A Preliminary Neural Network-Based Composite Method for Accurate Prediction of Enthalpies of Formation Authors: Gabriel César Pereira, Rogério Custodio Published: 2024 ABSTRACT AbstractA composite method, named ANN-G3S, is introduced, adapting from G3S theory and employing distinct sets of multiplicative scale factors. An artificial neural network (ANN)-based classification model is utilized to select optimal sets of four scale factors for electronic correlation and basis set expansion terms in electronic systems. The correlation and basis set terms are scaled by four parameters, two for atoms and the other two for molecules. The ANN model is trained on the G3/05 test set to identify the best parameter set for each electronic system. To validate the method, 10% of the structures from the test set are randomly excluded from training and optimization, forming a separate validation set. The method demonstrates a mean deviation of 1.11 kcal mol–1 for the G3/05 set and 0.89 kcal mol–1 for the validation set, close to the value presented by the G4 method and surpassing the accuracy of the G3 method of 1.19 kcal mol–1 with significantly reduced computational cost. This method shows advantages by eliminating the need for purely empirical corrections, thereby enhancing both efficiency and accuracy in predicting heats of formation. Introduction Many methods in the literature aim to determine thermochemical and spectroscopic properties for chemical systems of different sizes using composite approaches. The composite methods combine lower-level ab initio calculations yielding accurate results with reduced computational efforts compared to higher-level methods.1−5 Among the important composite methods are those in the Gaussian-n (Gn) family, developed by Pople, Curtiss, and colaborators.6−17 These methods estimate accurate electronic energy by applying linear corrections based on diverse conditions of electronic correlation effects and basis function. Often, in the case of Gn methods, empirical adjustments are employed to rectify inherent shortcomings, like high-level corrections (HLC), harmonic vibrational frequency scaling, atomization energy adjustments, among others.18−25 While effective in minimizing deviations from experimental data, it is sometimes unclear if these corrections address deficiencies in electronic correlation effects, basis functions, both, or other overlooked factors, including the quality of experimental data used in properties like enthalpies of formation. In an attempt to replace some of these corrections with empirical adjustments, the G3S7 method was developed. This method is based on the proposals by Gordon and Truhlar26−30 for scaling various electronic correlation terms, the PCI-X method by Siegbahn et al.31 for parametrized correlation, and the multi-coefficient correlation method (MCCM) by Truhlar et al.27,30 With a maximum of six parameters, the G3S method achieved an experimental accuracy level of 0.97 kcal mol–1 for calculating 148 enthalpies of formation, using the G2/9713,32 experimental reference set. Despite its effectiveness, G3S has limitations. Its representativeness was constrained by a relatively small structure set (148 molecules), and like many composite methods, it does not incorporate an external validation set to assess the generalizability of the method. This is a common issue, irrespective of whether the method involves adjustable parameters. These challenges highlight the need for new or improved composite methodologies. The objective of this work is to develop an adaptive composite method by directly adjusting correlation terms and basis functions using neural networks. This approach will utilize four parameters to refine electronic correlation terms and basis function effects to calculate enthalpies of formation. The efficacy and representativeness of the adjustments will be assessed through external validation. Computational Methods It is worth mentioning that the method described here is statistical in nature, employing artificial neural networks. Neural networks and other artificial intelligence methods are designed for classification, optimization, or both. In the first step, the HCA is used to establish a classification criterion based on the absolute error between calculated and experimental enthalpies of formation, grouping molecules into components that minimize