, and safety may outweigh the benefit [54]. Upon combustion, this strain is released in addition to the standard chemical enthalpy, boosting the net heat of combustion per unit mass or volume. As a result, the optimal range of E_se for HED fuels can be assumed between 10 and 60 kcal/mol, which is within the vicinity of compound 6 (exo-THDCPD) and 7 (quadricyclane). Compound 6 (exo-THDCPD) is already used as a military fuel, JP-10, while 7 (quadricyclane) is extensively investigated as a top candidate for HED fuel [55]. Figure 1 presents the molecular structures of the twenty studied compounds and compares E_se values for selected polycycloalkanes determined via RS and DFT-IR, with both methods showing consistent trends. The E_se of all compounds is under 100 kcal/mol except for compound 10 (prismane), which is estimated at 126.69 kcal/mol.

These observed E_se in Figure 1 are due to the ring structures of the compounds. The compounds with relatively high E_se actually hold the strained three- or four-membered cyclic structures. As discovered by Dragojlovic [25], the strain energy per –CH₂ group is highest in smaller rings, with 9.17 kcal/mol for cyclopropane (n = 3) and 6.58 kcal/mol for cyclobutane (n = 4). In polycyclic systems, however, additional factors such as ring fusion, bridging, and angular strain complicate the E_se landscape. For instance, high strain of compound 7 (quadricyclane) arises from its highly constrained geometry, comprising five rings with two three, two five, and one four-membered ring. In contrast, compound 16 (spirodecane) shows the lowest E_se at 5.59 kcal/mol among these compounds. It is noted that compound 6 (exo-THDCPD, i.e., JP-10) has a moderate E_se of 12.57 kcal/mol. The calculated E_se of the compounds is given in Table S2 of the SM.

3.2. Boiling Point and Flash Point Estimate

3.2.1. Property Dataset and ML Model Performance

An overview of the fuel property dataset and the molecular descriptors used to develop ML models is provided. Specifically, this section focuses on designing a model and validation to predict the BP and FP of the polycycloalkanes listed in Figure 1. The distribution of carbon numbers and ring sizes within the property dataset is shown in Figure 2. It includes a structurally diverse collection of monocyclic and polycyclic hydrocarbons, with carbon numbers ranging from C₃ to C₂₁ for BP data and C₃ to C₂₀ for FP data. The dataset covers ring sizes from three- to twelve-membered structures, encompassing a wide range of cyclic frameworks. Notably, within the carbon range typical of conventional jet fuels (C₇–C₁₈), the BP and FP distributions peak at total carbon numbers of C₈ and C₉, respectively, which are well represented in the dataset [3].

As shown in Figure 2, five- and six-membered rings are the most prevalent in both BP and FP datasets because they offer an optimal balance between thermodynamic stability and synthetic accessibility. Their conformations (e.g., envelope for cyclopentane, chair for cyclohexane) minimize both angle and torsional strain [25,36], making them energetically favorable. Additionally, they are readily formed in many synthetic and biosynthetic pathways and can accommodate diverse substitutions, thereby increasing their occurrence. In contrast, smaller rings suffer from high strain, while larger ones are less accessible and often less stable.

Figure 3 highlights the importance scores of the 10 selected molecular descriptors used in training the ML models. For BP prediction (Figure 3a), the number of carbon atoms (NC) was identified as the most influential descriptor, with an importance score of approximately 0.87 out of 1. The BertzCT graph-based descriptor appears as the second-most influential in predicting BP, which assesses the complexity of the molecules [56]. In contrast, FP prediction (Figure 3b) was influenced by the information content descriptor (Ipc) [57], which achieved an importance score of approximately 0.60. As demonstrated in SM (Figure S1), NC and Ipc are strongly correlated because both descriptors capture molecular size. In the preprocessing stage, NC was excluded from the FP model, with Ipc retained as the representative size-related variable. While feature selection in ML models is mainly determined by statistical correlations with the target property to improve predictive accuracy, it does not necessarily prioritize descriptors with clear physical significance [58]. It may instead choose less chemically informative variables. Therefore, although NC is more chemically intuitive, Ipc preferred for the FP model to strike a balance between statistical rigor and predictive performance. To maintain chemical interpretability and ensure consistency across both the BP and FP models, NC was still emphasized in our discussion of results.

