Local Quadrupole Ellipticity as Predictor of Anion‐Affinity in Nanographenes Authors: Omkar Charapale, Sergio Posada‐Pérez, Albert Poater, Miquel Solà Published: 2026 ABSTRACT ABSTRACTAnion binding to nanographenes is governed by noncovalent interactions, particularly anion–π interactions in electron‐deficient aromatic regions and CH‐‐‐anion hydrogen bonding in electron‐rich domains. These interactions are primarily driven by electrostatic effects, with the quadrupole moment of the aromatic system playing a central role in determining the strength and directionality of anion–π binding. The perpendicular component of the quadrupole moment (Qzz) correlates with binding energies for both anion–π and CH‐‐‐anion interactions, though polycyclic systems present challenges due to competing interaction modes. In this study, we investigate the role of the local quadrupole moment in anion binding across 171 Cl−–aromatic complexes, comparing various descriptors including aromaticity indices, Fukui functions, and electron density at ring critical points. We find that electrostatic descriptors, particularly the local quadrupole moment, provide a more consistent and robust explanation for binding energies than conventional descriptors. Specifically, two geometric descriptors derived from the local quadrupole moment—the scale factor (SRmax) and the ellipticity (ec′)—show good correlation with binding strength, with SRmax reflecting π‐acidity and ellipticity quantifying charge distribution anisotropy. These descriptors are validated across fluorinated naphthalenes and nanographenes, demonstrating their general applicability. Regression models based on SRmax and ec′ effectively predict binding energies, with enhanced accuracy when combined with polarization‐dependent penalty functions, especially for larger nanographene systems. While the predictive model is still somewhat constrained by polarization effects, its simplicity, robustness, and transferability across a wide range of systems offer distinct advantages over more complex, multilayered machine learning models. These results underscore the critical role of quadrupole moment anisotropy in anion–π interactions and offer a practical framework for predicting anion binding affinities and designing π‐acidic receptors. Introduction Anion binding [1, 2] in nanographenes is governed by a distinctive class of noncovalent interactions that may arise through multiple binding modes, including anion–π interactions [3] and ionic hydrogen bonding (C–H···X−, where X− denotes an anion) [4, 5]. These interactions depend on the local electronic structure of individual aromatic rings embedded within extended π‐conjugated frameworks. While nanographenes are generally π‐electron rich and thus well known for favorable cation–π interactions [6, 7, 8, 9, 10, 11], their interaction with anions has long remained largely unexplored, due to the intuitive expectation of electrostatic repulsion between negatively charged anions and aromatic π‐clouds [3, 12]. Early theoretical studies on benzene supported this view, predicting repulsive anion–π interactions and reinforcing the counterintuitive nature of anion binding above aromatic rings [13, 14, 15, 16]. However, subsequent investigations revealed that this picture is incomplete. Frontera and co‐workers demonstrated that, for benzene–chloride complexes, the interaction energy is nearly negligible due to a compensating balance between repulsive electrostatics and attractive ion‐induced polarization [17]. This “compensating effect” has subsequently become a recurring theme in anion–π chemistry, highlighting the subtle balance between electrostatics, induction, and dispersion contributions. Interest in anion–π interactions intensified following both experimental and theoretical studies that elucidated their physical origin [13, 15, 18, 19, 20]. Ab initio calculations established that anion binding is enhanced in electron‐deficient π‐systems, where depletion of electron density above the ring center leads to favorable electrostatic interactions [14, 17, 21, 22, 23]. Such depletion can be achieved by protonation or by introducing electron‐withdrawing substituents, resulting in positive regions of molecular electrostatic potential (MESP) [24, 25] above the aromatic plane [4, 7, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37]. Experimental observations of anions trapped within deep aromatic cavities of supramolecular hosts further confirmed the existence and relevance of anion–π interactions [12]. From a theoretical perspective, detailed computational studies have shown that anion–π interactions arise primarily from a combination of electrostatic [18, 38, 39, 40, 41] and induction effects [41, 42, 43, 44]. The electrostatic component is often correlated with the quadrupole moment component Qzz [13, 17, 19, 20, 30, 39, 45], while the induction contribution scales with molecular polarizability [42, 45, 46, 47, 48, 49, 50]. The molecular quadrupole moment for benzene and hexafluorobenzene is shown in Figure 1. Mascal et al. [26, 38] demonstrated a clear relationship between halide binding energies and the MESP above the ring center for systems such as 1,3,5‐triazine and trifluoro‐1,3,5‐triazine, where positive electrostatic potential facilitates anion binding. In this context, the quadrupole moment naturally emerges as a low‐order descriptor in the multipole expansion of the electrostatic potential [51, 52]. Visualization