Introduction

Thymine, like other nucleobases, undergoes ultrafast radiationless relaxation back to the ground state after being excited by ultraviolet radiation. This property has been tied to the resilience of genetic material against light-induced damage1. However, the exact mechanism of this decay is not fully understood and has been a subject of debate for several decades. Gas phase experiments have identified at least two excited state decay channels, one with a lifetime on the order of ≲ 100 fs, and one considerably longer, on the order of several ps2,3,4,5. Yet, the underlying mechanisms have been challenging to discern, with proposed explanations necessarily relying on simulations of the molecular dynamics. These simulations, in turn, introduce approximations with errors that are difficult to control. Different theoretical methods, and in particular different electronic structure methods, have therefore produced different explanations, and a consensus has yet to emerge.

Most reported simulations implicate two low-lying excited states in the relaxation: a dark nπ* state (S1) and the bright ππ* state (S2) into which the system is initially excited. The interplay of a bright ππ* and a dark nπ* state is central to the excited-state relaxation of a wide variety of other chromophores, including nucleobases such as uracil and adenine, although lifetimes and branching ratios differ. In the case of thymine, several simulations predict a ππ* trapping channel in which the initial excitation to the ππ* state is rapidly followed by relaxation into a minimum on the ππ* surface, where the wavepacket is trapped for tens or hundreds of fs6,7,8,9,10,11,12. These simulations disagree, however, on the amount of ππ* trapping, as well as the timescale and nature of the subsequent processes, with proposed mechanisms including ππ* to nπ* relaxation11 and direct ππ* relaxation to the ground state8. Some of these studies indicate that the amount of ππ* trapping is reduced by improving the description of dynamical correlation (instantaneous electron-electron interactions)8,11, a pattern that has also been found in the closely related nucleobase uracil13. In line with this, an early density functional theory (DFT) study found a significant nπ* population within the first 50 fs14 and a more recent mixed-reference spin-flip DFT study also found rapid barrier-less ππ*/nπ* transfer (τ = 30 fs), including subsequent nπ* trapping15.

Experimental evidence has implicated the nπ* state in the early dynamics. Indeed, by determining the gas phase oxygen-edge time-resolved X-ray absorption spectrum, Wolf et al.16 found a fast component (τ = 60 ± 30 fs) which was attributed to population of the nπ* state. This was further corroborated in a recent time-resolved photoelectron spectrum reported by Miura et al.17 (τ = 39 ± 1 fs). Thus, the wavepacket appears to already transfer some of its population to the nπ* state within the first 100 fs. Furthermore, Wolf et al. found that the nπ* signature lasts for several ps, revealing a second relaxation mechanism. After passing through the ππ*/nπ* conical intersection, parts of the wavepacket appears to get trapped in a minimum on the nπ* surface. On the timescale of a few ps, there is evidence that thymine further undergoes intersystem crossing from the nπ* singlet to a ππ* triplet18. While experiments have thus shown that the nπ* state is involved in the early sub-100 fs dynamics, it remains an open question whether or not there is some trapping in the ππ* state19,20. More accurate simulations of the dynamics are therefore essential to unravel the precise relaxation mechanisms in thymine.

Here we present the highest-level wavepacket simulation on the early dynamics of gas phase thymine to date. To the best of our knowledge, this is also the simulation with the highest level of single-reference electronic structure theory performed on a molecular system of this size. Thanks to recent developments21,22,23,24,25,26, we were able to describe the electronic structure with the highly accurate coupled cluster singles and doubles (CCSD)27 method in its equation of motion formulation for excited states (EOM-CCSD)28. To the best of our knowledge, this method has not been applied in nonadiabatic dynamics before. Coupled cluster (CC) theory is well-known for effectively capturing dynamical correlation, but it has been widely regarded as unsuited for excited state dynamics due to the presence of numerical artifacts at conical intersections21,29,30,31. Recent work has shown that these problems can be removed with similarity constrained CC (SCC) theory22,23,26. Here, we apply the SCC with singles and doubles (EOM-SCCSD) method to simulate the first 100 fs after photoexcitation using ab initio multiple spawning (AIMS)32,33, showing that it is possible to simulate nonadiabatic dynamics with a coupled cluster method that correctly describes conical intersections. This demonstrates that coupled cluster theory is a viable electronic structure method for simulating a range of photochemical processes.