the mean absolute error (MAE). Once the optimal number of groups is determined, the second step involves reoptimizing the parameters for each group to enhance the calculated enthalpies of formation for molecules within each group. Finally, the third step employs the ANN for classification and improvement of enthalpy predictions for new systems. With this preliminary outline in mind, the details of these three steps are elaborated in the following sections. General Adjustments The molecular internal energy expression is defined by a reference energy (Eref) modified by electronic correlation (ΔEcorr) and basis function effects (ΔEbasis), along with thermal contribution. Adjustments for atoms and molecules are independent, as shown in the following equations where ZPE33 represents harmonic zero-point energy and ESO indicates the spin–orbit correction for the atoms only. The reference energy is computed at the HF/aug-cc-pVDZ level. The electronic correlation correction is given by ΔEcorr = E[QCISD(T)/aug-cc-pVDZ] – E[HF/aug-cc-pVDZ], while basis function deficiency is addressed using Dunning type basis functions:34 ΔEbasis = E[MP2/aug-cc-pVTZ] – E[MP2/aug-cc-pVDZ]. Parameters a and c are for molecules, b and d are for atoms. Molecular geometries and all thermal effects for all single-point calculations were optimized at B3LYP/6-31G(d) level. Both the basis set increment and geometry optimization steps were performed following a scheme similar to that in the G3MP2(B3) method.10 All calculations were performed using 5d and 7f functions, which define the angular momentum components of d- and f-orbitals in standard basis sets, for basis set consistency with modern electronic structure codes. The use of 5d functions avoids the inclusion of an additional unphysical s orbital that can arise from 6d functions, mitigating potential basis set linear dependency issues. Preliminary tests showed negligible differences in the results compared to calculations using 6d functions, ensuring that the conclusions drawn remain valid. Thermal corrections were included only in the enthalpy of formation calculation, determined using a well-established method in the literature.35,36 The Gaussian 09 package37 was used for all calculations. Parameters a, b, c, and d were initially optimized for each molecule against G3/05 test set enthalpies of formation,16 aiming to minimize errors between calculated and experimental enthalpies of formation. Starting with an initial value of one for each parameter, the BFGS gradient method38 was used for optimization. The optimal parameters were arranged in a 248 × 4 matrix, containing the four values for each of the 248 molecules. Each column of the matrix was centered on its respective average value. Subsequently, a hierarchical clustering analysis using the Ward method39 was applied to generate up to four different groups. The clustering criterion was based on the similarity of the values of the four adjusted parameters. Following the characterization of each cluster through Hierarchical Cluster Analysis (HCA), 10% of the molecules from each cluster were set aside for validation of the final model. The remaining 90% were used to optimize the four parameters that represent the optimal settings for each respective group. Considering the available sample size of molecules, up to four different groups were formed during the clustering process. This approach helped to avoid parameter overfitting and ensured the representativeness of each group. As a result, four groups labeled G1, G2, G3, and G4 were established, containing 36, 94, 38, and 80 molecules, respectively. Once the training and validation sets were selected and the respective parameters for each training set were optimized, a classification model was constructed. This model assigns unknown molecules to one of the four groups to determine the optimal set of parameters (a, b, c, d) for use in eqs 1 and 2. The steps of this general method are summarized in Figure 1. Workflow of the ANN classification model, including parameter optimization, clustering, reoptimization, and classification. Classification Models To train a classification model for assigning unknown molecules to one of the generated groups, a set of features was designed based on the information available from the calculations necessary for implementing eqs 1 and 2. Table S1 in the Supporting