of quadrupole moment tensor (Q) for (A) benzene (left) and (B) hexafluorobenzene (right). The top figures represent the surface of the quadrupole moment tensors which are visualized with MESP values, red (blue) color represents negative (positive) tensor value, while bottom ones are quadric surfaces of Q (see Supporting Information). For benzene, the Qzz is negative and corresponding quadric is a hyperboloid of one sheet whereas for hexafluorobenzene, the Qzz is positive and corresponding quadric is a hyperboloid of two sheets. The convex regions of hyperbolas in both hyperboloids correspond to blue regions that is positive tensor value or depletion of charge. Nevertheless, the relationship between quadrupole moment, aromaticity, and anion affinity is far from straightforward. While positive Qzz values generally indicate π‐acidity [37, 53, 54, 55, 56, 57, 58, 59] and favor anion–π interactions [3, 13, 14, 39, 48], numerous exceptions have been reported. Aromatic rings with negligible or even negative quadrupole moments, such as 1,3,5‐trifluorobenzene or s‐triazine, can still exhibit favorable anion–π interactions due to dominant induction effects [41, 42, 43, 44]. In contrast, systems with strong electron‐withdrawing substituents may display negative quadrupole moments yet bind anions through additional stabilizing interactions, including non‐classical hydrogen bonding [60]. These observations indicate that global descriptors such as aromaticity indices or total quadrupole moments alone are insufficient to rationalize anion affinity across diverse π systems. The importance of polarization and compensating effects [17] is further illustrated in extended and substituted aromatic systems. For example, Garau, Frontera, and co‐workers reported attractive anion binding to triphenylene derivatives despite their permanent negative quadrupole moments, attributing the interaction to large molecular polarizability [47]. Similarly, in thiocyanuric acid derivatives, systematic substitution leads to decreasing quadrupole moments but increasing polarizability, while the chloride binding energy (BE) remains nearly constant—again reflecting a balance between opposing contributions [50]. Recently, Liu et al. have introduced a novel descriptor, the “orbital electrostatic energy” (OEE), to quantify the electrostatic effects between an ion and a π‐system [61, 62]. Their work demonstrated a strong correlation between OEE and ion‐binding energy for arenes, capturing both electrostatic and polarization effects. This approach provides further insight into the delicate balance between these two competing contributions, offering a more precise way to evaluate binding interactions in aromatic systems. A particularly influential reinterpretation of anion–π interactions was provided by Wheeler and Houk, who argued that anion binding to substituted benzenes is dominated by electrostatic interactions with substituent dipoles rather than with the π system itself [63, 64]. According to this view, the intrinsic anion–phenyl interaction remains repulsive, with favorable binding arising from additive electrostatic contributions of the substituents. While this interpretation successfully rationalizes trends in fluorinated benzenes, it has been challenged by subsequent studies suggesting that polarization, charge transfer [18, 19], and multicenter interactions [65] also contribute significantly. This argument was supported by energy decomposition analyses carried out by Thirman and Head‐Gordon [66], which indicate that upon fluorination, major changes are observed in frozen or steric (electrostatic + Pauli) component while polarization and dispersion remain constant. Also, Pauli repulsion dominates at short range, while electrostatics, induction, and dispersion govern stabilization beyond this regime. Spatially resolved investigations have further clarified the anisotropic nature of anion–π interactions [47, 67, 68]. Estarellas et al. [67] demonstrated that lateral displacement of an anion within a plane parallel to the aromatic ring results in relatively minor energy changes, whereas displacement along the axis perpendicular to the ring leads to pronounced reductions in binding energy. Moha et al. [68] extended this analysis by mapping the binding energy surface for anions above fluorinated and non‐fluorinated benzenes. Their results revealed a well‐defined energy minimum at the ring center for hexafluorobenzene, while less fluorinated systems exhibit flatter or saddle‐shaped energy landscapes, and benzene displays a repulsive maximum at the center. These findings underscore the importance of local electronic environments rather than global molecular properties. Despite extensive progress, a key challenge remains: the absence of a transferable, local descriptor capable of capturing anion affinity within extended nanographenes, where individual rings experience distinct electronic environments due to edge effects, heteroatom incorporation, and conjugation topology. In such systems, global aromaticity indices and molecular quadrupole moments tend to obscure chemically meaningful local variations. In this work, we address this challenge by introducing the local quadrupole moment tensor of individual six‐membered rings in nanographenes as a descriptor of anion binding. We employ a quadric‐surface representation to visualize the anisotropy of the local quadrupole tensor and relate its eigenvalues to anion binding