Two-dimensional electronic spectroscopy in the condensed phase using equivariant transformer accelerated molecular dynamics simulations

Joseph Kelly, Frank Hu, Arianna Damiani, Michael S. Chen, Andrew Snider, Minjung Son, Angela Lee, Prachi Gupta, Andrés Montoya-Castillo, Tim J. Zuehlsdorff, Gabriela S. Schlau-Cohen, Christine M. Isborn, Thomas E. Markland · 2025

Read Two-dimensional electronic spectroscopy in the condensed phase using equivariant transformer accelerated molecular dynamics simulations by Joseph Kelly, Frank Hu, Arianna Damiani, Michael S. Chen, Andrew Snider, Minjung Son, Angela Lee, Prachi Gupta, Andrés Montoya-Castillo, Tim J. Zuehlsdorff, Gabriela S. Schlau-Cohen, Christine M. Isborn, Thomas E. Markland (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-2503-22583.txt.

Introduction

Joseph Kelly and twelve co-authors sped up in 2025 the simulation of a rich light-based probe of molecules. The probe, two-dimensional electronic spectroscopy, reveals how electronic states in molecules and materials talk to each other and to their setting. Atom-level simulations explain the signals but cost too much.

The team trained a symmetry-aware learning model on a few thousand quantum calculations, covering both ground and excited states. The learned energy surfaces then drove fast simulations that keep the probe's detail. Cheaper, faithful simulation opens the method to proteins and devices, not just small molecules.

Catalog record

CREATORJoseph Kelly, Frank Hu, Arianna Damiani, Michael S. Chen, Andrew Snider, Minjung Son, Angela Lee, Prachi Gupta, Andrés Montoya-Castillo, Tim J. Zuehlsdorff, Gabriela S. Schlau-Cohen, Christine M. Isborn, Thomas E. Markland
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE62 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-22583