Results and discussion

The main photochemical pathway in our simulation is illustrated in Fig. 1. First, our simulation confirms an ultrafast ππ* to nπ* conversion. We find no significant ππ* trapping and no direct ππ* relaxation to the ground state in the first 100 fs. The population transfer from ππ* to nπ* is followed by trapping in an nπ* minimum on S1. This nπ* trapping channel constitutes the main (87%) photochemical channel of the simulation. In addition to this ultrafast ππ* to nπ* conversion, our simulation predicts a minor πσ*-mediated N-H dissociation channel (13%).

Fig. 1: Main photochemical pathway in the dynamics simulation.
figure 1

Potential energy surfaces are shown for the first and the second excited states (S1 and S2). Following photoexcitation to the bright ππ* state (S2), the simulation predicts two channels. The main channel is the nπ* trapping channel. Here, the wavepacket passes through the S1/S2 intersection and heads toward an nπ* minimum on the S1 surface. This minimum is reached in two ways, either by heading to the minimum directly (solid line) or by reaching it indirectly through a ππ* region on S1 (dashed line).

Full size image

The nπ* trapping channel can be experimentally identified from the gas phase time-resolved oxygen-edge X-ray absorption spectrum. In Fig. 2, we compare our simulated spectrum with the experimental X-ray absorption spectrum reported by Wolf et al.16. The most striking feature in these spectra is a bright signal, at around 526 eV, that grows in intensity in the first 50–70 fs. The theoretical and experimental spectra are in agreement for this feature (see panels C and D), which we find to be due to population transfer from ππ* to nπ*, in agreement with the mechanism proposed in ref. 16. By analyzing the simulation data, we find a characteristic time of τ = 41 ± 14 fs for this conversion (see Suppl. Note 1), which fits well with experimental estimates (60 ± 30 fs16 and 39 ± 1 fs17). In the simulated spectrum, we also find weak but visible features that are associated with the initial ππ* dynamics: in the first 20 fs, there are broad and diffuse features at around 526–528 eV and at 534 eV, where the 534 eV feature is partially hidden by the ground state bleach (see Fig. 2B). These features reflect the rapid movement of the wavepacket away from the Franck-Condon region and towards the S1/S2 conical intersection. Moreover, we find oscillations in the signal at 526 eV associated with dynamics on the nπ* state. These oscillations have a period of about 20 fs and an amplitude of about 1 eV (see panel B). The limited 100 fs FWHM time resolution of the experiment16 washes these oscillations out (see Fig. 2C), but we predict that they would be observed with improved time resolution. The oscillations are similarly washed out when we apply the same time broadening to the simulated spectrum (see Fig. 2D).

Fig. 2: Thymine oxygen-edge X-ray absorption spectrum.
figure 2

Panel A illustrates the pump-probe scheme for thymine, with an ultraviolet (UV) pump and an X-ray probe, and the numbering used for the important atoms in the dynamics. Panels B–D show oxygen-edge X-ray absorption spectra for thymine, simulated (B and D) and experimental (C)16. Color bars express the intensities in arbitrary units. Simulated spectra were computed using CC3/cc-pVDZ transition strengths and AIMS dynamics with CCSD/cc-pVDZ and SCCSD/cc-pVDZ, together with Gaussian broadening with a full width at half maximum (FWHM) of 0.3 eV/10 fs (panel B) and 0.3 eV/100 fs (panel D). For more details, see Suppl. Note 2. The most prominent feature in the spectrum is a bright signal at 526 eV, which is due to the nπ* state and reflects the rapid ππ*/nπ* internal conversion after photoexcitation to the ππ* state (at time t = 0). The oscillations in the nπ* feature (see panel B) are washed out by a larger time broadening (see panel D), reflecting the time resolution of the experiment (see panel C).