strength. Using kernel ridge regression and multilinear models, we demonstrate that anion affinity can be accurately predicted from the tensor eigenvalues or from derived geometric parameters such as ellipticity and scale factor. Our results establish the local quadrupole moment as a physically motivated and quantitative descriptor, providing a robust framework for understanding and predicting anion–π interactions in nanographenes in the gas‐phase. Computational Details The charge neutral systems under consideration are coronene family members such as coronene (C24H12), circumcoronene (C54H18), circumcircumcoronene (C96H24) and circumcircumcircumcoronene (C150H30) and fluorarenes especially all possible fluorine substituted benzenes (C6H6‐xFx, 0 ≤ x ≤ 6) and naphthalene (C10H8‐xFx, 0 ≤ x ≤ 8). The nomenclature for isomers of fluorobenzenes and fluoronaphthalenes is given in Tables S3 and S4 of the Supporting Information. Wavefunction analyses were carried out using AIMAll [69] software package, while for Fukui and aromaticity indicator calculations, MultiWFN [70] and ESI‐3D [71] packages were used, respectively. For computation and visualization of local quadrupole moment of symmetrically non‐equivalent rings, the in‐house python program, PyArcuits [11] is employed. This program integrates several open‐source libraries such as ESIpy [72], NetworkX [73], RDKit [74], SciKit‐learn [75], and ASE [76], enabling the construction of various nanographenes, the introduction of doping and defects, and the detailed analysis of local topology and aromatic π‐circuits within the systems. The PyArcuits package is freely available on GitHub. The Cl− adsorbed complexes are also optimized and validated with Hessian analysis at the PBE0‐DH‐D3(BJ)/6‐31 + G(d,p) level of theory. The Cl‐interaction energy is calculated using the following equation, where ECl−−NG is the electronic energy of the Cl− adsorbed at nanographenes, and ECl− and ENG are the energies of isolated species, obtained by performing single‐point energy calculations on optimized geometries at the MN15/6‐311++G(2df,2p) level of theory under gas‐phase conditions. According to Equation (1), positive values denote attractive binding energies between the selected nanographenes and Cl−, whereas negative values indicate repulsive interactions. Further, with the obtained database of Cl− interaction energies of different rings by using DFT, kernel‐ridge and multilinear regression are performed using Robert [77] and PyArcuits. The local quadrupole moment for a fragment or ring R was computed with the formula [78]: here ΘA, μA, and qA are atomic quadrupole moment, atomic dipole moment, and charge of an atom A in a molecule respectively extracted by using AIMAll software package, while rA is the position vector originated from centroid of ring R. The summation goes over all the atoms participating in the fragment or ring R. It is important to note that ΘR is origin‐dependent while ΘA, μA, and qA are atomic partition‐dependent. The visualization of local quadrupole moment tensor (ΘR) is discussed in section 1.2 of Supporting Information. Based on the visualization of local quadrupole moment, two descriptors scale factor (SRmax) and ellipticity (ec′) are defined as follows: where qa and qb are non‐negative eigenvalues of ΘR which form an ellipse upon intersection with molecular plane (see Figure S1) which can be given as, Geometrically, ec′ (unitless) is the ellipticity of this ellipse, while SRmax (unit B1/2, i.e., Buckingham1/2, where in SI units 1 B = 3.34 × 10−40 cm2) is the curvature of circle inscribing largest cyclic quadrilateral in an ellipse that is rhombus formed by endpoints of axes of an ellipse (see Figure S2). These two descriptors are further used as features to train regression models on obtained database of Cl− interaction energies of different rings. Results and Discussion The dataset comprises the binding energies of Cl− interacting with symmetrically non‐equivalent six‐membered rings (6‐MRs) in three classes of fluorinated aromatic systems: 13 fluorobenzenes (C6H6₋xFx, 0 ≤ x ≤ 6), 76 fluoronaphthalenes (C10H8₋xFx, 0 ≤ x ≤ 8), and four nanographenes from the coronene family (C6x2H6x, 2 ≤ x ≤ 5). In total, the dataset contains 171 data points, including 13 from fluorobenzenes, 137 from fluoronaphthalenes, and 21 from coronene‐based systems. The binding motifs can be classified into four distinct interaction types based on the dominant mode of anion binding, as illustrated in Figure 2: (i) C–H‐‐‐Cl− (blue), (ii) C–H‐‐‐Cl−‐‐‐H–C (red), (iii) C–H‐‐‐Cl−‐‐‐H–C′ (golden), and (iv) Cl−‐‐‐π interactions (green). Figure 3 illustrates the interaction energy between Cl− and 13 fluorobenzenes against Qzz component of molecular quadrupole moment. Nomenclature and absolute values of Qzz for fluorobenzenes are given in Table S3 of the Supporting Information. As shown in Figure 3, the strength of Cl−‐‐‐π interactions (green points) increases with increasing Qzz values. In contrast, systems with negative Qzz (red points) exhibit an inverse correlation between interaction strength and Qzz magnitude: the smaller the absolute value of Qzz, the stronger the interaction. This behavior has been previously rationalized by Thirman and Head‐Gordon [63], using angled path of anion approaching fluorobenzene instead of path perpendicular to molecular plane. A blue point in Figure 3 corresponds to 1,3,5‐trifluorobenzene, which has Qzz = 0. This does not