Information lists the features used as input for training the model. Three classification methods were tested: K-nearest neighbors (KNN),40 support vector machine (SVM),41 and artificial neural networks (ANN).42 Each method was employed to develop the best classification model through a semiexhaustive search for optimal model parameters. For SVM models, the regularization parameter C, which controls the trade-off between model complexity and classification accuracy, and the kernel coefficient γ, which defines the influence range of a training example, were scanned across linear, radial, and polynomial kernels. The C and γ parameters were scanned in the range of 0.1–100 and 0.01–10, respectively. In the KNN model, the sole adjusted parameter was the number of nearest neighbors. Here, schemes with 3, 5, and 7 neighbors were experimented. For the ANN model, given the large number of parameters such as the number of layers, neurons, regularizers, etc., a manual search for the best set of parameters was carried out. Preliminary tests indicated that as the complexity of the classification problem increased with the number of groups considered, the ANN classification showed superior performance (Figure 2). The optimized ANN model contained 2 hidden layers, each using the Rectified Linear Unit (ReLU) activation function and containing 8 fully connected neurons plus 1 bias corrector. The output layer contained 4 units, each representing one of the groups considered in the classification. A key feature of ANN classification is its use of a softmax distribution in the output. This allows for the application of a weighted average of the adjusted parameters a, b, c, and d from each group, based on the respective probabilities of a molecule belonging to each group, rather than using the parameters of the most likely one. Comparison of the different classification methods with respect to the number of mistakes made by each model as the number of possible groups increases. Results and Discussion ANN-Driven Predictions Different ANN classification models were created, categorizing molecules into 2, 3, and 4 groups. Table 1 details the mean absolute error (MAE) results for the enthalpies of formation of test set molecules across these three levels of grouping. These results were obtained using two approaches: the most probable group and the weighted average according to the softmax distribution. The MAEs for both the test set and the entire G3/05 set predominantly ranged from 0.8 to 2.0 kcal mol–1. The uniformity of MAE values across all groups suggests a degree of consistency in the calculated errors. Table 1Mean Absolute Errors (MAE in kcal mol–1) in Calculated Enthalpies of Formation Compared to Experimental Data: (a) for the Test Set, and (b) for All Molecules in the G3/05 Group test setcomplete G3/05number of groupsMax.Prob.averageMax.Prob.average12.152.152.152.1521.901.731.991.9531.271.171.311.2841.170.891.261.12 Mean Absolute Errors (MAE in kcal mol–1) in Calculated Enthalpies of Formation Compared to Experimental Data: (a) for the Test Set, and (b) for All Molecules in the G3/05 Group Analysis of the data in Table 1 enabled the identification of the most effective grouping method for the G3/05 set, a critical step in finalizing the proposed configuration method. A trend of improved accuracy in adjustments was observed with an increased number of optimization groups. However, this improvement trend is not immediately apparent since more groups or classes for the molecules also complicate the classification process, potentially increasing the mean absolute error. To avoid diminishing the generalizability of the method, the number of groups was not excessively increased due to the limited size of the available molecule set. Additionally, when increasing the number of groups beyond 4, some groups contained too few molecules, leading to increasingly unreliable optimized parameters for those groups. Moreover, no significant improvement in the mean absolute error (MAE) was observed for either the training or validation sets to justify the inclusion of additional groups. For demonstration purposes, the MAE values for the validation set with 4, 5, and 6 groups are as follows: 4 groups: 0.89 kcal mol–1, 5 groups: 0.85 kcal mol–1, and 6 groups: 0.84 kcal mol–1. It is worth noting that for these tests, the same ANN model structure