Full size image

The ππ*/nπ* conversion time of τ = 41 ± 14 fs was determined by analyzing the growth of the 526 eV signal in the simulated spectrum. This time constant is consistent with the rate of ππ*/nπ* conversion in the simulated dynamics, that is, from the observed change in electronic character from ππ* to nπ*. We find a rapid adiabatic population transfer from S2 to S1 (τ = 17 ± 1 fs) in our simulation. However, when the adiabatic states are decomposed into their diabatic components, and in particular into their ππ* and nπ* components, we see that the growth in the nπ* character (τ = 37 ± 9 fs) is in close agreement with the time constant determined from the simulated spectrum (τ  = 41 ± 14 fs). This shows that the 526 eV signal in the spectrum is due to the electronic nπ* character.

To better understand the dynamics, and in particular the nπ* trapping channel, we identify stationary points on the S1 and S2 surfaces in the vicinity of the Franck-Condon region. In agreement with previous calculations, we find an nπ* minimum on the S1 surface at an extended C4–O8 bond length11,16. This extension is due to the anti-bonding character of the π* orbital along the bond. We also find an S1/S2 minimum energy conical intersection (MECI) that can be reached from the Franck-Condon region through C5–C6 elongation. Close to such intersections, CCSD exhibits numerical artifacts (unphysical complex energies and distorted potential energy surfaces) which can be removed with SCCSD. Figure 3 details a simulated initial condition that encounters such artifacts and shows how the issue is averted with the SCCSD method.

Fig. 3: An example of a numerical artifact encountered during the dynamics with EOM-CCSD (panels A and B) and their correction with EOM-SCCSD (panels C and D).
figure 3

Panels A and C show a branching or gh plane for a conical intersection encountered during the simulation of one of the initial conditions. The positive and negative energy difference between the first and second excited states (S1 and S2) divided by two (ΔE/2) is shown. The g and h vectors are energy-rescaled and the displacements g and h are in arbitrary units. For CCSD (A), there is an ellipse of degeneracy in the gh plane with unphysical complex-valued energies in the interior of the elliptical boundary; for SCCSD (C), we instead see a single point of degeneracy and no unphysical energies. In CCSD simulations, the wavepacket may approach the intersection too closely and end up in the region with complex-valued energies (the interior of the ellipse in the gh plane). Whenever this happens, we re-run the simulation with SCCSD. Panels B and D show the corresponding potential energy curves for the center of a nuclear trajectory basis function with CCSD (B) and the same trajectory basis function with SCCSD (D). At 71 fs, the CCSD simulation enters the complex-valued region and is terminated (B). The SCCSD simulation, on the other hand, does not encounter any problems (D). The expansion point used in the branching plane calculation is the geometry with the smallest ΔE (as given by SCCSD) in the nonadiabatic event at 71 fs.

Full size image

By analyzing stationary points on S1 and S2, Wolf et al.16 suggested that the excited state decay of thymine follows a two-step process in the C5-C6 and C4-O8 coordinates (see Fig. 2A for atom labeling): following photoexcitation, the C5–C6 bond is first elongated, and along this stretching coordinate, the S1/S2 intersection seam is accessible; then, after interconversion to the S1 state, the C4–O8 bond elongates as the wavepacket heads towards the nπ* minimum on S1. This picture is borne out by our dynamics simulation. In Fig. 4, we show the time evolution of the nuclear density in the C5–C6 and C4–O8 bond coordinates, and we indeed see this two-step process unfolding in real time. For an initial condition simulated using an extended basis set (aug-cc-pVDZ), we also find that the initial dynamics follows the same path from the Franck-Condon region to the S1/S2 intersection region (see Suppl. Note 3).

Fig. 4: Total nuclear density snapshots in the C5-C6 and C4-O8 coordinates.
figure 4

Three important stationary points are shown: the ground state minimum \({S}_{0}^{\min }\), the minimum energy conical intersection S1/S2 MECI, and the S1 minimum \(n{\pi }_{\min }^{*}\). The wavepacket quickly moves away from the Franck-Condon region (\({S}_{0}^{\min }\)) and towards the minimum-energy conical intersection, where it starts transferring population to the nπ* surface (19 fs), eventually causing the wavepacket to split (38 fs). At longer times, after almost all of its population has transferred to the lower surface (58 fs), the wavepacket settles in the vicinity of a minimum on the nπ* surface (77 and 96 fs).

Full size image

The most intriguing result in the dynamics simulation is the presence of a hydrogen dissociation channel (13%). This finding comes as a surprise because no such dissociative channel has been previously reported. In fact, total kinetic energy release (TKER) spectra across a broad range of wavelengths, 270–230 nm, showed no signature of ultrafast N-H dissociation34, and recent experimental and theoretical investigations (see, e.g., refs. 11,15,18,19) have not invoked this pathway to explain the molecular dynamics. However, a similar decay mechanism has been proposed in adenine, where a dark πσ* state is accessed through interconversion from the ππ* state35,36. More generally, the importance of these low-lying dissociative πσ* states are well-recognized in a wide variety of heteroaromatic systems37,38,39,40. In adenine, the πσ* state is dissociative along the N-H bond coordinate and intersects with the ground state (S0) at larger N-H distances. Following excitation to the ππ* state, the wavepacket may reach this ππ*/πσ* conical intersection within 20 fs. Once on the repulsive πσ* state, N-H dissociation proceeds rapidly, resulting in a characteristically anisotropic TKER spectrum of H atom fragments41. Given the finding of a similar πσ* channel in our simulation on gas phase thymine, it may be that certain regions of the ππ* surface—in particular, regions accessed by a highly excited N-H stretching mode—lead to rapid relaxation through an intersection with a πσ* state, followed by hydrogen dissociation at the N-H bond.

Figure 5 characterizes this channel in terms of the normal mode associated with the N1-H stretch (see panel A). When thymine is displaced along the mode away from the equilibrium geometry (Q = 0), a dissociative πσ* state comes down and intersects with nπ* and ππ* states at around Q = 0.4 (see panel B), which corresponds to an N1-H bond length of around 1.3 Å. At longer N1-H bond lengths, the character of S1 is πσ* (see panel C). The two dissociating initial conditions both have a short initial N1-H bond length, and they both transfer population to S1 at long N1-H bond lengths (see panel D). The simplest explanation for the dissociative channel, therefore, is that initial conditions with short N1-H bonds experience a rapid extension of the bond in the initial phase of the nuclear dynamics, allowing them to access the πσ* state. This suggests that initial conditions with shorter N1-H bonds at t  = 0 would be more likely to dissociate. This indeed appears to be the case: out of 17 additional initial conditions specifically selected to have N1-H bond lengths shorter than 0.9 Å, we find that 7 of the conditions (41%) have access to parts of the potential energy surfaces with πσ* character (see Suppl. Note 4).

Fig. 5: The πσ*-mediated N-H dissociation pathway.
figure 5

Panel A illustrates the normal mode associated with the N-H stretch, where Q denotes the displacement (in Bohr) along this mode relative to the ground state equilibrium geometry (Q = 0 corresponds to the ground state (S0) minimum geometry). Panel B shows potential energy curves along this normal mode for the ground and first five excited states (S0, S1, S2, S3, S4, and S5), with the associated N-H bond length R(N1-H) given on the secondary axis. The shaded area denotes the values of Q for which we see spawning events in the dynamics. Panel C shows the natural transition orbitals of S1 for Q = 1.0, where the state is of πσ* character. Panel D shows the initial geometries (at time t = 0) and the spawning geometries (where population is transferred between S2 and S1), distinguishing between the dissociative geometries and the non-dissociative geometries. For each geometry, we give the N-H bond length and the projected value of the mode displacement Q. Dissociating trajectories have negative initial values of Q (short N-H distances) and spawn at high positive values of Q (long N-H distances), corresponding to regions of the potential energy surfaces where the πσ* state becomes accessible. Panel E shows the recorded ultraviolet gas phase spectrum for thymine with a comparison to a simulated coupled cluster singles, doubles and perturbative triples (CC3) spectrum (see Suppl. Note 5). The shaded area corresponds to wavelengths where total kinetic energy release (TKER) spectra have not been recorded.

Full size image

In the dynamics simulation, only 2 out of 16 initial conditions (13%) lead to hydrogen dissociation at the N1-H bond. Given the limited number of initial conditions, it is difficult to estimate how common this pathway is. Inspection of the natural transition orbitals (NTOs) of one of the conditions shows how the dissociation happens (see Suppl. Note 4). The other initial condition has similar behavior. Already at 2.5 fs, S3 is mainly of πσ* character, while S2 is still of ππ* character and S1 of nπ* character. As the wavepacket moves along the N1-H stretching coordinate, the πσ* state is stabilized, ultimately falling below the ππ* and nπ* states. The S3/S2 gap is about 0.5 eV at 2.4 fs, but by 3.4 fs, this gap has decreased to 0.1 eV. At the same time, S3 is of mainly πσ* character at 2.4 fs, but by 4.0 fs, the character of S2 and S3 has flipped. Eventually, at around 4.1 fs, there is an intersection between S1 and S2, where most of the population is transferred to S1. Moving away from the intersection, S1 is dominated by πσ* character. N1-H dissociation occurs rapidly over the next 15 fs, eventually reaching a ground state intersection at an extended N1-H bond length (greater than 2.0 Å).

Additional calculations support the predicted πσ* pathway. Both of the dissociative initial conditions display identical behavior when the S3 state is included in the dynamics simulation, showing that its inclusion does not suppress the channel (see Suppl. Note 6). Moreover, while the πσ* state has Rydberg character in the Franck-Condon region, this character decreases as the N1-H bond extends and the state becomes involved in the dynamics. Potential energy curves along the N1-H bond with a diffuse basis set (aug-cc-pVDZ) and with a higher-order correlation treatment (coupled cluster with perturbative triples, CC342) suggest that the pathway is still present with more accurate treatments (see Suppl. Note 7).

Current experimental data does not allow us to verify or exclude the existence of the dissociative channel. We do not find any clear signature of the πσ* channel in our simulated spectrum (see Fig. 2). However, the channel appears to be small (~13%), and an oxygen-edge spectrum is not expected to be highly sensitive to changes at the N1-H bond. A nitrogen-edge spectrum, on the other hand, should be sensitive to these changes, as indicated by calculations at selected geometries (see Suppl. Note 8). However, no experimental time-resolved nitrogen-edge spectrum has been reported.

Some experimental data have appeared to disconfirm the existence of a dissociative πσ* channel, as reported TKER spectra showed only smooth, isotropic H-atom kinetic energy, indicating no involvement of ultrafast, πσ*-mediated N-H dissociation34. However, these spectra only scanned over excitation wavelengths from 270–230 nm, and dissociation might become more common at shorter wavelengths. For adenine, the TKER spectra for wavelengths from 280–234 nm also show smooth, isotropic H-atom kinetic energies. Only for wavelengths 233 nm and shorter do anisotropic, fast H-atom peaks consistent with ultrafast πσ*-mediated N-H dissociation appear, and this signature becomes more intense with increasing excitation energy41. This behavior for adenine, coupled with the presence of the πσ* N-H dissociation in our dynamics, leads us to suggest that if the TKER experiments for thymine described by Schneider and coworkers34 are carried out using excitation wavelengths in the range 230–200 nm (as also suggested in ref. 40), then one might observe anisotropic, fast H-atom peaks consistent with this dissociative channel. However, wavelengths in the range 230–200 nm may not be directly comparable to our simulations as it would also excite the system to states above the S2(ππ*) state, in particular, to the ππ* band that lies about 1.0 eV above S2(ππ*) (see Fig. 5E and Suppl. Note 5). This is similar to adenine, where several states may contribute to the πσ* dissociation channel41. The utility of TKER experiments for confirming/denying the presence of the πσ* N-H dissociation channel in thymine has previously been suggested by others40,43.