follow the trend as the interaction is dominated by polarization as mentioned by Frontera et al. [50] Similar to Qzz, scatter plots correlating binding energies with the aromaticity descriptor MCI and the reactivity descriptor Fukui function (fR+) are presented in Figure S3 and Table S3 of the Supporting Information. In both cases, only weak correlations are observed, indicating that binding energies are not strongly governed by aromaticity or local reactivity descriptors. This behavior is consistent with the predominantly electrostatic nature of the interactions considered here. Furthermore, the difference in the electron density value at the ring critical point (ρ(rCP)) upon fluorination is of the order of 10−3 a.u. (refer to Table S3) which is very negligible for comparison and hence correlation between ρ(rCP) and binding energies (BE) is practically meaningless. Schematic representation of different modes of interactions of nanographenes with a chloride ion (light green circles), three anionic hydrogen bonding interactions and the Cl−–anion‐π interaction. In general, the strength of the ionic hydrogen bonding is higher than that of the anion‐π bond. Plot between molecular quadrupole moment in direction perpendicular to plane of 6‐MR of series benzene to hexafluorobenzene, and their interaction energy with chloride. It is clearly seen that, upon fluorination, the quadrupole moment (Qzz) increases. For systems with negative Qzz, Cl− interacts via ionic hydrogen bonding while systems with positive Qzz, it shows Cl−‐‐‐π interaction. Both patterns show linear regression with Qzz. The nomenclature and absolute values of Qzz for fluorobenzenes are given in Table S3 of the Supporting Information. The values of the scale factor (SRmax) and the ellipticity (ec′) of the local quadrupole moment for the 6‐membered rings (6‐MRs) of fluorobenzenes are provided in the Supporting Information, Table S3. The value of SRmax increases stepwise with fluorination: SRmax ≈ 4.24 and 4.53 B1/2 for benzene and fluorobenzene, respectively; ~4.8, ~5.1, and ~5.4 B1/2 for di‐, tri‐, and tetrafluorinated derivatives; and ≈ 5.65 and 5.92 B1/2 for the penta‐ and hexafluoro derivatives. This trend confirms that the scale factor correlates with the number of electron‐withdrawing groups (here, F atoms) attached to the 6‐MR. Additionally, as shown in Figure S2 of the Supporting Information, the size (curvature) of the orange dotted circle decreases (increases) from A to E with increasing fluorine substitution, indicating that the 6‐MR becomes progressively π‐electron deficient (π‐acidic). From this observation, it can be concluded that SRmax is directly related to the π‐acidity of the 6‐MR. Interestingly, the ec′ values for benzene, 1,3,5‐trifluorobenzene, and hexafluorobenzene are close to zero, and the BE increases with increasing SRmax regardless of the interaction mode. In principle, the ec′ values should be exact zero for abovementioned molecules due to their symmetry. However, the observed numerical error arises from the integration scheme used in the AIM partition, as the grid is not fully resolved. A similar trend in BE is observed for fluorobenzene and 1,2‐difluorobenzene (ec′ ≈ 0.3) as well as for 1,2,3,4‐tetrafluorobenzene and pentafluorobenzene (ec′ ≈ 0.25). Ellipticity ec′ can be interpreted as a measure of the “flatness” or distortion of the ellipse describing the local quadrupole distribution, and it strongly depends on the substitution pattern of fluorine atoms on the ring. For example, in difluorinated derivatives (C6H4F2) with SRmax ≈ 4.8, the ortho‐ and meta‐substituted isomers show ec′ ≈ 0.3 with BEs of ~50 kJ mol−1, whereas the para‐substituted isomer exhibits a significantly larger ellipticity (ec′ ≈ 0.57) accompanied by a higher BE of ~56 kJ mol−1. A similar trend can also be observed for the tri‐ and tetrafluorinated derivatives, where variations in substitution pattern lead to noticeable changes in ec′ and the corresponding interaction strength. The ellipticity parameter can therefore be associated with the anisotropy of charge distribution within the molecular plane. Greater in‐plane anisotropy leads to stronger charge separation across the aromatic ring, which enhances the electrostatic component of the anion–π interaction with the corresponding 6‐MR. From a physical perspective, this indicates that not only the overall π‐electron deficiency of the ring but also the directional distribution of electron density plays a crucial role in determining interaction strength. Taken together with the trends discussed above for SRmax, these observations suggest two complementary factors governing the interaction: (i) the scale factor (SRmax), which reflects the overall π‐acidity of the ring and increases with the number of electron‐withdrawing substituents, and (ii) the ellipticity (ec′), which captures the anisotropic redistribution of charge caused by specific substitution patterns. Consequently, rings possessing both a larger scale factor and higher ellipticity tend to exhibit stronger interactions, independent of the specific interaction mode. To further test the generality of this relationship, we extended our analysis to a broader dataset. For this purpose, a dataset consisting of 171 data points was considered, where each point corresponds to the interaction of Cl− with a specific six‐membered ring (6‐MR). In this way, the calculated binding energies can be directly associated with the local