was applied in terms of the number of hidden layers and type of activation functions, with the only modification being the number of output units to match the number of groups. The model was not further optimized for 5 or 6 groups, as it was for the 4-group model. Figure 3 illustrates the optimized neural network configuration that presented the best performance. This network, tasked with selecting the parameters a, b, c, and d for each molecule, demonstrated R2 values of 0.90 and 0.89 for the training and test sets, respectively. Table 2 presents the calculated enthalpies of formation for the G3/05 set molecules, alongside deviations from experimental values. Framework of the ANN classification model. ReLU functions were used in the hidden layers, and softmax functions were used in the output layer. Table 2Experimental (Exp.) and Calculated (Calc.) Enthalpies of Formation (in kcal mol–1)amoleculeExp.Cal.moleculeExp.Calc.SiF4–386.00.35CH2CO–11.4–0.67PF5–381.12.42CH3CH2NH2–11.30.51C2F6–321.32.59cycC5H10NH–11.3–0.21Al2Cl6–309.70.44C2H5SH-ethanethiol–11.10.35SF6–291.72.27NH3–11–0.45AlF3–289.0–0.65CBrCl3–10–1.89BF3–271.42.05CH3SCH3–8.90.58PF3–229.10.11C4H4O–8.31.89C6F6–228.42.56C4H8S–8.2–1.31CF4–223.01.98CH3SiH3–70.50C6F5Cl–194.1–2.09(CH3)3N–5.70.56BeF2–190.3–0.03CH3SH–5.50.79CF3Cl–169.52.08CH3NH2–5.5–0.62CHF3–166.61.01ClFO3–5.10.08SiCl4–158.0–2.15H2S–4.90.64C2F4–157.43.69(CH3)2NH–4.4–0.43CF3Br–155.01.63SCl2–4.2–0.98COF2–149.1–4.23H2COH–4.10.19AlCl3–139.7–0.36C4H8,isobutane–4.0–0.04CH3–C=OOC=OCH3–136.80.85S2Cl2–4.0–0.16POCl3–133.89.95CH3CH2O–3.71.00CF3CN–118.4–1.58C4H6O–3.3–0.98CH3C=OOCH(CH3)2, isopropylacetate–115.1–0.25C2Cl4–3.00.37CF3–111.31.25CH3CO (2A)–2.4–0.22CH2F2–107.70.23C4H8NH–0.80.28CH3COF–105.71.00F20.00.80CH3COOH–103.40.40H20.00.71CHF2Br–101.60.12Cl20.00.50CH3COOCH3–98.40.28O20.00.31BCl3–96.3–1.01N20.0–0.35SO3–94.6–0.17SO1.2–0.80CO2–94.10.70PH31.3–0.38MgCl2–93.8–0.33N2O42.2–0.31CH3–CH(OCH3)2–93.12.63ClNO22.90.80HCOOH–90.51.37CH3O4.1–0.01(CH3)2SO2–89.7–1.36CH3–CH=CH24.80.11PCl5–86.11.53(CH3)2CHCN5.6–0.66HCOOCH3–85.02.48F2O5.92.10SO2Cl2–84.86.63C4H8,cyclobutane6.80.77LiF–80.1–1.18NO27.9–0.45(CH3)2CHOCH(CH3)2–76.3–0.02SiH48.21.16C4H8O2,(para)–75.51.40CH2=CHCl8.91.75(CH3)3COH–74.70.26OH9.40.99C6H4F2, 1,3-difluorobenzene–73.9–1.40HCO10.0–1.02C6H4F2, 1,4-difluorobenzene–73.3–1.43C6H5–CH3,toluene12.0–1.01SO2–71.00.67(CH3)3C12.3–0.08NaF–69.4–2.11C6H5Cl12.43.12PCl3–69.0–1.85ClNO12.4–0.15(CH3)3COCH3–67.80.38C2H412.5–0.06(CH3)2CHOH–65.20.76C3H6,cyclopropane12.7–0.24HF–65.10.71P414.1–1.33AlF–63.50.25CH3C=OCCH15.6–0.96CH3CH2COCH2CH3–61.6–0.85C5H8,isoprene18.0–0.46CH3CH2OCH2CH3–60.30.54CH3CN18.00.56CH3COCl–58.0–0.49C2H3Br18.9–1.55H2O–57.80.11Si2H619.11.79CH3COCH2CH3, methylethylketone–57.1–0.47N2O19.6–0.53CH3CONH2–57.00.08C2H4S19.60.41CH3CH2OH–56.21.04C6H619.7–1.13Si(CH3)4–55.7–3.97N2O319.8–0.16C5H10O–53.4–1.11C5H6S,methylthiophene,20.0–2.59Cl2CO–52.4–0.73C4H6S20.8–0.79CH3COCH3–51.9–0.34C6H5NH220.8–1.35C2H5OCH3methyl–ethyl-ether–51.71.50(CH3)2CH21.50.76(CH3)2CHCHO,isobutanal–51.6–1.82NO21.6–0.91HCOCOH–50.72.48H2NNH222.8–0.90C8H18–49.9c1.45ClO24.2–0.91H3COH–48.01.51C5H7N24.6–0.91C5H8O–45.9–1.47C6H8,1,4-cyclohexadiene25.0–0.80C7H16–44.91.30C6H5Br25.2–2.97CH3OCH3–44.01.52C6H8,1,3-cyclohexadiene25.4–0.44C4H8O (tetrahydrofuran)–44.0–0.76C4H5N25.9–1.07NaCl–43.8–1.58CH2CHCHCH2,butadiene26.3–0.59(CH3)3CCl–43.51.38C4H4S27.5–1.52C6H14,methylpentane–41.10.54CS228.0–1.09SiCl2–40.3–2.43C2H528.90.67C5H12,neopentane–40.2–0.08CH3S29.80.96C6H14–39.91.18C2H4NH30.2–0.10CH3CHO–39.70.23S230.7–1.30CH3CH2CH(CH3)NO2, nitro-s-butane–39.10.86HCN31.50.43ClF3–38.0–19.68ClCN32.9–0.83CH3CH2CH2CH2Cl–37.00.72PH233.11.00(CH3)2SO–36.20.08LiH33.30.37C6H13Br–35.4–1.33C5H5N33.6–0.51C5H12,n-pentane–35.10.77Na234.00.26CH2=CHF–33.20.40O334.10.92OCS–33.00.64HS34.20.45H2O2–32.52.01P234.3–1.02C4H10,ISO–32.10.37C4H6,2-butyne34.8–0.18NF3–31.60.02CH335.00.35CH3CH2CH2Cl–31.51.99C10H8,naphthalene35.9–0.89C4H10,trans–30.00.75C4H6,cyclobutene37.4–0.92C6H12–29.50.38CH3–CH=C=CH238.80.36C6H4O2–29.4–3.49CH2CHCN43.2–1.47(CH3)3CNH2,t-butylamine–28.90.13LiNa43.40.33C6H5F–27.74.53CH3CCH44.20.27COBr2–27.1–3.29C5H8,spiropentane44.30.02C2H5Cl–26.81.43NH245.11.43CO–26.4–0.47CH2=C=CH2, (allene)45.50.50(CH3)3CSH–26.20.49C4H4N2,1,4-dipyridine46.8–3.08H2CO–26.00.46C4H4N2,pyrimidine46.91.48C3H8,propene–25.00.56SiH347.90.99CHCl3–24.7–0.94C4H6-methylenecyclopropane47.91.90SiO–24.6–1.20NCCH2CH2CN50.10.68CH3CH=CHCHO–24.0–0.39Li251.60.27C3H7Br–23.8–2.35C4H6,bicyclo51.9–1.67C6H5OH–23.0–1.01C2H254.2–0.28CCl4–22.9–1.06SiH2,singlete65.21.33CH2Cl2–22.8–0.58C3H466.2–1.30HCl–22.10.80CS66.9–0.69C2H6–20.10.38C10H8,azulene69.1–2.92CH3Cl–19.5–0.13C8H870.7–1.18C5H10,cyclopentane–18.30.17C2H371.60.40CH3CH2SSCH2CH3–17.9–1.21NCCN73.3–0.46CH4–17.90.03C6H5, phenil_radical81.2–2.87CH3NO2–17.8–1.10BeH81.7–0.42HOCl–17.81.13NH85.20.77C3H6Br2–17.1–0.24SiH2,triplet86.21.00CH3ONO–15.90.30CH2,triplet93.7–0.15C5H10S–15.2–1.55CH2,singlet102.80.46C2H5Br–14.8–0.94CN104.9–2.21ClF–13.20.39CCH135.1–1.14C5H8Br2–13.1–0.32Si2139.9–1.15C2H4O–12.60.94CH142.50.43aThe Calc. column refers to the mean absolute error (MAE) between the calculated and the experimental values. Experimental (Exp.) and Calculated (Calc.) Enthalpies of Formation (in kcal mol–1)a The Calc. column refers to the mean absolute error (MAE) between the calculated and the experimental values. Group Formation and Adjustments The rationale behind the adaptive composite method lies in creating a calculation approach that is (a) versatile across various molecular systems without significantly compromising accuracy; (b) representative, having undergone validation against an external set of test molecules; (c) capable of providing a basic level of physical rationale in its corrections; and (d) computationally efficient. The rigor and versatility of the method are highlighted in Table 1, which shows a 20% reduction in the mean error for the test set when using two adjustment possibilities for the a, b, c, and d parameters in enthalpy of formation calculations. Expanding to four adjustment possibilities leads to a 46% reduction in MAE. Notably, increased adjustment options introduce complexity, potentially complicating the accurate classification of molecules and affecting enthalpy of formation calculations. This underscores the importance of a robust classification algorithm. An attempt to understand how groups are formed in relation to the chemical composition of the molecules can be insightful. Table 3 presents the elemental composition of groups 1 and 2. The division of the G3/05 set into two groups, as analyzed in Table 3, reflects the chemical composition of the molecules in each group. This information aids in identifying where the effects of electronic correlation and basis function enhancement, implemented in the method, are either under or overestimated. Table 3Normalized Rate of Occurrence of Each Element in Each GroupaatomsGroup1Group 2H0.504.57C0.502.58N0.030.23O0.190.48F1.140.25Si0.060.06P0.000.06S0.030.14Cl0.610.26Br0.140.04aThe count was made over the complete G3/05 set. Normalized Rate of Occurrence of Each Element in Each Groupa The count was made over the complete G3/05 set. Comparing the parameters of each group (Table 4), it is noted that the differences between scaling parameters a and b, associated with the electronic correlation term, and c and d, associated with the basis set correction term are very small for Group 1. However, both differences (amolecule – batoms) and (cmolecule – datoms) are large for Group 2. Table 3 shows that Group 1 contains most molecules with fluorine, chlorine, and bromine. The presence of these electron-rich elements justifies the need for approximately equal scaling of the correlation correction term for atoms and the molecules containing these elements in Group 1. On the other hand, Group 2 has a higher presence of nitrogen, carbon, hydrogen, and oxygen. These elements are associated with the larger organic molecules contained in Group 2, which require a higher scaling of the basis set correction term in the molecule due to the greater spatial span of their electronic density. Moreover, the greater value of batoms when compared to amolecule in Group 2 seems to counter the overestimation of the correlation correction term for the molecules in this group. Table 4Optimal Parameters Obtained for Groups 1 and 2 in the Two-Group Clustering StageparametersGroup 1Group 2a0.9627020.971345b0.9706291.015386c1.3913221.420572d1.3990961.348204 Optimal Parameters Obtained for Groups 1 and 2 in the Two-Group Clustering Stage Moving to the grouping in 3 groups (Table 5), a separation of the organic molecules in Group 2 is observed, with two further subgroups: oxygenated and halogenated molecules, and nitrogenated molecules with almost no halogens. Further subgrouping into 4 groups results in the division of Group 3 into molecules richer in oxygen, nitrogen, and sulfur from the smaller molecules containing fluorine and chlorine. Similarly to what was observed in the first grouping (Table 6). Similarly to the grouping in 2 groups, the difference (amolecule – batoms) is smaller for the groups containing more halogens (Groups 1 and 3) and higher for groups without these elements. Additionally, the difference (cmolecule – datoms) is higher in the groups containing molecules with longer carbon chains (Groups 2 and 4). Further subgrouping would likely increase chemical specificity within each group but also raise the risk of overfitting. A complete breakdown of the molecules present in each of the four final groups is presented (Table S2) in the Supporting Information along with all the optimized parameters for each group (Table S3). Table 5Occurrence Rate of Each Element in the Molecules Present in the Generated GroupsatomsGroup 1Group 2Group 3H0.505.953.47C0.503.561.79N0.030.330.14O0.190.220.69F1.140.050.42Si0.060.020.08P0.000.090.04S0.030.070.19Cl0.610.090.40Br0.140.030.05 Occurrence Rate of Each Element in the Molecules Present in the Generated Groups Table 6Occurrence Rate of Each Element in the Molecules Present in the Generated 4 Final GroupsatomGroup 1Group 2Group 3Group 4H0.505.950.874.70C0.503.560.662.33N0.030.330.240.10O0.190.220.580.74F1.140.050.820.22Si0.060.020.110.08P0.000.090.030.05S0.030.070.130.22Cl0.610.091.000.11Br0.140.030.050.05 Occurrence Rate of Each Element in the Molecules Present in the Generated 4 Final Groups Hierarchical clustering analysis reveals that groupings of more than two categories emerge as branches of the initial grouping. This branching provides insights into the chemical specificity of each group and its constituent molecules, as well as the necessary corrections to be applied when calculating the properties of these groups (Table 7). Table 7Optimal Parameters Obtained for Groups 1, 2, and 3 in the Three-Group Clustering StageparametersGroup 1Group 2Group 3a0.9627020.8596440.999648b0.9706290.9810001.028248c1.3913221.3625461.300952d1.3990961.2372921.265845 Optimal Parameters Obtained for Groups 1, 2, and 3 in the Three-Group Clustering Stage General Implementation This section outlines the process for determining the a, b, c, and d parameters for a given molecule. As an example, we will calculate these parameters for SiF4. The method involves several steps, which are detailed below: Generation of the Feature Array The feature array is constructed using the descriptors listed in Table S1. For SiF4, the feature array is defined as SiF4: features = [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 4.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, −0.93, 0.38, 1.31, −0.39, 0.49, −0.44, 0.87, 0.01, −0.0, 0.07, 0.81, 180.48, 4.96] Normalization of the Feature Array The feature array is normalized using the mean and standard deviation values obtained from the training step (provided in the Supporting Information). The normalized feature array is calculated as normalized = (features – mean)/std. Feeding the Normalized Array into the ANN Model The normalized feature array is then input into the pretrained Artificial Neural Network (ANN) model. The output is a probability distribution representing the likelihood of the molecule belonging to each of the four predefined groups: Prob. = softmax(ReLU(ReLU([normalized]·[W1])·[W2] + b2)·[W3] + b3), where softmax(xi) = exi–max(x)/(∑j = 1n exj–max(x)) and ReLU(xi) = max(0, xi). Calculation of the Final Parameters The final a, b, c, and d parameters are computed as the weighted average of the parameters listed in Table S3, based on the probability distribution calculated in the previous step. For SiF4, the calculation is as follows Note: For clarity, the matrix values have been truncated. For practical calculations, all significant figures should be used to ensure accuracy. A plug-and-play implementation of the present method is accessible through the web application available at: https://gabrielcesar9.github.io/ANNG3S/. Users can input energy values from single point calculations to obtain the optimal parameters a, b, c, and d, as well as the final energy, atomization energy, and heat of formation for their system. Additionally, the parameters a, b, c, and d