The ultimate fate of the dissociative πσ* channel cannot be resolved in this work, as CC theory is known to break down when the πσ* state becomes degenerate with the ground state (the dissociation limit). Because of this, there may also be a pathway back to the ground state through an intersection with the ground state, which might compete with N-H bond fission40. However, the presence and accessibility of the πσ* channel is well-described in our simulations. Both the early-time involvement of the πσ* state, and the initial rapid N-H stretch, are treated correctly with CC theory, and this aspect of the simulated dynamics paints a picture in line with both theoretical and experimental studies of πσ* states in other heteroaromatics39,40.

Conclusion

The photorelaxation pathways of thymine are still under debate and a clear consensus has yet to emerge, despite numerous theoretical and experimental investigations. Here, we have simulated the ultrafast dynamics of thymine using high-level nonadiabatic dynamics simulations (ab initio multiple spawning) combined with coupled cluster theory (with single and double excitations) for the electronic structure. This represents, to the best of our knowledge, the highest level simulation performed on thymine to date, and the simulation with the highest level of single-reference electronic structure on a molecular system of this size. From the data, we simulated the oxygen-edge X-ray absorption spectrum of Wolf et al.16, obtaining excellent agreement on the appearance and position of a signal unequivocally associated with the nπ* state, as well as for the associated ππ*/nπ* conversion time: our theoretical estimate of 41 ± 14 fs agrees quantitatively with experimental estimates (60 ± 30 fs and 39 ± 1 fs16,17). Furthermore, we find no significant ππ* trapping.

Interestingly, our simulation predicts an additional, minor channel in which the population is rapidly transferred to a πσ* state, leading to N-H dissociation. This type of channel has been implicated in the ultrafast excited state dynamics of other heteroaromatic systems, including adenine39,40. In view of this surprising finding, we believe further experiments—for example, measuring the nitrogen-edge X-ray absorption spectrum—are warranted to confirm or disprove the existence of this dissociative channel as one of the relaxation pathways in thymine.

This work was made possible by recent developments in coupled cluster theory. Earlier work by some of the authors21,22,23 had already indicated that the method could be modified to correctly describe conical intersections and therefore also nonadiabatic dynamics. Here, we have shown that the method introduced in these papers, the similarity constrained coupled cluster method, can in fact be applied in nonadiabatic dynamics simulations, opening up a range of applications that may now be studied with the hierarchy of coupled cluster methods.

All electronic structure methods have limitations and this is also true for coupled cluster theory. As a single-reference method, it does not give a proper description of conical intersections with the ground state, and high-order treatments are needed to describe excited states with significant double excitation character. Nevertheless, given their ability to capture dynamical correlation and to provide systematically improvable predictions, we expect that coupled cluster methods will shed new light on the photochemistry of several important systems.

