-
Due to its simple electronic structure (three protons and three electrons), H3 system has long been a benchmark system for both quantum chemistry calculations and reaction dynamic studies. The accurate treatment of its electronic structure enables the detailed examination of various microscopic phenomena and mechanisms in reaction collisions. In addition, the intrinsic stability of the nonpolar H
$ _{2} $ molecule’s orientation also makes H+H$ _{2} $ system particularly suitable for investigating stereodynamics in reaction dynamics [1–4]. Extensive experimental data on H+H$ _{2} $ and its isotopic reactions provide strong data support for validating theoretical models, highlighting the essential role of H3 system in advancing the field of molecular reaction dynamics [5–10]. Therefore, H3 system is highly regarded among researchers seeking to unravel molecular reaction mechanisms through theoretical analysis.The exchange reactions of H+H
$ _{2} $ and its isotopes exhibit a conical intersection between the two lowest electronic states when the three nuclei form an equilateral triangle configuration, with the minimum energy being around 2.7 eV. Due to its simplicity, this reaction system is recognizedas an optimal model for the examination of geometric phase in chemical reactions [11]. Early theoretical studies suggest that although a subtle geometric phase effect is observed in state-to-state reaction probabilities, this effect appears to diminish when calculating state-to-state integral cross sections [12]. Until recently, Yuan and coworkers have confirmed the geometric phase effect in the H+HD exchange reaction at 2.77 eV, through a combinational study of theory and experiment [13]. Further works revealed geometric phase effects at energies of 2.28 eV and 1.72 eV, affirming their relevance beyond the Born-Oppenheimer approximation, especially in altering differential cross-section angular distributions [14, 15]. In 2023, Wang and cowokers extended the inquiry into the H3 system at collision energies of 0.5 eV, 1.2 eV, and 2.07 eV, unveiling the crucial role of quantum interference in perpendicular collision configuration reactions. This insight has significantly advanced the understanding of stereodynamics in molecular reactions [4]. These collective efforts have substantially advanced the understanding of molecular reaction dynamics, demonstrating the synergy between experimental validation and theoretical innovation in addressing complex chemical phenomena.Current research on the H3 system predominantly focuses on collision energies below 3 eV. The investigations into higher energy collisions are rather rare. In 2008, Bouakline et al. expanded the scope of theoretical analysis including collision energies up to 4.5 eV, revealing that beyond 3.5 eV, geometric phase effects significantly influence state-to-state reaction probabilities [16]. These findings support earlier observations that such effects are compensated for integral cross sections (ICS), indicating a discrepancy between theoretical predictions and experimental observations at higher energies. The advancement of experimental techniques now enables studies at these higher energies, which demands a more comprehensive theoretical framework [17, 18]. While the BKMP2 (Boothroyd-Keogh-Martin-Peterson-2) potential energy surface (PES) [19] offers remarkable accuracy for lower energy collisions, its applicability becomes restricted at higher energies where the conversion of kinetic energy to potential energy may involve excited states. This situation underscores the limitation of solely relying on adiabatic potential energy surfaces for accurate theoretical predictions. Therefore, the development of a diabatic potential energy matrix, designed for high-energy collisions, with better accuracy, is essential. This advancement is critical for bridging the gap between theoretical models and experimental observations, paving the way for a deeper understanding of the H3 system across a wider energy range.
In 2002, Abrol et al. developed a diabatic potential energy matrix (DPEM) for the H3 system using derivative coupling data, by solving the three-dimensional Poisson equation [20]. In 2018, Guan et al. employed neural network techniques to solve the Poisson equation to construct the DPEM. This study directly integrated the BKMP2 PES [19] for the ground state, while the excited state surface was derived by adding the energy differece between two electronic states. Utilizing this DPEM, they conducted quantum dynamic simulations and examined them against the experimental findings to explore the geometric phase effect [13]. Both of these work addressed the system’s symmetry through solutions of the Poisson equation. Recently, Yin et al. proposed an innovative method for handling the symmetry of H3 system, integrating the DPEM elements relationships associated with the transformations of the permutation group S3 into the functional form of the DPEM [21]. This approach strictly adheres to the symmetry constraints of the system, offering a novel framework for constructing the DPEM of H3.