quadrupole moment of the corresponding ring (see Supporting Information, section 2), thereby providing a ring‐resolved description of Cl− anion affinity. The binding energies are then analyzed as a function of the scale factor (SRmax) and ellipticity (ec') of the ellipse formed by the intersection of the local quadrupole moment tensor with the molecular plane (see Supporting Information). This relationship is illustrated in Figure 4A,B, where the dependence of interaction strength on these two descriptors is explicitly shown. The data points are observed to align along several distinct vertical dotted lines, each corresponding to a specific value of the scale factor. These lines reflect the number of electron‐withdrawing substituents attached to the corresponding six‐membered ring (6‐MR). For instance, points located along SRmax ≈ 4.25 correspond to 6‐MRs without fluorine substituents, including Cl− interactions with benzene, naphthalene, and all 6‐MRs in the coronene family. Upon progressive fluorination, the value of SRmax increases, reaching values close to 6 for fully fluorinated rings. In particular, hexafluorobenzene exhibits the highest SRmax approaching ~6 B1/2. Notably, the majority of the green data points corresponding to anion–π interactions are located either at SRmax ≈ 4.25 or at SRmax > 5.35. The former region corresponds to polarization‐dominated anion–π interactions, whereas the latter reflects electrostatically dominated anion–π binding. For SRmax > 5.35, the interactions are exclusively anion–π in nature, while in the intermediate range (4.3 < SRmax < 5.3) the dominant binding motif is ionic hydrogen bonding. An exception is the C–H‐‐‐Cl−‐‐‐H–C′ interaction mode, which represents a bifurcated ionic hydrogen bond involving hydrogens from two fused rings. This kind of interaction cannot be uniquely assigned to the local quadrupole moment of a single isolated 6‐MR, and thus, the corresponding data points (golden symbols) were excluded from further analysis. After this exclusion, the resulting dataset contains 150 data points. Mapping of different motifs of chloride‐nanographene interaction (left) and their binding energy (right) in the plane traced by scale factor (SRmax) and ellipticity (ec′) of an ellipse in the molecular plane of local quadrupole moment tensor (units are B1/2). These plots help to identify dispersion of points depending upon type of interaction and substitution. The dotted vertical lines determine the extent of substituent groups (F) attached to 6‐MR and their corresponding effect on binding energy. BE increases with increase in scale factor (SRmax) and ellipticity (ec′), in general. Golden points correspond to bifurcated anionic hydrogen bonding involving hydrogens from two fused rings and are therefore not assignable to a single 6‐MR quadrupole moment hence these are discarded for further assessment even though they are energetically greater and darker in above plots. As depicted in Figure 4, the BE generally increases with increasing ellipticity and scale factor, independently of the specific interaction mode. This fact confirms the generality explained above. The next step is therefore to quantify the relationship between these descriptors and the binding energy using regression models. Based on this observation, we model the binding energy as an explicit function of the scale factor (SRmax) and ellipticity (ec′). Deriving an exact analytical expression for this relationship is nontrivial; therefore, we employ regression‐based approaches to capture the underlying trends. Several regression models available in the scikit‐learn library were evaluated, including multilinear regression, logistic regression, and kernel ridge regression (KRR). Because the relationship between binding energy and the descriptors is clearly nonlinear, linear regression models provide insufficient accuracy. Therefore, a more flexible model such as kernel ridge regression (KRR) was employed. KRR combines ridge regularization with nonlinear kernel mapping; in this work, a radial basis function (RBF) kernel was employed (performance of different kernel functions is provided in the Supporting Information, Table S1). Details of kernel hyperparameters and optimization procedures are provided in the Supporting Information in section 2. Given the relatively small size of the dataset, model validation was performed using cross‐validation to assess robustness with respect to data partitioning. A four‐fold cross‐validation strategy was adopted, in which approximately 70%–75% of the data were randomly selected for training and the remaining data were used for testing. Multiple independent train–test cycles yielded consistent performance. Across these trials, the mean absolute error (MAE) in the predicted binding energies remained below 8 kJ mol−1, while the root mean square error (RMSE) remained below 12 kJ mol−1. Representative results from four such trials are shown in Figure 5, along with their corresponding MAE and RMSE values. To further assess model performance, the distribution of absolute errors, sorted in ascending order, is shown in Figure 5 (golden curve). This analysis reveals that a small subset of approximately 4–6 data points (large outliers) accounts for the majority of the deviation between predicted and computed energies, with 2–3 points exhibiting errors exceeding 20 kJ mol−1. Notably, these outliers correspond predominantly to polarization‐dominated anion–π interactions, indicating that polarization effects are not fully captured by the present descriptor set. Performance of kernel ridge regression model after trained and tested on 4 randomly selected datasets similar to 4‐fold cross validation. The scatter plots of calculated energy using DFT versus predicted energy (red color) show the accuracy of model with MAE (6–8 kJ·mol−1) and RMSE (8–12 kJ·mol−1), while distribution of absolute error (golden color) infers that 4–6 points (corresponding to polarization dominant anion‐π interaction) are responsible for limiting the accuracy of the model. The parameters of this model are given in Supporting Information. The dataset contains only 150 points and because of the relatively small size, removing outliers would significantly reduce the available information. Instead of this, adding corrections to outliers can increase the performance of the KRR model. Here, extensive hyperparameter tuning could increase the risk of overfitting. To improve model accuracy without overfitting, we employed a residual learning strategy [79, 80], which states that another regression model should be trained on the residuals coming from the earlier trained dataset. This is structurally analogous to a penalty function formulation in non‐linear programming (Supporting Information, equation 19). Hence our model could be further improved by introducing an additional penalty term or descriptor accounting explicitly for polarization. Here, the penalty function is implemented as a separate model that is trained on the prediction errors of the earlier KRR model, using additional features to improve the correction of those errors. Since larger nanographenes exhibit higher molecular polarizability than smaller systems, the extent of anion‐induced polarization is expected to depend on nanographene size, which is not explicitly included in the current formulation. To determine an appropriate penalty function accounting for polarization effects, we isolated all data points corresponding to anion–π interactions (green symbols in Figure 4A). A multilinear regression model was then trained on this subset using analytical functions of the eigenvalues of the local quadrupole moment tensor as regressors such as qa,qb,qa,qb,qaqb,qaqb,qa/qb,qb/qa,qa/qb,qb/qa, while fractional powers of the number of carbon atoms (n C ) or the number of six‐membered rings (n r ) in the molecule were introduced as penalty functions (Supporting Information, equation 20). A detailed description of the methodology and functional forms is provided in the Supporting Information. The regressors were selected based on the Pearson correlation coefficient matrix presented in Table S2 in Supporting Information, six regressors qa,qb,qa,qb,qa/qb,qb/qa were selected as the most suitable variables for training the dataset. Figure 6 illustrates the dispersion of the trained data for two representative model specifications along with their corresponding coefficients of determination (R 2). In both cases, RMSE remains below 10 kJ mol−1. However, noticeable differences in model performance are observed upon increasing the number of regressors and modifying the penalty function. Scatter plots between calculated BE using DFT and predicted BE using multilinear regression (1) with 6 regressors (See Supporting Information) and penalty term n C and (2) with 7 regressors and (n C )1/16 as penalty term. The difference is quite straightforward, lower the power of n C in penalty term, stronger the correlation. The regressors are analytic functions of eigenvalues of local quadrupole moment qa,qb (i.e., SRmax, ec′) while the penalty term is added to compensate induction effects in larger nanographenes. Additional information is given in Supporting Information. A systematic comparison of model accuracy across different numbers of regressors and different penalty functions based on n C and n r is presented in Figure 7. For both variables, decreasing the fractional power of the penalty function leads to an increase in R 2, with the curves approaching a plateau at a fractional power of 1/16. Additionally, increasing the number of regressors from six to nine improves the model performance, although the improvement is nonuniform. A significant increase in R 2 is observed when moving from six to seven regressors, whereas further increases beyond seven yield marginal gains. In fact, models employing seven or eight regressors show nearly overlapping performance, indicating that the essential physical variations are already captured. When nine regressors are used, the corresponding curve converges with the others only at lower fractional powers (≤ 1/8). Overall, these results indicate that models with six regressors suffer from underfitting, whereas seven regressors are sufficient to accurately describe the data. Regarding the choice of penalty function, both n C 1/8 and n r 1/8 yield consistently high accuracy, with R 2 values exceeding 0.85. Although penalty functions based on n r exhibit slightly better performance, both descriptors provide a robust means of incorporating polarization effects into the regression framework. Comparison of performance of multilinear regression after choosing different powers of number of C centers (left) or number of 6‐MRs (right) present in system and increasing number of regressors. In all cases, the curve gets flattened as k increases and there is significant increase in R 2 as an additional