can be calculated using the ANN mathematical model, which involves the dot product of the matrices containing the weight parameters of the model (eq 3) as implemented in a simple python function made available through the script composite.py in the Supporting Information. Conclusions This study highlights the increasing importance of machine learning algorithms in determining chemical properties and developing high-accuracy, cost-effective computational methods. An adaptive composite method was devised, utilizing advancements in computational chemistry to effectively balance accuracy, representativeness (externally validated), and processing efficiency. The validity of the method is demonstrated by its ability to approximate enthalpies of formation with an accuracy of about 0.89 kcal mol–1 for molecules in the validation set, i.e., those outside the training set. This reflects the robustness of the model when applied to unseen data. Additionally, the method offers computational efficiency comparable to the most accurate Gn methods. It provides a slight improvement over G3 with respect to experimental data and approaches the performance of the G4 method, while reducing computational costs by utilizing smaller basis sets and lower-order perturbation computations. To quantify this, the mean absolute errors (MAE) compared to experimental values for some Gaussian-n methods and the proposed ANN-G3S method on the training set are as follows: G3:1.19 kcal mol–1, G3(MP2): 1.40 kcal mol–1, G4:0.80 kcal mol–1, G4(MP2): 0.99 kcal mol–1, and ANN-G3S: 1.11 kcal mol–1. These results demonstrate that the ANN-G3S method performs slightly better than G3 and approaches G4′s accuracy, while maintaining greater computational efficiency. A key feature of this method is the transparent implementation of empirical adjustments in electronic correlation terms and basis functions, which make clear the limitations of the theoretical levels used. This insight lays the groundwork for future optimizations, aimed at more precisely modeling specific system types, informed by identified deficiencies in parameters a, b, c, and d concerning modeling correlation and basis set effects. Future efforts will focus on refining this method, particularly its core formulation (eq 1) and the classification process crucial for deriving optimal parameters with maximum confidence. These attempts will aim to enhance the adaptability and predictive power of the method, further solidifying its role in electronic calculations. Acknowledgments The authors acknowledge financial support from FAPESP (Fundação de Amparo à Pesquisa do Estado de São Paulo - Center for Computational Engineering and Sciences, Grant 2013/08293-7, and Grant 2017/11485-6). This study was also financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - Brasil (CAPES) - Finance Code 001. Supporting Information Available The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jctc.4c01351. All tools and data necessary to ensure reproducibility and ease of access; data vector used as the information source for model training (Table S1); parameter values of each of the four final groups (Table S2); parameter values of each of the four final groups (Table S3); additionally, a Python script (composite.py) is provided to enable users to calculate molecular properties and classify molecules locally, without relying on external web servers, this program offers a straightforward implementation of the proposed methodology (PDF) All tools and data necessary to ensure reproducibility and ease of access; data vector used as the information source for model training (Table S1); parameter values of each of the four final groups (Table S2); parameter values of each of the four final groups (Table S3); additionally, a Python script (composite.py) is provided to enable users to calculate molecular properties and classify molecules locally, without relying on external web servers, this program offers a straightforward implementation of the proposed methodology (PDF) The Article Processing Charge for the publication of this research was funded by the Coordination for the Improvement of Higher Education Personnel - CAPES (ROR identifier: 00x0ma614). The authors declare no competing financial interest. Supplementary Material ct4c01351_si_001.pdf References Associated Data Supplementary Materials ct4c01351_si_001.pdf