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Two-dimensional electronic spectroscopy in the condensed phase using equivariant transformer accelerated molecular dynamics simulations Joseph Kelly, Frank Hu, Arianna Damiani, Michael S. Chen, Andrew Snider, Minjung Son, Angela Lee, Prachi Gupta, Andrés Montoya-Castillo, Tim J. Zuehlsdorff, Gabriela S. Schlau-Cohen, Christine M. Isborn, Thomas E. Markland Abstract Two-dimensional electronic spectroscopy (2DES) provides rich information about how the electronic states of molecules, proteins, and solid-state materials interact with each other and their surrounding environment. Atomistic molecular dynamics simulations offer an appealing route to uncover how nuclear motions mediate electronic energy relaxation and their manifestation in electronic spectroscopies, but are computationally expensive. Here we show that, by using an equivariant transformer-based machine learning architecture trained with only 2500 ground state and 100 excited state electronic structure calculations, one can construct accurate machine-learned potential energy surfaces for both the ground-state electronic surface and excited-state energy gap. We demonstrate the utility of this approach for simulating the dynamics of Nile blue in ethanol, where we experimentally validate and decompose the simulated 2DES to establish the nuclear motions of the chromophore and the solvent that couple to the excited state, connecting the spectroscopic signals to their molecular origin. Extracting information about the interplay between electronic and nuclear dynamics in complex condensed phase materials, ranging from light-harvesting complexes to quantum dots, is often greatly enhanced by going beyond linear electronic spectroscopy. Two-dimensional electronic spectroscopy [ 1 , 2 ] (2DES) is a nonlinear spectroscopic technique that provides access to features that encode population dynamics within and between states as well as the couplings of these states distributed across a range of excitation and emission frequencies. Characterizing these features and attributing them to specific molecular motions, particularly in condensed phase environments where these can be highly influenced by the structure and dynamics of the environment (e.g. solvent), is challenging because of the complexity of these systems. Understanding these observations has been deeply enriched by leveraging atomistic simulation. It is therefore important to develop efficient simulation methods that can accurately predict 2DES signals capturing phenomena such as anharmonic and vibronic effects, thermal broadening, and electronic interactions with the environment. Capturing these effects requires generating atomistic molecular dynamics (MD) on timescales of 50 to 100 picoseconds, along with the corresponding electronic excitation energies, to sample the nuclear fluctuations of both the chromophore and its environment. Assuming an ab initio treatment of both the dynamics and the environment, achieving this requires hundreds of thousands of ground and excited state electronic structure calculations for systems that may exceed 500 atoms owing to the desire to include potential charge transfer and polarization between the chromophore and its condensed phase environment. Such a large number of calculations is typically intractable at anything but the lowest levels of electronic structure theory, such as semiempirical methods. Recent work has suggested opportunities to use machine learning (ML) approaches to provide a tractable route to efficiently compute and understand electronic properties ranging from electronic excitation energies in the gas phase to condensed phase linear and 2D electronic spectra. For gas-phase systems these ML approaches have been used to predict orbital energy gaps [ 3 , 4 , 5 , 6 ] and excitation energies [ 7 , 8 , 9 , 10 , 11 , 12 ] for databases of molecules at their optimized geometries. To obtain gas phase linear spectra that incorporate the shape of the electronic absorption peaks arising from molecular geometries beyond their optimized structure, recent work has trained ML models to predict electronic energy gaps, i.e. the difference in energy between the electronic states, at multiple geometries and combined these with the ensemble approach [ 13 , 14 , 15 , 16 , 17 , 18 , 19 ] to compute linear spectra [ 20 , 21 , 22 , 23 , 24 , 25 , 26 ] . ML potential energy surfaces have also been trained to the quality required to perform nonadiabatic dynamics simulations to study electronic relaxation through conical intersections in gas phase systems [ 27 , 28 , 29 , 30 ] . Machine learned potentials also show considerable promise in condensed phase studies such as in the construction of model Hamiltonians for large networks of coupled chromophores [ 31 , 32 , 33 , 34 ] , calculating ensemble approach linear electronic absorption spectra by machine learning electronic energy gaps in implicit solvent [ 35 , 36 ] , and using electrostatic machine learning embedding to accurately capture environmental and vibronic effects [ 37 ] . Our recent work has sought to employ ML models to compute the excited state electronic energy gaps for chromophores in condensed phase environments enabling the combination of a dynamical treatment of linear and 2D electronic spectra via the second-order cumulant approach [ 38 , 13 , 39 ] , allowing for the inclusion of vibronic and anharmonic effects [ 39 ] , combined with an explicit treatment of the solvent [ 40 ] . This work has allowed us to show that one can train an accurate ML model of the electronic energy gap using only \sim 2000 energies [ 40 ] or just \sim 300 electronic energy gap gradients [ 41 ] . By exploiting transfer learning we have also shown that, starting from ML energy gap surfaces trained using computationally affordable excited-state electronic structure methods including time-dependent density-functional theory (TDDFT) or configuration interaction singles (CIS), one can adapt the model to the more accurate but more expensive equation-of-motion coupled cluster with singles and doubles (EOM-CCSD) level of theory using only 100s of energies [ 41 ] . This has allowed us to show that the explicit treatment of the hydrogen bonds (H-bonds) to the anionic green fluorescent protein chromophore in liquid water combined with a high-level wavefunction-based description of the excited state electronic structure is essential to capture the experimentally observed breadth of the linear electronic absorption spectrum and that this, in turn, leads to large changes in the 2DES signals [ 41 ] . Here we show that using an Equivariant Transformer (EqT) architecture [ 3 , 4 , 5 ] , which naturally encodes permutational, rotational, and translational symmetries within the machine-learned model, one can capture both the ground state electronic potential energy and electronic energy gap surfaces to simulate atomistic MD, the excitation energy gaps, and the corresponding linear and 2D electronic absorption spectra of a solvated chromophore using the energies and gradients obtained from only 2546 ground state and 90 excited state electronic structure calculations (SI Sec. IB-ID) compared to \sim 100,000 of both without the use of ML. We demonstrate this approach by simulating the electronic spectroscopy of Nile blue in ethanol and comparing it to recent 2DES experiments [ 1 ] . In doing so, we reveal how atomistic MD simulations can be used to unravel the rich information present in 2DES signals and highlight the additional understanding that one can extract about the role of the atomistic environment. The EqT approach offers significant improvements over our previous work as an EqT model trained with less than 100 excited state calculations can achieve a root mean squared error (RMSE) of less than 10 meV compared to 1000’s of training calculations achieving RMSEs of 50-80 meV

Read the complete text