Aside from size-related descriptors, other features also played a significant role in FP prediction, reflecting the interaction of topological and electronic factors. Including VE1_A, BalabanJ, and GATS1Z emerged as key predictors (Table S6, SM). The VE1_A descriptor, derived from graph theory, describes molecular connectivity and tends to increase with branching, which may affect ignition behavior [56]. The BalabanJ topological index measures molecular complexity by considering both atomic connectivity and distances [2,56], while the GATS1Z descriptor captures autocorrelation of atomic properties (such as atomic number) across a specified topological distance, emphasizing the role of electronic distribution [56]. Collectively, these descriptors improved predictive power compared to models relying only on size descriptors. Their combination demonstrates how ML-based selection of complementary features can produce strong predictive models, while also highlighting the importance of balancing chemical interpretability with statistical performance in developing property prediction models.

Figure 4 shows the relationship between the descriptor NC (number of carbon atoms in a single ring) and the BP dataset, which contains only experimental values. For monocyclic alkanes with three- to twelve-membered rings, the BP increases approximately linearly with ring size. The experimental BP value for cycloundecane (eleven-membered ring) in the property dataset of the NIST Chemistry WebBook [39] shows a slight deviation from this trend. For consistency in the analysis, a more recent value (491.55 K) from Guidechem [59] was adopted. This trend reflects the influence of increasing molecular mass and van der Waals surface area, which enhances intermolecular dispersion forces and raises the BP. Smaller rings such as cyclopropane (C₃H₆) and cyclobutane (C₄H₈) have significantly lower BPs, while larger rings exhibit progressively higher values due to their greater surface area and stronger cohesive forces [60].

The predictive performance of the BP and FP using the SVR models was evaluated using three standard metrics: the coefficient of determination (R²), mean absolute error (MAE), and root mean squared error (RMSE), which are similar to our previous study [2], and details are given in the SM. Figure 5 reports the model performance for BP and FP across all datasets. In Figure 5a, the model shows a consistent and high R² value of 0.99 across the training, test, and validation sets for BP prediction, indicating an excellent linear correlation between the predicted and reported values in the property dataset. For instance, the test set yielded an MAE of 6.92 K and an RMSE of 8.46 K, suggesting the model accurately predicts BP with minimal deviation. Even in the validation set, which serves as a test of model reliability, the error metrics remain low (MAE = 5.04 K, RMSE = 6.21 K), highlighting the model’s ability to generalize well to unseen data. While a few minor outliers can be observed (e.g., particularly at higher BP values >490 K), they do not significantly affect the overall predictive performance, and the majority of data points fall close to the parity line (y = x).

Similarly, in Figure 5b, the FP prediction model also exhibits strong performance, albeit with slightly lower R² values compared to the BP model. The training, test, and validation sets show very similar R² values of 0.98, 0.96, and 0.98, respectively, indicating consistent model performance. The test set in particular demonstrates high accuracy with an MAE of 9.58 K and an RMSE of 12.50 K, supporting the model’s strength in FP estimation. However, a few deviations are noticeable in the validation set. Despite these discrepancies, the majority of predictions remain closely aligned with the reported values. Hence, the strong agreement between predicted and reported values in both BP and FP models, particularly with consistently high R² values and low error values across all datasets, demonstrates the efficacy of the SVR approach.

To ensure the robustness and real-world applicability of the developed machine learning models, multiple validation strategies were employed. Specifically, three repeated 5-fold cross-validations [48] were conducted on the training dataset, providing reliable estimates of model generalizability. The models demonstrated strong predictive performance for both BP $(MAE=5.65\pm 1.70\mathrm{K},RMSE=8.08\pm 2.86\mathrm{K},R^2=0.98\pm 0.01)$ and FP $(MAE=8.28\pm 2.41\mathrm{K},RMSE=11.05\pm 4.01\mathrm{K},R^2=0.94\pm 0.05)$ as summarized in SM (Table S8). Because the target property dataset is relatively small, we further assessed model stability and confidence intervals through a sensitivity analysis using 100 bootstrap iterations [49]. The bootstrap analysis further confirmed the stability of the models, with the BP model achieving a mean R² of 0.99 (σ = 0.01) and the FP model achieving a mean R² of 0.95 (σ = 0.02). Consistently low MAE and RMSE values across bootstrap samples indicate reliable predictive accuracy and minimal variability, as detailed in SM (Table S9). Slightly higher variation observed for FP can be attributed to the inclusion of both experimental and computational data in the property dataset.