In this study, we aim to construct a new DPEM for high-energy collisions based on the method proposed by Yin and coworkers. This new DPEM will use a more accurate ab initio treatment and incorporate more data points. The ab initio calculations of the H3 system in this study are performed using the full configuration interaction (FCI) method with aug-cc-pVnZ series of basis sets. To enhance accuracy, a two-parameter formula is utilized for complete basis set (CBS) extrapolation [22–26]. Building upon the point set from Yin et al. [21], additional data points are incorporated within higher energy range based on a systematic selection procedure [27, 28]. Finally, the ab initio data are fitted and reproduced by a symmtrized DPEM expressed through fundamental invariant neural network (FI-NN) [21, 29, 30].
In the Methodology section, we will provide a brief overview of the symmetry of the DPEM elements for H3 and the diabatization procedure with FI-NN. In the Results and Discussion section, we will analyze and discuss the properties of the newly constructed DPEM, including the reproduction of ab initio data and quantum dynamic calculations. The Conclusion and Outlook section provides a concise summary.
-
In this study, the Molpro software was employed for full configuration interaction (FCI) calculations with aug-cc-pVTZ and aug-cc-pVQZ basis sets [22–25]. All calculations were performed with
$ C_{\rm{s}} $ point group. A two-parameter complete basis set (CBS) extrapolation technique given by Eq.(1) was utilized to improve the accuracy of ab initio energy [26].In Eq.(1), the variable n corresponds to n within the aug-cc-pVnZ basis set series. Throughout the extrapolation process, the Hartree-Fock energy was directly obtained from the results with aug-cc-pVQZ basis set [22]. The extrapolation for the correlation energy was done by setting n=3 and n=4 in Eq.(1), and
$ E_{\infty }^{\rm{corr}} $ can be obtained from Eq.(2).In this research, we employed the point set previously prepared by Yin et al. [21], which was selected using grid sampling with the hyperspherical coordinates. The hyperradius
$ \rho $ , ranged from 1.4 bohr to 18.0 bohr, while the hyperangles,$ \theta $ and$ \phi $ , were confined to the ranges of 0° to 90° and 60° to 120°, respectively. It was observed that they discarded data points with energies higher than 10.0 eV relative to the ground state H+H$ _{2} $ asymptote. To expand the energy range of the DPEM, we have added additional data points in higher energy regions by utilizing a methodical selection process inspired by their work. Consequently, a total of 10985 data points were assembled, which fully saturated the configuration space we are interested in. -
In the case where the three nuclei form an equilateral triangle, degeneracy occurs between the 1
$ ^{2}A' $ and 2$ ^{2}A' $ electronic states, which are subsequently classified under the two-dimensional irreducible representation E of the D$ _{\rm{3h}} $ point group. The complete nuclear permutation inversion (CNPI) group for H3 is identified as the doublet group associated with S3, which is isomorphic to$ D_{\rm{3h}} $ point group. The symmetry properties of DPEM elements under CNPI operations has been fully resolved previously [21]. Utilizing$ r_{1} $ ,$ r_{2} $ ,$ r_{3} $ to represent bond length and$ f_{1} $ ,$ f_{2} $ ,$ f_{3} $ to denote intermediate functions which will be represented by the output layer of the neural network in this work, the DPEM$ \mathbf{H}^{d} $ for H3 takes the following form:For a two-state DPEM
$ \mathbf{H}^{d} $ ,$ {H}^{d}_{21}={H}^{d}_{12} $ . -
Neural networks have been proven very effective in constructing DPEM [13, 21, 30–33]. Therefore, in this work, the fundamental invariant neural network (FI-NN) method is employed to construct the DPEM for H3. FIG. 1 shows the structure of feedforward neural network used in this work. Initially, the bond lengths were transformed into fundamental invariant (FI) polynomials, which constitutes the input layer of the neural network. To obtain better fitting results, the FI polynomials were generated with the inverse of bond lengths, which are given by Eq.(6).
The neural network is optimized by reproducing ab initio data. Different from previous derivative coupling based methods, the loss function in this study is formulated to include only adiabatic energies errors (i.e., diabatization by ansatz). The structure of the loss function is given by
where
$ Q $ represents the total number of geometries in the dataset,$ \mathrm{\mathit{a}} $ denotes the adiabatic representation, (nn) indicates the fitting results from the neural network model, and (ab) corresponds to the results obtained from ab initio calculations.The diabatization by ansatz procedure is employed based on two considerations: firstly, adiabatic energies obtained via complete basis set (CBS) extrapolation might not seamlessly align with derivative coupling and gradient data, posing unnecessary difficulties in the fitting process; secondly, the computational resources required for accurate derivative couplings are notebly high. Nonetheless, the symmetry framework established in this study is designed to alleviate the potential loss of information resulting from the exclusion of derivative couplings.