regressor is added to 6 regressors, and ε(n C ) = (n C )1/k or ε(n r ) = (n r )1/k . From this figure, it can be concluded that 7 regressors are sufficient to fit data with (n C )1/8 or (n r )1/8 as penalty function (R 2 ≥ 0.85). For comparison, an automated model‐selection protocol was also applied using the Robert software package. Several algorithms were evaluated, including random forest (RF), multivariate linear (MVL), gradient boosting (GB) and neural network (NN). Among these models, GB provided the best performance (Figures S4–S11); however, the resulting correlation remained relatively weak (R 2 ≈ 0.42) with a mean absolute error of approximately 5.6 kJ mol−1 and a root‐mean‐square error of 7.7 kJ mol−1. These results confirm that generic machine‐learning approaches without physically motivated descriptors perform significantly worse than the descriptor‐based regression model developed in this work. Conclusions This study demonstrates that the strength of anion binding to aromatic systems can be rationalized using simple descriptors derived from the local quadrupole moment of six‐membered rings. Analysis of a dataset containing 171 Cl−–aromatic complexes revealed that conventional descriptors such as aromaticity indices, Fukui functions, and electron density at ring critical points show only weak correlations with binding energy, at variance with what we found for the cation‐π interactions [11]. This indicates that the interactions considered here are primarily governed by electrostatic effects rather than aromaticity or local reactivity. A more consistent description emerges when the local quadrupole moment is analyzed in terms of its geometrical representation in the molecular plane. Two descriptors derived from this representation—the scale factor (SRmax) and the ellipticity (ec′)—capture complementary physical effects. The scale factor reflects the overall π‐acidity of the aromatic ring and increases systematically with the number of electron‐withdrawing substituents, while the ellipticity quantifies the anisotropy of charge distribution induced by specific substitution patterns. Rings with both larger SRmax (large π‐acidity) and higher ellipticity (large anisotropy) exhibit stronger interactions with the chloride anion, largely independent of the specific binding motif. Extending this analysis to fluorinated naphthalenes and nanographenes confirms the generality of this relationship. Simple regression models based on SRmax and ec′ effectively replicate the observed binding energies, demonstrating that these descriptors offer a compact representation of the electrostatic features that govern anion binding. When combined with polarization‐dependent penalty functions, this regression model shows enhanced predictive performance, particularly for larger nanographene systems, where anion‐induced polarization plays a more significant role. This suggests that while the predictive model is still somewhat constrained by polarization effects, its simplicity, robustness, and transferability across a wide range of systems offer distinct advantages over more complex, multi‐layered machine learning models. Future work may involve incorporating additional polarization‐sensitive descriptors or developing models that explicitly account for these effects in order to improve accuracy for polarization‐dominated anion–π interactions. However, a notable limitation of this theoretical study is the experimental exclusion of anion‐π interactions in less‐fluorinated aromatic rings in the gas‐phase, as observed by Weber et al. [81, 82] In their study of anion‐arene complexes by infrared photodissociation spectroscopy, anion‐π bonding was detected only in perfluorinated units, whereas ionic hydrogen bonding was seen in other cases, despite the presence of π‐acidic character in both types of rings. Overall, the present results highlight that the magnitude and anisotropy of the local quadrupole moment constitute key determinants of anion–π interactions. The proposed descriptor framework provides both physical insight and a practical tool for predicting anion binding affinities in extended aromatic systems, potentially facilitating the design of π‐acidic receptors and functional materials for anion recognition. However, computational efficiency still remains a challenge for future applications. Author Contributions O.C. performed the calculations, the formal analysis, data curation, and original draft writing. A.P., M.S., and S.P.‐P. contributed to supervision, provided resources, participated in manuscript review and editing, and secured funding. M.S. conceived the study. Funding This work was supported by Universitat de Girona, IFUdG2023/12; Horizon 2020 Framework Programme, Marie Skłodowska‐Curie grant agreement No. 945380; Ministerio de Ciencia, Innovación y Universidades, PID2023‐147424NB‐I00, PID2024‐155989NB‐I00, RED2024‐154178‐T; Direcció General de Recerca, Generalitat de Catalunya, 2021SGR623, ICREA Acadèmia 2024, ICREA Excel·lència 2025; Barcelona Supercomputing Center, QHS‐2025‐1‐0018, QHS‐2025‐2‐0038. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Figure S1: Visualisation of local quadrupole moment tensor (ΘR) for (A) Benzene and (B) 1,4‐Difluorobenzene. It is clearly visible that upon intersection with the molecular plane, the local quadrupole moment tensor forms an ellipse. Figure S2: Ellipse formed by intersection of local quadrupole moment (ΘR) with molecular plane for (A) Benzene, (B) 1,4‐Difluorobenzene, (C) 1,3,5‐Trifluorobenzene, (D) 1,2,4,5‐Tetrafluorobenzene and (E) Hexafluorobenzene. The parameter ec, is an ellipticity of an ellipse while SRmax is curvature of orange dotted circle shown above. Upon fluorination, the size of orange dotted circle is getting reduced as its curvature (SRmax) is getting increased. The values of ec,=1−qbqa and SRmax=qa+qb are given in Table S3. Table S1: Performance of kernel ridge regression model applied on test data with different descriptor set and different kernel functions with parameters α = 1 and γ = 0.1. The MAE and RMSE values are compared to select most reliable kernel function. Along with these number of points with deviation greater than MAE are presented to understand skewness of distribution. Table S2: Pearson correlation coefficient matrix for EBECl−−π, regressors (analytic functions of qa, qb) and P. Figure S3: Scatter plots between aromaticity descriptor MCI (Blue), Fukui function for 6‐MR (Red) and Cl−‐BE for fluorobenzenes (C6H6 −xFx for 0 ≤ x ≤ 6). There is no direct correlation exist between these two pairs. Table S3: Series of Fluorobenzenes (C6H6−xFx for 0 ≤ x ≤ 6) and their electronic properties such as. Table S4: Series of Fluoronapthalenes (C10H8−xFx for 0 ≤ x ≤ 6) their names, and number of symmetrically non‐equivalent 6‐MRs (N). Figure S4: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S5: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S6: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S7: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S8: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S9: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S10: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S11: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Acknowledgments This work is dedicated to Prof. Boyd Russell on his 80th birthday, in recognition of his significant contributions to the theoretical and computational chemistry field. The authors thank the Consorci de Serveis Universitaris de Catalunya (CSUC) and the Barcelona Supercomputing Center: Red Española de Supercomputación (BSC:RES) for computational resources with project IDs QHS‐2025‐1‐0018 and QHS‐2025‐2‐0038. O.C. is grateful to the Universitat de Girona for the doctoral fellowship IFUdG2023/12. S.P.‐P. acknowledges the received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Skłodowska‐Curie grant agreement No. 945380. M.S. and A.P. thank the Spanish MICIU (MICIU/AEI/10.13039/50110001103) for projects PID2023‐147424NB‐I00, PID2024‐155989NB‐I00, and RED2024‐154178‐T, and the Generalitat de Catalunya for project 2021SGR623. M.S. and A.P. are grateful to the Generalitat de Catalunya for the ICREA Academia prizes 2024 and 2025, respectively. Data Availability Statement The data set of computational results such as xyz coordinates of molecular structures are available in the ioChem‐BD repository [83, 84] and can be accessed via https://doi.org/10.19061/iochem‐bd‐6‐640. References Associated Data Supplementary Materials Figure S1: Visualisation of local quadrupole moment tensor (ΘR) for (A) Benzene and (B) 1,4‐Difluorobenzene. It is clearly visible that upon intersection with the molecular plane, the local quadrupole moment tensor forms an ellipse. Figure S2: Ellipse formed by intersection of local quadrupole moment (ΘR) with molecular plane for (A) Benzene, (B) 1,4‐Difluorobenzene, (C) 1,3,5‐Trifluorobenzene, (D) 1,2,4,5‐Tetrafluorobenzene and (E) Hexafluorobenzene. The parameter ec, is an ellipticity of an ellipse while SRmax is curvature of orange dotted circle shown above. Upon fluorination, the size of orange dotted circle is getting reduced as its curvature (SRmax) is getting increased. The values of ec,=1−qbqa and SRmax=qa+qb are given in Table S3. Table S1: Performance of kernel ridge regression model applied on test data with different descriptor set and different kernel functions with parameters α = 1 and γ = 0.1. The MAE and RMSE values are compared to select most reliable kernel function. Along with these number of points with deviation greater than MAE are presented to understand skewness of distribution. Table S2: Pearson correlation coefficient matrix for EBECl−−π, regressors (analytic functions of qa, qb) and P. Figure S3: Scatter plots between aromaticity descriptor MCI (Blue), Fukui function for 6‐MR (Red) and Cl−‐BE for fluorobenzenes (C6H6 −xFx for 0 ≤ x ≤ 6). There is no direct correlation exist between these two pairs. Table S3: Series of Fluorobenzenes (C6H6−xFx for 0 ≤ x ≤ 6) and their electronic properties such as. Table S4: Series of Fluoronapthalenes (C10H8−xFx for 0 ≤ x ≤ 6) their names, and number of symmetrically non‐equivalent 6‐MRs (N). Figure S4: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S5: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S6: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S7: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S8: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S9: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S10: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Figure S11: Report from ROBERT[1] with best selected model trained on the dataset with abovementioned variables. Data Availability Statement The data set of computational results such as xyz coordinates of molecular structures are available in the ioChem‐BD repository [83, 84] and can be accessed via https://doi.org/10.19061/iochem‐bd‐6‐640.