Methods

Theoretical

The excited state dynamics simulation was performed with the ab initio multiple spawning (AIMS) method32,33,44. We prepared 16 initial conditions (ICs) and simulated the dynamics for a total of 4000 au (~100 fs). The initial trajectory basis functions (TBFs) were sampled from a 0 K harmonic Wigner distribution obtained from the ground state equilibrium geometry at the CCSD/aug-cc-pVDZ level. We have not applied any sampling restriction to an excitation window. Both the geometry and the frequencies were obtained with CCSD/aug-cc-pVDZ. The dynamics simulation was performed at the CCSD/cc-pVDZ level. The initial positions and momenta of the individual samples are publicly available45. Details about the numbering of the trajectories are given in Suppl. Note 9. For the ICs, we adopt the independent first generation approximation and average the results over 16 independent AIMS simulations. In these simulations, the S1 and S2 states were included in the dynamics. Of these, 12 ran normally for the whole simulation. Two ICs (the two belonging to the πσ* pathway) reach a near-degeneracy between S0 and S1 as the N1H bond extends beyond 2 Å and are terminated at this point because CCSD cannot describe the dissociation limit (termination is caused by convergence problems at around 20 fs). A further two ICs enter a defective region surrounding the S1/S2 intersection, at which point the energies become complex-valued and the simulations are terminated. These two ICs were re-run with similarity constrained CC theory (EOM-SCCSD) using the \({{{{\mathscr{E}}}}}\) with T = 0 projection26. For both CCSD and SCCSD, the coupling elements were evaluated with the nuclear derivative acting on the right vector without normalization; for more details, see Refs. 25,26. In the time-integration, we have used a default timestep of 20 au (0.5 fs) and a smaller timestep of 5 au (0.1 fs) in regions of high coupling. The spawning threshold was set to 0.05, where the spawning criterion is given as the norm of the coupling times the velocity. This produced a total of 67 TBFs over the course of the simulation. Nuclear density snapshots were simulated with weights given by the squared TBF amplitudes and a 0.08 Å FWHM Gaussian broadening.

Using the data from the dynamics simulation, we simulated the oxygen-edge X-ray absorption spectrum using the CC342 method with the cc-pVDZ basis (additional spectra are reported in Suppl. Note 2). Core excited states were obtained with the core-valence separation approximation46. To simulate the spectrum, we have applied the incoherent approximation47, where the spectrum is calculated as an average of the spectra computed at the centers of the TBFs, weighted with the corresponding TBF amplitude. To avoid state assignment ambiguities, we have used CCSD for the valence excited states and CC3 for the core excited states, where the strengths are evaluated with CCSD (excluding the approximate triples in the CC3 states). Similarly, absorption energies are evaluated from energies obtained with CCSD and CC3 (see Suppl. Note 2). All X-ray absorption spectra were shifted by −0.5 eV, corresponding to the required shift needed to align the first ground state peak with the experimental value at 531.4 eV. The spectra were smoothed with Gaussian broadening (10 and 100 fs FWHM and 0.3 eV FWHM) to match experimental widths and uncertainties. Some of the spectra reported in Suppl. Note 2 made use of matplotlib’s48 imshow for Gaussian interpolations. The UV-vis absorption spectrum was calculated at the CC3/aug-cc-pVDZ+KBJ(3–4)49 level for the ground state equilibrium geometry, determined with CCSD/aug-cc-pVDZ, see Suppl. Note 5. We have further simulated the nitrogen-edge X-ray absorption spectrum at Franck-Condon, at the nπ* minimum, and at extended N-H (Q = 1.0), where we align the first ground state peak with experiment50, see Suppl. Note 8. In Suppl. Notes 10 and 11, we provide spawning geometries and the stationary geometries shown in Fig. 4, respectively.

Additional simulations were performed with S1 and S2 included in the dynamics for 17 initial conditions with short initial NH bond lengths (see Suppl. Note 4), for the two conditions in the πσ* pathway with S1, S2, and S3 included in the dynamics (see Suppl. Note 6), and for the initial S2 dynamics of one condition in nπ* pathway with aug-cc-pVDZ (see Suppl. Note 3).

All electronic structure calculations were performed using development versions of the eT program51. The AIMS dynamics was run with the FMS program and an interface to eT. See Suppl. Note 12 for a description of the interface.

Experimental

The experimental UV spectrum of thymine was taken with a Cary 5E UV-Vis-NIR spectrometer using an in-house developed gas cell described in ref. 52. The sample was purchased from Sigma-Aldrich with >99% purity and used without further refinement. The cell was heated to 150 °C to obtain sufficient absorption. Spectra were recorded over a range of 200–400 nm with a step size of 0.5 nm and an integration time of 0.5 s per data point. A background spectrum of the empty cell has been recorded at the same temperature and settings and subtracted from the spectra including the sample.