The output layer of NN comprises intermediary functions (
$ f_1 $ ,$ f_2 $ and$ f_3 $ ) in Eq.(3), Eq.(4), and Eq.(5). It is important to note that excluding gradient information from the loss function increases the risk of overfitting. To address this, early stopping is employed during the fitting process to effectively mitigate overfitting. The structure of neural network used in this work is 3-50-50-3, which has 3 FIs in the input vector, 50 neurons in the two hidden layers and 3 elements in the output layer. The activation functions for the two hidden layers are hyperbolic tangent sigmoids, and a linear activation function is used for the output layer. -
The root mean square error (RMSE) for the adiabatic energies is determined to assess the quality of neural network fitting, with RMSEs for the ground and excited states being 0.228 meV and 0.611 meV, respectively. FIG. 2 illustrates the error distribution in the fitting of adiabatic energies for both the ground and first excited states. It becomes evident that for both the ground state and the first excited state, a significant portion of the data points in the error distribution falls into the absolute error range less than 1 meV. However, more significant deviations are observed for the first excited state with the potential energy being around 10 eV. In this particular energy range, there exists degeneracy with higher excited states. The current two-state DPEM is inadequate to describe such conical intersections (CIs), which thus leads to larger fitting errors. However, considering their high energies, these conical intersections are irrelevant to the exchange reactions of H+H
$ _{2} $ that we are interested in.An accurate description of conical intersections is essential for an accurate DPEM. For a clear representation of the PESs in the vicinity of CI, we introduce the hyperspherical coordinates (
$ \rho, \theta, \phi $ ) of Kuppermann [34], which are defined as follows:where
$ S_\lambda= d_\lambda R_\lambda $ and$ s_\lambda= r_\lambda/d_\lambda $ , and$ d_\lambda $ is a dimensionless scaling factor,where
$ \mu $ is the three-body reduced mass and$ M $ is the total mass of the system. The hyperspherical coordinates are connected to Cartesian coordinates ($ x, y, z $ ) via the following equation:When
$ x=z=0 $ , it corresponds to an equilateral triangle configuration, where the two electronic states become degenerate. The utilization of hyperspherical coordinates facilitates a clear visualization of these intersections. FIG. 3(B) provides the three-dimensional plot of$ {H}^{d}_{11} $ and$ {H}^{d}_{22} $ for H3 with hyperspherical radius ($ \rho $ ) being 2.5 bohr. Correspondingly, FIG. 3(A) shows the three-dimensional plot for E$ _{1} $ and E$ _{2} $ . The conical intersection can be clearly observed with energy being 2.74 eV. As can be seen from the figure, the cusps in the PESs due to conical intersections have been replaced by two smooth and continuous surfaces ($ {H}^{d}_{11} $ and$ {H}^{d}_{22} $ ), thus justifying the process of diabatization.FIG. 4 presents a one-dimensional cut of adiabatic energy and derivative coupling norm ||f1,2|| obtained from FI-NN based DPEM on the
$ xz $ plane, with$ z $ set to 0 and$ x $ spanning from −1.5 bohr to 1.5 bohr. Notably, at the coordinate origin, which corresponds to an equilateral triangular configuration, an exact intersection of the two electronic states is observed, accompanied by a divergent peak of derivative couplings.In the DPEM constructed by Guan et al. [13], adiabatic energy of the ground state is obtained from the BKMP2 PES, while the energy of the first excited state is obtained by adding an energy difference on top of the BKMP2 surface. This approach ensures exceptional accuracy in determing adiabatic energies. However, the symmetry of DPEM is achieved by solving the Poisson equation constrained by boundary conditions. To be specific, the mixing angles on the boundaries are symmetrically connected. Different symmetrized diabatization methods lead to different results in reproducing the derivative couplings. FIG. 5 compares the derivative couplings for H3 obtained from ab initio calculation, the DPEM constructed by Guan et al., and the DPEM constructed in this work. The derivative couplings are plotted in the orthogonal and normalized coordinate system Eq.(11) proposed by Guan et al. [13].