Even with multiple validation strategies, the risk of overfitting or underfitting cannot be entirely eliminated, which remains a key challenge in ML models when applied to new data [61]. In this study, the dataset is relatively small and restricted to strained cycloalkanes only; therefore, while reliability was assessed by comparing ML predictions with values obtained from the GC method and available experimental data in SM (Table S10), a formal Applicability Domain (AD) analysis (e.g., Williams plot) was not performed. Given the limited chemical space considered here, such an analysis would be more meaningful for larger and more diverse datasets [62]. Future studies involving broader chemical datasets will likely incorporate AD approaches. The close agreement across these comparisons supports the robustness of the ML models within the present domain and highlights their practical relevance. In addition to predictive accuracy, the ML models provide rapid estimations and reveal meaningful structure–property relationships, thereby enhancing both efficiency and interpretability in molecular property prediction.

3.2.2. Structure-Property Trends in Polycycloalkanes

To assess the thermal properties of aviation fuels and evaluate the structural influence of the polycycloalkanes listed in Figure 1, trained models were applied to predict their BP and FP, as summarized in Table S10 of the SM. Figure 6 shows the BP/FP relationship, which follows a well-established linear trend [24]. The green boundary in Figure 6 marks the polycycloalkanes that meet both BP and FP requirements. Approximately 65% of the compounds fall within the operational BP/FP range for Jet A (BP: 433.15–573.15 K; FP ≥ 311.15 K) [12,35]. Thirteen polycycloalkanes (1–6, 9, 13, 15–17, 19, and 20) meet these criteria, supporting their strong potential as candidates for aviation fuel design. Notably, compound 6 (JP-10), a well-established military fuel, is also identified by these criteria as a promising candidate, thereby validating the effectiveness of the method. Since both boiling point (BP, reflecting volatility) and flash point (FP, reflecting ignition characteristics) influence fuel safety, a balanced BP/FP profile is critical. Excessive volatility (low BP) can lead to evaporative losses at altitude, while insufficient volatility (high BP) can impair combustion efficiency [63].

A detailed analysis of the molecular structure within the defined boundary in Figure 6 was conducted to evaluate its influence on the studied properties. Figure 7 presents the structures of polycycloalkanes that satisfy both BP and FP requirements, together with the corresponding predictions from the GC and ML approaches. The strong agreement between GC and ML validates the predicted thermal properties. Most candidates feature a dominant five-membered ring structure, often accompanied by a few smaller three- or four-membered rings. For example, compound 2 (THTCPD; NC = 15, BP = 558.68 K, FP = 422.29 K), which contains five five-membered rings, exhibits the highest BP and FP in the studied set. These results indicate that the number of carbon (NC)—a proxy for molecular size—and the presence of ring structures are primary factors influencing BP and FP. Furthermore, Figure S2 in the SM in the illustrates a nearly linear relationship between BP/FP and NC, which likely arises from the increasing contribution of van der Waals volume with molecular size and associated packing effects, as noted in a prior study [64]. Volatility and ignition performance are observed in the C₉–C₁₅ range, consistent with previous reports that cycloalkanes in this window generally meet jet fuel specifications [65].

In addition to carbon count, E_se may also modulate thermal behavior. The majority of the candidates within the boundary exhibit moderate strain (E_se < 60 kcal/mol), which appears beneficial for balancing stability and volatility. For instance, compound 5 (tetracyclononane; NC = 9, BP = 439.75 K, FP = 323.27 K, E_se = 63.46 kcal/mol) displays higher BP and FP than compound 18 (5-ethylnorbornane; NC = 9, BP = 424.07 K, FP = 307.30 K, E_se = 12.13 kcal/mol), despite identical NC values, placing 5 inside the boundary and 18 outside. This indicates that constrained topologies with moderate strain can enhance intermolecular interactions, elevating volatility and ignition thresholds. In contrast, excessive strain (>60 kcal/mol), typically arising from multiple small rings, reduces both BP and FP by limiting effective NC and molecular flexibility. Such effects are evident in compound 7 (quadricyclane; NC = 7, BP = 386.89 K, FP = 272.75 K, E_se = 78.59 kcal/mol), which fails to meet BP/FP requirements despite satisfying other fuel criteria and being widely researched as a potential aviation fuel [51,55]. Thus, a balance between molecular size, ring topology, and moderate strain emerges as the key design principle for identifying viable polycycloalkane fuel candidates, as determined by this study. These findings indicate the dual role of E_se in polycyclic fuel performance. Moderate strain improves molecular rigidity and elevates BP and FP, while excessive strain compromises both properties. Polycyclic structures dominated by five-membered rings, with controlled incorporation of three- and four-membered units, strike an effective balance between energy density and operational safety (BP&FP), making them strong candidates for SAF design. Figure 6 and Figure 7 provide a structural rationale for targeted molecular screening.

While this study emphasizes the influence of E_se on volatility, ignition, and stability, it is essential to acknowledge that real-world fuel performance also depends on additional parameters beyond the scope of this work. The cyclic and strained nature of the candidate molecules is particularly relevant, as literature reports that highly strained polycyclic structures can alter reaction pathways, rupture C–C bonds, and thereby promote the formation of polycyclic aromatic hydrocarbons (PAHs) and increase soot production during combustion [66]. For example, while compounds like 1 (pentacyclododecane) and 2 (THTCPD) exhibit promising energy, BP, and FP profiles, their complex polycyclic structures may act as precursors to PAHs, resulting in undesirable sooting behavior in engines. Similarly, strained cyclic motifs, while beneficial for energy density, can influence ignition kinetics and flame propagation in ways that are not captured by the current screening. Incorporating these combustion considerations into future ML- and DFT-based frameworks will be essential for a more accurate assessment of fuel suitability. DFT studies of high-potential molecules, such as 6 (JP-10), could provide additional insight into internal bond strain and reactivity, complementing empirical observations and guiding the design of SAFs.

In addition to combustion performance, cold-flow properties, such as FrP and viscosity, are critical determinants of aviation fuel suitability [3,12,35]. Molecular compactness, ring size, and rigidity strongly affect freezing behavior, potentially disqualifying otherwise promising candidates [18]. For instance, while compound 6 (JP-10, exo-THDCPD) remains suitable with a FrP of 194.15 K [67], its isomer endo-THDCPD solidifies at 350.15 K, making it unsuitable [18,67] for aviation fuel. Likewise, compact molecules such as 5 (tetracyclononane) may exhibit elevated FrP, which can limit performance in extreme cold, despite meeting volatility and ignition criteria. These trade-offs support the importance of integrating cold-flow analysis into future screening frameworks, using computational modeling and ML methods combined with quantum chemistry to ensure that candidate molecules meet the full spectrum of operational requirements for SAFs.

4. Conclusions

This study demonstrates the utility of combining DFT with ML to study fuel properties of strained polycycloalkanes as potential HED SAF. Machine learning using SVR models provided fast and reasonably accurate predictions of BP and FP, enabling the preliminary screening of candidate molecules. Thirteen (approximately 65%) of the compounds, including JP-10, met key volatility and ignition criteria. Although E_se showed no direct correlation with fuel suitability, the limited dataset suggests that promising candidates may cluster below approximately 60 kcal/mol in E_se, indicating a potential balance between energy content and stability that warrants further investigation. The identified C₉–C₁₅ compounds, predominantly five-membered ring structures, fall within the typical Jet A range for energy, BP, and FP, suggesting practical potential in aviation fuel. However, these results represent only an initial computational screening, and the actual viability of these candidates as SAFs will depend on a thorough evaluation of their combustion behavior and cold-flow properties. Overall, this work illustrates how integrating ML-based screening with quantum-level analysis can provide a cost-effective and scientifically grounded approach for the early-stage assessment of SAF candidates.