Yinan Shu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yinan Shu has a total of 207 co-authors. Here are the top co-authors by number of shared publications:
- Donald G. Truhlar
Department of Chemistry, Chemical Theory Center, and Minnesota Supercomputing Institute, University of Minnesota
53 shared publications - Benjamin G. Levine
Stony Brook University
16 shared publications - Linyao Zhang
School of Energy Science and Engineering
13 shared publications - Zoltan Varga
University of Minnesota
13 shared publications - Dayou Zhang
University of Minnesota 207 Pleasant St SE
12 shared publications - Shaozeng Sun
Harbin Institute of Technology
7 shared publications
Latest Publications
- 7 October 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (20) , 10548-10564
Electronically Nonadiabatic Dynamics of O2 + O Collisions on Sixteen New Machine-Learned 3 A ″ Global Potential Energy Surfaces
High-energy O2 + O collisions on the 3A″ potential surface manifold of O3 provide a challenge for multistate nonadiabatic dynamics because covalent valence states and ionic configurations become strongly mixed in dynamically accessible regions. We developed a global machine-learned compatible potential-energy matrix (CPEM) for a 16-state nonadiabatic description of O2(X) + O(3P) dynamics with 3A″ symmetry, trained on XMS-CASPT2 data and augmented with physically motivated treatments of O2+ + O– ionic character where they are important. The adiabatic surfaces obtained by diagonalization have the analytic gradients and the correct conical-intersection topology, and they achieve good accuracy in both triatomic and O2 + O asymptotic limits. This establishes a general strategy for learning dense manifolds of coupled potential-energy surfaces with mixed covalent–ionic character. Using these surfaces with the asymptotically extended curvature-driven coherent switching with decay of mixing (AE-κCSDM) dynamics method, we computed electronically inelastic and reactive cross sections over a wide range of translational energies and O2 rovibrational states. The simulations reveal exchange reactivity, modest electronically nonadiabatic branching, and systematic trends in electronic-state-resolved excitation as functions of vibrational quantum number, rotational excitation, and dynamical energy.
- 7 October 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (20) , 10548-10564
Electronically Nonadiabatic Dynamics of O2 + O Collisions on Sixteen New Machine-Learned 3 A ″ Global Potential Energy Surfaces
High-energy O2 + O collisions on the 3A″ potential surface manifold of O3 provide a challenge for multistate nonadiabatic dynamics because covalent valence states and ionic configurations become strongly mixed in dynamically accessible regions. We developed a global machine-learned compatible potential-energy matrix (CPEM) for a 16-state nonadiabatic description of O2(X) + O(3P) dynamics with 3A″ symmetry, trained on XMS-CASPT2 data and augmented with physically motivated treatments of O2+ + O– ionic character where they are important. The adiabatic surfaces obtained by diagonalization have the analytic gradients and the correct conical-intersection topology, and they achieve good accuracy in both triatomic and O2 + O asymptotic limits. This establishes a general strategy for learning dense manifolds of coupled potential-energy surfaces with mixed covalent–ionic character. Using these surfaces with the asymptotically extended curvature-driven coherent switching with decay of mixing (AE-κCSDM) dynamics method, we computed electronically inelastic and reactive cross sections over a wide range of translational energies and O2 rovibrational states. The simulations reveal exchange reactivity, modest electronically nonadiabatic branching, and systematic trends in electronic-state-resolved excitation as functions of vibrational quantum number, rotational excitation, and dynamical energy.
- 22 September 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (19) , 9901-9910
ML26: New Ingredients for an Integral-Features Density Functional with Broad Chemical Accuracy
Integral-features density functional theory (IF-DFT) replaces the conventional integration over local ingredients with a nonlinear mapping from global integral descriptors, enabling machine-learned density functionals with broad transferability. Here we report ML26@MN15, a new IF functional that extends the earlier ML25@MN15 model by incorporating 12 additional correlation features─including overlap-projected rung-3.5 terms, CS1 dynamic-correlation ingredients, and VV10 nonlocal correlation─and by treating Hartree–Fock exchange as an integral feature rather than as a fixed hybrid percentage. ML26@MN15 employs 79 integral features evaluated on MN15 ingredients and a single-hidden-layer neural network with 5000 nodes to produce a size-extensive exchange–correlation energy. Trained on 185 databases comprising 7242 reference data, ML26@MN15 achieves a data-averaged energetic mean unsigned error of 0.89 kcal/mol, improving upon ML25@MN15 by 15% and outperforming a group of previous leading functionals across all eight chemical partitions examined. For the widely used MDB2019S, MGCDB84, and GMTKN55 data sets, ML26@MN15 yields the lowest averaged errors among all tested functionals, including the results for GMTKN55 by ML-based Skala-1.1 and DM21 density functional approximations. The new functional also maintains strong performance for systems containing heavy elements and exhibits improved behavior on a challenging self-interaction benchmark. These results show that expanding the integral-feature space with physically motivated nonlocal ingredients yields a density functional with unprecedented accuracy across diverse chemical benchmarks. The broad and consistent accuracy of ML26@MN15 highlights the promise of integral-features density functional approximations as a scalable framework for incorporating additional nonlocal physics into future functional development.
- 22 September 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (19) , 9901-9910
ML26: New Ingredients for an Integral-Features Density Functional with Broad Chemical Accuracy
Integral-features density functional theory (IF-DFT) replaces the conventional integration over local ingredients with a nonlinear mapping from global integral descriptors, enabling machine-learned density functionals with broad transferability. Here we report ML26@MN15, a new IF functional that extends the earlier ML25@MN15 model by incorporating 12 additional correlation features─including overlap-projected rung-3.5 terms, CS1 dynamic-correlation ingredients, and VV10 nonlocal correlation─and by treating Hartree–Fock exchange as an integral feature rather than as a fixed hybrid percentage. ML26@MN15 employs 79 integral features evaluated on MN15 ingredients and a single-hidden-layer neural network with 5000 nodes to produce a size-extensive exchange–correlation energy. Trained on 185 databases comprising 7242 reference data, ML26@MN15 achieves a data-averaged energetic mean unsigned error of 0.89 kcal/mol, improving upon ML25@MN15 by 15% and outperforming a group of previous leading functionals across all eight chemical partitions examined. For the widely used MDB2019S, MGCDB84, and GMTKN55 data sets, ML26@MN15 yields the lowest averaged errors among all tested functionals, including the results for GMTKN55 by ML-based Skala-1.1 and DM21 density functional approximations. The new functional also maintains strong performance for systems containing heavy elements and exhibits improved behavior on a challenging self-interaction benchmark. These results show that expanding the integral-feature space with physically motivated nonlocal ingredients yields a density functional with unprecedented accuracy across diverse chemical benchmarks. The broad and consistent accuracy of ML26@MN15 highlights the promise of integral-features density functional approximations as a scalable framework for incorporating additional nonlocal physics into future functional development.
- 28 August 2026
- Published by AIP Publishing in The Journal of Chemical Physics
- Vol. 165 (8)
Convenient implementation of rung-3.5 density functionals in standard Gaussian-basis codes
Developing more broadly accurate density functional approximations can be accomplished not only by new parameterizations with existing ingredients but also by using new ingredients encoding new physics, such as rung-3.5 energy densities that allow a better balance of static correlation and electron delocalization and also allow an explicit treatment of static correlation. Here, we introduce a convenient implementation of rung-3.5 energy densities as ingredients for new density functionals. With this implementation, overlap-projected rung-3.5 functionals can be implemented in any code that uses Gaussian basis functions by computing a set of auxiliary basis functions with the same form as Gaussian basis functions but having modified orbital exponents. We illustrate the method with M11po, a density functional constructed from the broadly accurate M11plus density functional by replacing all Coulomb-projected rung-3.5 ingredients with overlap-projected rung-3.5 ingredients, without reparameterization. M11po gives performance approaching that of the harder-to-implement M11plus functional, and it is freely available as an extension to the PySCF electronic structure package.
- 28 August 2026
- Published by AIP Publishing in The Journal of Chemical Physics
- Vol. 165 (8)
Convenient implementation of rung-3.5 density functionals in standard Gaussian-basis codes
Developing more broadly accurate density functional approximations can be accomplished not only by new parameterizations with existing ingredients but also by using new ingredients encoding new physics, such as rung-3.5 energy densities that allow a better balance of static correlation and electron delocalization and also allow an explicit treatment of static correlation. Here, we introduce a convenient implementation of rung-3.5 energy densities as ingredients for new density functionals. With this implementation, overlap-projected rung-3.5 functionals can be implemented in any code that uses Gaussian basis functions by computing a set of auxiliary basis functions with the same form as Gaussian basis functions but having modified orbital exponents. We illustrate the method with M11po, a density functional constructed from the broadly accurate M11plus density functional by replacing all Coulomb-projected rung-3.5 ingredients with overlap-projected rung-3.5 ingredients, without reparameterization. M11po gives performance approaching that of the harder-to-implement M11plus functional, and it is freely available as an extension to the PySCF electronic structure package.
- 25 June 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (13) , 6295-6303
Reinventing Density Functional Theory with Machine Learning on Integral Features
Kohn-Sham density functional theory remains central to electronic structure modeling, because of its efficiency and the steady gains in accuracy that have been achieved through improved density functionals. Here, we introduce a conceptually new strategy in which a multilayer perceptron learns a nonlinear functional of integrated local descriptors ("integral features"). In contrast to grid-wise neural-network approaches, the integral-features formulation applies the nonlinear model only once per calculation, preserving the computational efficiency of standard density functionals while enabling a substantially richer functional dependence. Using this framework, we develop ML25@MN15, an integral-features functional trained on 185 databases spanning thermochemistry, kinetics, noncovalent interactions, spin-flip energies, and molecular structures. Enabled by its expanded representational flexibility, ML25@MN15 achieves a mean unsigned error of 1.05 kcal/mol over 7232 energetic data and consistently outperforms leading modern functionals across all major categories, including systems containing transition metals. Because the neural network operates on integrated descriptors rather than grid-point values, the computational cost of ML25@MN15 remains comparable to MN15. These results demonstrate that introducing integral features and allowing the machine to determine the density functional's dependence on them yields a high-accuracy density functional without increasing computational cost.
- 25 June 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (13) , 6295-6303
Reinventing Density Functional Theory with Machine Learning on Integral Features
Kohn-Sham density functional theory remains central to electronic structure modeling, because of its efficiency and the steady gains in accuracy that have been achieved through improved density functionals. Here, we introduce a conceptually new strategy in which a multilayer perceptron learns a nonlinear functional of integrated local descriptors ("integral features"). In contrast to grid-wise neural-network approaches, the integral-features formulation applies the nonlinear model only once per calculation, preserving the computational efficiency of standard density functionals while enabling a substantially richer functional dependence. Using this framework, we develop ML25@MN15, an integral-features functional trained on 185 databases spanning thermochemistry, kinetics, noncovalent interactions, spin-flip energies, and molecular structures. Enabled by its expanded representational flexibility, ML25@MN15 achieves a mean unsigned error of 1.05 kcal/mol over 7232 energetic data and consistently outperforms leading modern functionals across all major categories, including systems containing transition metals. Because the neural network operates on integrated descriptors rather than grid-point values, the computational cost of ML25@MN15 remains comparable to MN15. These results demonstrate that introducing integral features and allowing the machine to determine the density functional's dependence on them yields a high-accuracy density functional without increasing computational cost.
- Open Access
- 22 June 2026
- Published by AIP Publishing in The Journal of Chemical Physics
- Vol. 164 (24)
Nonadiabatic dynamics starting from reactant and transition state in luminol chemiluminescence
The role of the starting configurations in nonadiabatic dynamics is investigated, and it is found for the first time that the dissociation times of the surface hopping trajectories initializing with the reactant are much greater than those starting from the transition state; however, the two sets of chemiexcitation yields are almost the same within standard errors in luminol chemiluminescence. The significant singlet chemiexcitation yield is explained by the difference between the number of S0 to S1 excitations and S1 to S0 de-excitations, and the negligible triplet can be explained similarly. The substantial singlet chemiexcitation yield obtained via the trajectory surface hopping method is attributed to the existence of the chemiexcitation region, through which the intrinsic reaction path does not pass.
- 22 June 2026
- Published by AIP Publishing in The Journal of Chemical Physics
- Vol. 164 (24)
Open AccessNonadiabatic dynamics starting from reactant and transition state in luminol chemiluminescence
The role of the starting configurations in nonadiabatic dynamics is investigated, and it is found for the first time that the dissociation times of the surface hopping trajectories initializing with the reactant are much greater than those starting from the transition state; however, the two sets of chemiexcitation yields are almost the same within standard errors in luminol chemiluminescence. The significant singlet chemiexcitation yield is explained by the difference between the number of S0 to S1 excitations and S1 to S0 de-excitations, and the negligible triplet can be explained similarly. The substantial singlet chemiexcitation yield obtained via the trajectory surface hopping method is attributed to the existence of the chemiexcitation region, through which the intrinsic reaction path does not pass.
- Open Access
- 12 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Reinventing Density Functional Theory with Machine Learning on Integral Features
Kohn–Sham density functional theory remains central to electronic structure modeling because of its efficiency and the steady gains in accuracy that has been achieved through improved density functionals. Here we introduce a conceptually new strategy in which a multilayer perceptron learns a nonlinear functional of integrated local descriptors (“integral features”). In contrast to grid-wise neural‑network approaches, the integral‑features formulation applies the nonlinear model only once per calculation, preserving the computational efficiency of standard density functionals while enabling a substantially richer functional dependence. Using this framework, we develop ML25@MN15, an integral-features functional trained on 185 databases spanning thermochemistry, kinetics, noncovalent interactions, spin-flip energies, and molecular structures. Enabled by its expanded representational flexibility, ML25@MN15 achieves a mean unsigned error of 1.05 kcal/mol over 7232 energetic data and consistently outperforms leading modern functionals across all major categories, including systems containing transition metals. Because the neural network operates on integrated descriptors rather than grid‑point values, the computational cost of ML25@MN15 remains comparable to MN15. These results demonstrate that introducing integral features and allowing the machine to determine the density functional’s dependence on them yields a high‑accuracy density functional without increasing computational cost.
- 12 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessReinventing Density Functional Theory with Machine Learning on Integral Features
Kohn–Sham density functional theory remains central to electronic structure modeling because of its efficiency and the steady gains in accuracy that has been achieved through improved density functionals. Here we introduce a conceptually new strategy in which a multilayer perceptron learns a nonlinear functional of integrated local descriptors (“integral features”). In contrast to grid-wise neural‑network approaches, the integral‑features formulation applies the nonlinear model only once per calculation, preserving the computational efficiency of standard density functionals while enabling a substantially richer functional dependence. Using this framework, we develop ML25@MN15, an integral-features functional trained on 185 databases spanning thermochemistry, kinetics, noncovalent interactions, spin-flip energies, and molecular structures. Enabled by its expanded representational flexibility, ML25@MN15 achieves a mean unsigned error of 1.05 kcal/mol over 7232 energetic data and consistently outperforms leading modern functionals across all major categories, including systems containing transition metals. Because the neural network operates on integrated descriptors rather than grid‑point values, the computational cost of ML25@MN15 remains comparable to MN15. These results demonstrate that introducing integral features and allowing the machine to determine the density functional’s dependence on them yields a high‑accuracy density functional without increasing computational cost.
- Open Access
- 8 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Reinventing Density Functional Theory with Machine-Learned Integral Features
Kohn–Sham density functional theory remains central to electronic structure modeling because of its efficiency and the steady gains in accuracy that has been achieved through improved density functionals. Here we introduce a conceptually new strategy in which a multilayer perceptron learns a nonlinear functional of integrated local descriptors (“integral features”). In contrast to pointwise neural-network approaches, the integral-features formulation applies the nonlinear model only once per calculation, preserving the computational efficiency of standard density functionals while enabling a substantially richer functional dependence. Using this framework, we develop ML25@MN15, an integrated-features functional trained on 185 databases spanning thermochemistry, kinetics, noncovalent interactions, spin-flip energies, and molecular structures. Enabled by its expanded representational flexibility, ML25@MN15 achieves a mean unsigned error of 1.05 kcal/mol over 7232 energetic data and consistently outperforms leading modern functionals across all major categories, including systems containing transition metals. Because the neural network operates on integrated descriptors rather than grid-point values, the computational cost of ML25@MN15 remains comparable to MN15. These results demonstrate that introducing integral features and allowing the machine to determine the density functional’s dependence on them yields a high-accuracy density functional without increasing computational cost.
- 8 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessReinventing Density Functional Theory with Machine-Learned Integral Features
Kohn–Sham density functional theory remains central to electronic structure modeling because of its efficiency and the steady gains in accuracy that has been achieved through improved density functionals. Here we introduce a conceptually new strategy in which a multilayer perceptron learns a nonlinear functional of integrated local descriptors (“integral features”). In contrast to pointwise neural-network approaches, the integral-features formulation applies the nonlinear model only once per calculation, preserving the computational efficiency of standard density functionals while enabling a substantially richer functional dependence. Using this framework, we develop ML25@MN15, an integrated-features functional trained on 185 databases spanning thermochemistry, kinetics, noncovalent interactions, spin-flip energies, and molecular structures. Enabled by its expanded representational flexibility, ML25@MN15 achieves a mean unsigned error of 1.05 kcal/mol over 7232 energetic data and consistently outperforms leading modern functionals across all major categories, including systems containing transition metals. Because the neural network operates on integrated descriptors rather than grid-point values, the computational cost of ML25@MN15 remains comparable to MN15. These results demonstrate that introducing integral features and allowing the machine to determine the density functional’s dependence on them yields a high-accuracy density functional without increasing computational cost.
- 13 May 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (10) , 4908-4921
Dynamics of Electronically Inelastic and Reactive Collisions of O( 1 D) with N 2 Based on Machine-Learned Coupled Potential Energy Surfaces
Electronically nonadiabatic transitions can play important roles in the pathways for reactive and energy-transfer processes, especially in high-energy collisions, and their accurate description is essential for many applications, including the simulation of thermal energy content and heat flux in shock-heated air. In the present work, we use extended multistate restricted-active-space second-order perturbation theory (XMS-RASPT2) with automatic learning via compatibilization by deep neural network (CDNN) to obtain the analytic representation of coupled potential energy surfaces (PESs) in the 11-state 1A″ manifold of N2O. The compatible potential energy surfaces are analytically tractable, with smooth and continuous behavior and with intersections in regions where the adiabatic curves display discontinuous derivatives and locally avoided crossings. We find that the convenient compatible matrix procedure is successful in globally fitting a dense manifold of states over a wide range of energies. The procedure yields not only a fit to the compatible potential energy matrix (CPEM) and its gradient but also analytic representations of the adiabatic potential energy surfaces and their gradients, including the correct topology of conical intersection seams. By combining the analytic representation of the present global PESs for an 11-state 1A″ manifold and a slightly modified previously determined representation of a 13-state 1A′ manifold of N2O, we have enough information to study electronically nonadiabatic and reactive collisions of O(1D) with by using the semiclassical asymptotically extended curvature-driven coherent switching with decay of mixing (AE-κCSDM) method, and we used this method to calculate cross sections for these collisions as functions of relative translational energy and initial vibrational and rotational energy. This reveals that reactive collisions dominate over electronically inelastic nonreactive collisions, with the most competitive electronically inelastic nonreactive channel leading to formation. The calculated cross sections of high-energy collisions of N2 with O(1D) can be used for simulating thermal energy content and heat flux in shock-heated air.- 13 May 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (10) , 4908-4921
Dynamics of Electronically Inelastic and Reactive Collisions of O( 1 D) with N 2 Based on Machine-Learned Coupled Potential Energy Surfaces
Electronically nonadiabatic transitions can play important roles in the pathways for reactive and energy-transfer processes, especially in high-energy collisions, and their accurate description is essential for many applications, including the simulation of thermal energy content and heat flux in shock-heated air. In the present work, we use extended multistate restricted-active-space second-order perturbation theory (XMS-RASPT2) with automatic learning via compatibilization by deep neural network (CDNN) to obtain the analytic representation of coupled potential energy surfaces (PESs) in the 11-state 1A″ manifold of N2O. The compatible potential energy surfaces are analytically tractable, with smooth and continuous behavior and with intersections in regions where the adiabatic curves display discontinuous derivatives and locally avoided crossings. We find that the convenient compatible matrix procedure is successful in globally fitting a dense manifold of states over a wide range of energies. The procedure yields not only a fit to the compatible potential energy matrix (CPEM) and its gradient but also analytic representations of the adiabatic potential energy surfaces and their gradients, including the correct topology of conical intersection seams. By combining the analytic representation of the present global PESs for an 11-state 1A″ manifold and a slightly modified previously determined representation of a 13-state 1A′ manifold of N2O, we have enough information to study electronically nonadiabatic and reactive collisions of O(1D) with by using the semiclassical asymptotically extended curvature-driven coherent switching with decay of mixing (AE-κCSDM) method, and we used this method to calculate cross sections for these collisions as functions of relative translational energy and initial vibrational and rotational energy. This reveals that reactive collisions dominate over electronically inelastic nonreactive collisions, with the most competitive electronically inelastic nonreactive channel leading to formation. The calculated cross sections of high-energy collisions of N2 with O(1D) can be used for simulating thermal energy content and heat flux in shock-heated air.- 22 April 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (9) , 4276-4287
New Framework for Multi-Electronic-State Dynamics and Its Validation for Photoisomerization of 1,3-Cyclohexadiene
We present a new paradigm for modeling electronically nonadiabatic molecular dynamics by extending the specific reaction parameter (SRP) approach to active space configuration interaction with single, double, and triple excitations in order to model coupled potential energy surfaces. Using the photoisomerization of 1,3-cyclohexadiene (CHD) as an example application, we develop ODM3.25(CHD), a system-specific reparameterization of the semiempirical ODM3.25 method trained against XMS-CASPT2 energies for the S0, S1, and S2 states at 1001 geometries. The resulting model achieves a mean unsigned error of 0.195 eV─comparable to the intrinsic accuracy limits of multireference perturbation theory─while automatically preserving (without diabatization) the correct topology of conical intersections through its configuration-interaction framework. Coupled with curvature-driven coherent switching with decay of mixing (κCSDM), ODM3.25(CHD) enables large-ensemble, 800 fs nonadiabatic trajectory simulations that reproduce key experimental and theoretical features of CHD photochemistry, including excited-state lifetimes of 60–70 fs, early passage through S1/S2 intersections, and ring-opening quantum yields of 43–53%. ODM3.25-SRP enables nonadiabatic dynamics of cyclohexadiene with XMS-CASPT2 quality and an expanded active space while reducing the computation time by a factor of 65. This work demonstrates that treating ODM3.25 as a domain-specific foundation model and fine-tuning it via SRP yields a computationally efficient, chemically accurate multistate potential capable of describing complex photodynamics at a fraction of the cost of high-level ab initio methods. This establishes SRP-based custom excited-state models as a powerful new tool for large-scale nonadiabatic dynamics.
- 22 April 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (9) , 4276-4287
New Framework for Multi-Electronic-State Dynamics and Its Validation for Photoisomerization of 1,3-Cyclohexadiene
We present a new paradigm for modeling electronically nonadiabatic molecular dynamics by extending the specific reaction parameter (SRP) approach to active space configuration interaction with single, double, and triple excitations in order to model coupled potential energy surfaces. Using the photoisomerization of 1,3-cyclohexadiene (CHD) as an example application, we develop ODM3.25(CHD), a system-specific reparameterization of the semiempirical ODM3.25 method trained against XMS-CASPT2 energies for the S0, S1, and S2 states at 1001 geometries. The resulting model achieves a mean unsigned error of 0.195 eV─comparable to the intrinsic accuracy limits of multireference perturbation theory─while automatically preserving (without diabatization) the correct topology of conical intersections through its configuration-interaction framework. Coupled with curvature-driven coherent switching with decay of mixing (κCSDM), ODM3.25(CHD) enables large-ensemble, 800 fs nonadiabatic trajectory simulations that reproduce key experimental and theoretical features of CHD photochemistry, including excited-state lifetimes of 60–70 fs, early passage through S1/S2 intersections, and ring-opening quantum yields of 43–53%. ODM3.25-SRP enables nonadiabatic dynamics of cyclohexadiene with XMS-CASPT2 quality and an expanded active space while reducing the computation time by a factor of 65. This work demonstrates that treating ODM3.25 as a domain-specific foundation model and fine-tuning it via SRP yields a computationally efficient, chemically accurate multistate potential capable of describing complex photodynamics at a fraction of the cost of high-level ab initio methods. This establishes SRP-based custom excited-state models as a powerful new tool for large-scale nonadiabatic dynamics.
- 14 February 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (5) , 2141-2155
Semiclassical Multi-State Dynamics of High-Energy O + O 2 Collisions: Influence of Electronic Excitation on Energy Relaxation
Electronic excitation is particularly significant in high-energy or high-radiation environments. In this study, we use mixed quantum-classical dynamics calculations to simulate O + O2 electronically nonadiabatic collision dynamics involving multiple electronic states. We calculate rate constants for electronic-vibrational energy transfer in collisions with 3A' symmetry, and we use them to calculate the effect of electronically nonadiabatic transitions on energy relaxation. We find that although electronic excitation quantitatively reduces the rate constants for vibrational energy transfer, the dependence of the electronic-vibrational energy transfer on the collision energy can be accounted for by an extension of the activation-saturation (AS) model previously proposed (Zhao, X. J. Chem. Phys. 2024, 161, 231101) for electronically adiabatic atom-diatom collisional energy transfer. Multi-electronic-state master equation calculations with electronic-vibrational energy transfer rate constants described by the vibronic AS (VAS) model show that for the high-temperature nonequilibrium energy relaxation process of O + O2, the electronically nonadiabatic effect increases the energy relaxation time by about 20%. By including electronically nonadiabatic transitions in the master equation, this research sheds light on the microscopic mechanisms of electron-vibration coupled energy transfer in high-temperature nonequilibrium flows, thereby providing guidance for energy relaxation prediction.
- 14 February 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (5) , 2141-2155
Semiclassical Multi-State Dynamics of High-Energy O + O 2 Collisions: Influence of Electronic Excitation on Energy Relaxation
Electronic excitation is particularly significant in high-energy or high-radiation environments. In this study, we use mixed quantum-classical dynamics calculations to simulate O + O2 electronically nonadiabatic collision dynamics involving multiple electronic states. We calculate rate constants for electronic-vibrational energy transfer in collisions with 3A' symmetry, and we use them to calculate the effect of electronically nonadiabatic transitions on energy relaxation. We find that although electronic excitation quantitatively reduces the rate constants for vibrational energy transfer, the dependence of the electronic-vibrational energy transfer on the collision energy can be accounted for by an extension of the activation-saturation (AS) model previously proposed (Zhao, X. J. Chem. Phys. 2024, 161, 231101) for electronically adiabatic atom-diatom collisional energy transfer. Multi-electronic-state master equation calculations with electronic-vibrational energy transfer rate constants described by the vibronic AS (VAS) model show that for the high-temperature nonequilibrium energy relaxation process of O + O2, the electronically nonadiabatic effect increases the energy relaxation time by about 20%. By including electronically nonadiabatic transitions in the master equation, this research sheds light on the microscopic mechanisms of electron-vibration coupled energy transfer in high-temperature nonequilibrium flows, thereby providing guidance for energy relaxation prediction.