In FIG. 5,
$ \beta $ is fixed at 0.01 bohr, and$ s_{1} $ is fixed at 2 bohr, which approximates the molecular structure to an equilateral triangular configuration. The ab initio derivative couplings were obtained from multi-reference configuration interaction with single and double excitations (MRCI-SD) wavefunctions. The COLUMBUS computational package [35–37] was used to perform MRCI-SD calculations and the corresponding basis set was aug-cc-pVQZ [22]. As can be seen from the figure, although the overall range of derivative couplings being relatively small, all the derivative couplings exhibit similar pattern. Howerver, the derivative couplings obtained from this work closely match the ab initio results. This observation indicates that the DPEM constructed in this work effectively captures derivative coupling terms, despite the fact that derivative couplings were not explicitly included in the fitting process. The accuracy of diabatization by$ ansatz $ is limited because it primarily considers adiabatic energies in constructing the DPEM and neglects other pertinent information. However, the DPEM model used in this study is formulated using rigorous group representation theory. This theoretical framework guarantees that the exact degeneracy with D$ _{3\mathrm{h}} $ symmetry can be achieved. Moreover, the well reproduced adiabatic energies also guarantee a good description of the local topography of conical. This method effectively compensates for the limitations of diabatization by ansatz, thus enabling the accurate reproduction of derivative couplings even without using the information of derivative coupling data.To further test the quality of DPEM constructed in this work, quantum reaction dynamic calculations were performed. FIG. 6 illustrates the J=0 reaction probability of H+H2(
$ v $ =0,$ j $ =0) as a function of collision energies. In addition to the results from DPEM constructed in this work, FIG. 6 also compares the total reaction probabilities obtained from DPEMs previously constructed by Guan et al. [13] and Yin et al. [21]. As depicted in the figure, the three reaction probability curves almost coincide. Notably, in the region where the collision energy exceeds 1.5 eV, the results from this study align more closely with those of Guan et al. [13]. This outcome suggests that the DPEM developed in this work has achieved the anticipated dynamical accuracy.In FIG. 7, we compare the angular distribution of differential cross-sections for the reaction H + D2(
$ v $ =0,$ j $ =0)$ \rightarrow $ HD($ v $ '=0,$ j $ '=8) + D with a collision energy of 2.38 eV, based on three DPEMs. As can be seen from the figure, significant differences between adiabatic and diabatic differential cross-sections at smaller scattering angles can be observed, affirming the influence of the geometric phase effect. Moreover, the results from three DPEMs show considerable agreement, suggesting that all the diabatic models have attained a high level of accuracy. Nonetheless, subtle discrepancies persist in the differential cross-sections at larger scattering angles, suggesting the need to further enhance the precision of the DPEM. -
In this study, an accurate two-state DPEM for the H3 system has been constructed. The elements of DPEM are symmetrized to preserve CNPI symmetry of H3 system and fundamental invariant neural network (FI-NN) is introduced to provide the flexibility to fit high level ab initio energies, which are obtained by complete basis set (CBS) extrapolation of full configuration interaction (FCI) energies based on a two-parameter formula. This DPEM effectively captures the derivative coupling terms, despite that derivative couplings are absent in the fitting process. Quantum dynamic results demonstrate that the outcomes of this work are closely aligned with those obtained from previous benchmark DPEMs. This two-state DPEM extends the accurate representation of adiabatic energy up to 18 eV, making it the most accurate DPEM for H3 system to date. Ultimately, this work provides the scientific community with a reliable DPEM, advancing the exploration of reaction mechanisms at higher collision energies.
A New Diabatic Potential Energy Matrix for the Two Lowest Electronic States of H3
- Received Date: 17/04/2024
- Available Online: 28/02/2025
-
Key words:
- Diabatic potential energy matrix /
- Conical intersection /
- Fundamental invariant neural network
Abstract: A new two-state diabatic potential energy matrix (DPEM) for H3 has been constructed, based on the fundamental invariant neural network (FI-NN) diabatization method proposed in our previous work [Phys. Chem. Chem. Phys. 21, 15040 (2019)]. In that initial effort, a two-state DPEM was constructed only with a 10 eV energy threshold. The current work aims to expand the energy range and improve the accuracy of DPEM. This is achieved by the utilization of full configuration interaction (FCI) with aug-cc-pVnZ basis sets and complete basis set (CBS) extrapolation. The original dataset is augmented with additional points with higher adiabatic energies, which give rise to a total of 10985 data points. The DPEM constructed in this work now enables accurate representation of adiabatic energies up to 18 eV. Quantum dynamic calculations based on this DPEM are nearly identical to those obtained from benchmark surfaces, which makes it the most accurate DPEM for the H3 system to date, therefore facilitating detailed exploration of reaction mechanisms at higher collision energies.
首页
登录
注册


DownLoad: