Flexible Framework for Surface Hopping: From Hybrid Schemes for Machine Learning to Benchmarkable Nonadiabatic Dynamics
Jakub Martinka, Mikołaj Martyka, Biman Medhi, Jiří Pittner, Pavlo O. Dral · 2025
Read Flexible Framework for Surface Hopping: From Hybrid Schemes for Machine Learning to Benchmarkable Nonadiabatic Dynamics by Jakub Martinka, Mikołaj Martyka, Biman Medhi, Jiří Pittner, Pavlo O. Dral (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-19152.txt.
Introduction
Jakub Martinka, Mikołaj Martyka, Biman Medhi, and Jiří Pittner present a flexible surface-hopping framework inside the MLatom software. Surface hopping simulates molecules that jump between energy surfaces after absorbing light, a core process in photochemistry. Their setup hosts Tully's fewest-switches method and a time-dependent Baeck-An variant.
Three examples show the workflow, from computing energies and couplings to following many trajectories at once. Custom user models can cut compute time and serve as fair tests for machine learning models. In their comparison, a Landau-Zener style scheme beat the Baeck-An variant. Analysis tools for single paths and whole ensembles round out the package.
Catalog record
| CREATOR | Jakub Martinka, Mikołaj Martyka, Biman Medhi, Jiří Pittner, Pavlo O. Dral |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-19152
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Flexible Framework for Surface Hopping: From Hybrid Schemes for Machine Learning to Benchmarkable Nonadiabatic Dynamics Jakub Martinka, Mikołaj Martyka, Biman Medhi, Jiří Pittner, Pavlo O. Dral Abstract Nonadiabatic molecular dynamics is a key technique for investigating a broad range of photochemical and photophysical processes. Among the established approaches, surface hopping schemes are widely used and can be easily integrated with various quantum chemistry programs or machine learning models. We present a flexible framework in MLatom that includes a newly implemented Tully’s fewest-switches surface hopping algorithm and its time-dependent Baeck–An variant. The capabilities of this framework are demonstrated through three representative examples corresponding to typical stages of a surface hopping study. First, we focus on methods providing energy, energy gradients and nonadiabatic couplings. We show that the flexibility of user-defined custom models can save computational time and that it is useful for benchmarking machine learning models. Next, we compare curvature-driven surface hopping schemes and show that Landau–Zener approach outperforms the time-dependent Baeck–An scheme. Finally, we showcase easy-to-use analysis tools for both individual trajectories and trajectory ensembles. This framework enables accelerated development of machine learning models and provides deeper insight into nonadiabatic dynamics. It is available as a part of the open-source MLatom package. keywords American Chemical Society, L a T e X † † affiliation: J. Heyrovský Institute of Physical Chemistry, Academy of Sciences of the Czech Republic, v.v.i. , Dolejškova 3, 18223 Prague 8, Czech Republic † † affiliation: Department of Physical and Macromolecular Chemistry, Faculty of Science, Charles University, Hlavova 8, 128 43 Prague 2, Czech Republic † † affiliation: Faculty of Chemistry, University of Warsaw, Pasteura 1, Warsaw, 02-093, Poland † † affiliation: Department of Chemistry, Indian Institute of Technology, Guwahati, Assam, India † † affiliation: J. Heyrovský Institute of Physical Chemistry, Academy of Sciences of the Czech Republic, v.v.i. , Dolejškova 3, 18223 Prague 8, Czech Republic † † email: [email protected] † † affiliation: State Key Laboratory of Physical Chemistry of Solid Surfaces, Department of Chemistry, College of Chemistry and Chemical Engineering, and Fujian Provincial Key Laboratory of Theoretical and Computational Chemistry † † affiliation: Institute of Physics, Faculty of Physics, Astronomy, and Informatics, Nicolaus Copernicus University in Toruń, ul. Grudziadzka 5, 87-100 Toruń, Poland † † affiliation: Aitomistic, Shenzhen 518000, China † † abbreviations: IR,NMR,UV 1 Introduction The computational tools are an essential part of chemical research. They allow deep insight into complex processes in atomistic systems, complementing experimental investigation. Unfortunately, every tool, regardless of whether it is experimental or theoretical, is inherently limited to a certain degree. One of the limiting factors is the level of theory that is used within a simulation. In a simplified view of quantum chemistry (QC), the more accurate method or larger basis set we choose, the more precise the result, and more computational time has to be spent. While performing simulations like molecular dynamics, one is therefore forced to sacrifice accuracy to be able to describe a large system or to reduce the time of performing the simulations. Machine learning (ML) helped push these limits by making fast predictions while maintaining the accuracy of the reference training data, and is now a well-established part of computational chemistry 24 . The MLatom package 22 is an open-source computational chemistry toolbox that facilitates this connection between QC methods, state-of-the-art ML models and simulations. It provides a wide range of different simulation techniques, e.g. geometry optimisation, reaction exploration, spectra calculation or molecular dynamics. A variety of interfaces to QC packages and ML models make it a particularly useful tool to study excited-state phenomena. Here, we present our latest development in the direction of nonadiabatic molecular dynamics (NAMD) simulations, particularly various schemes of Tully’s fewest-switches surface hopping (FSSH) algorithm, enlarging MLatom’s repertoire of simulation techniques. The FSSH is a popular tool for the investigation of photophysical and photochemical processes in molecular systems, especially due to its simplicity and scalability. It is based on separating the system into classical (nuclei) and quantum (electrons) degrees of freedom. For a given molecular geometry {\bm{R}}=\{{\bm{R}}_{A}\}_{A=1}^{N_{\mathrm{nucl}}} , the time-independent electronic Schrödinger equation is solved to obtain ground and excited state energies, which serve as a potential for the motion of nuclei obtained by integration of Newton’s equations of motion. This leads to a nuclear trajectory on an excited potential energy surface (PES), with the possibility of a stochastic ‘hop’ to a different PES. A large ensemble of these trajectories provides a statistical picture of population transfer between electronic states as well as configurational changes associated with such a process. Since its development in 1990s 85 ; 35 the large number of software performing FSSH simulations emerged, such as Newton-X 7 , SHARC 58 ; 57 , PyRAI2MD 49 ; 48 , PYXAID 1 , MNDO 26 , JADE 25 , ANT 77 , Libra 75 , ABIN 80 , FISH 69 , Hefei-NAMD 101 , and others. The application of ML methods to NAMD is a challenging task. At each classical step, FSSH requires the energies of excited electronic states, their gradients and nonadiabatic couplings (NACs) between them. The PESs are non-differentiable on the conical intersection (CI) seam, and the PESs in its vicinity must be accurately described by the ML model in order to correctly reproduce the excited-state deactivation process. Therefore, special care must be taken when assembling the training set, e.g. by employing active learning (AL) strategies 92 ; 49 ; 93 ; 63 . Energy gradients are typically obtained from the energy model as analytical derivatives, and they are commonly included in the training procedure. Last quantity to fit, and probably the most problematic one, is the first-order nonadiabatic coupling between states i and j defined as {\bm{d}}_{ij}=\bra{\Psi_i}\nabla_{\bm{R}}\ket{\Psi_j}=\frac{{\bm{h}}_{ij}}{E_{i}-E_{j}}\quad\forall\,i\neq j, (1) where {\bm{h}}_{ij}=\bra{\Psi_i}\nabla_{\bm{R}}\hat{H}_{\text{el}}\ket{\Psi_j} and \Psi_{i} represents an electronic state with corresponding energy E_{i} for geometry \bm{R} . The first difference is that energies are scalars, so they are rotationally invariant quantities, while NACs are of a vectorial nature. This means that NACs have to transform equivariantly with respect to the transformation of the input structure, posing a constraint for ML models. Next, when E_{i}-E_{j}\rightarrow 0 , see Eq. 1 , the NACs diverge to infinity. Last, due to the arbitrary phase of wavefunctions \Psi_{i} , the sign of NAC corresponding to one geometry \bm{R} is not well defined. This means that having a data set is not enough for ML to be successful; one has to phase-correct the data or train the models in a way to remove dependence on sign. The complicated nature of NACs and their unavailability for some QC methods led to the development of a numerical approach based on the overlap integral of electronic wave functions 35 . This approach has been used in FSSH simulations based on CC2, ADC(2), TDDFT, as well as semiempirical methods 30 ; 83 ; 70 ; 8 ; 73 ; 59 . However, it cannot be combined with ML models, and several schemes based on the topography of PESs, the so-called curvature-driven surface hopping schemes, have been developed. These schemes require only energies and energy gradients, reducing the problem to fitting the PESs. Nevertheless, fitting of excited PESs is still challenging due to the existence of the CIs, but it builds on the years of experience