WorksPhysics and Astronomy
Toward universal excited-state machine-learned potentials: a local excitation Hamiltonian for size consistency and transferability
Abstract
Universal machine-learned interatomic potentials could make excited-state simulations across chemical space routine, but excitation energies, unlike ground-state energies, are not sums of atomic contributions. The excited states of noninteracting molecules are the union of their states, which pooled multistate readouts reproduce at best for the lowest state. We introduce LEH-MACE, whose excitation energies are the eigenvalues of a local excitation Hamiltonian assembled from atom-centred blocks, so this union holds exactly. We generated TD-DFT energies and forces for about 300,000 equilibrium, thermal, and excited-state geometries of over 100,000 molecules. LEH-MACE's advantage over direct readouts grew with the distance from the training data and was largest above , with about half their and errors for long conjugated oligomers. For molecule pairs, only LEH-MACE kept its single-molecule accuracy, while a softmin readout remained superior for dynamics of larger chromophores. Exact size consistency is thus one step toward such potentials.
Introduction
Machine-learned interatomic potentials (MLIPs) reach the accuracy of their quantum-chemical reference methods at a small fraction of the cost and have become standard tools for simulations in the electronic ground state [1]. For electronically excited states, MLIPs can make photodynamics simulations of molecules orders of magnitude cheaper than on-the-fly electronic-structure calculations [2, 3, 4]. Since the first nonadiabatic dynamics simulations with MLIPs [5, 6, 7], excited-state MLIPs have enabled photodynamics on nanosecond time scales [8, 9], mechanistic studies of photochemical reactions [10], the virtual screening of photoswitches [11, 12], and studies of the photodegradation of materials for organic light-emitting diodes [13], with architectures that also predict the couplings between states [14, 15]. Most of these models are trained for one molecule or a small family of related molecules. A diabatic neural network trained on azobenzene derivatives transferred to azobenzenes with substitution patterns outside its training set and predicted their photoisomerization quantum yields [11], and models that transfer across chemical space more broadly have been proposed only recently [16, 17, 18, 19]. Such a model has to solve two problems that ground-state MLIPs do not face.
The first problem is the structure of the target. Ground-state energies are, to a good approximation, sums of local contributions, which makes atom-centred MLIPs transferable to larger systems. Excitation energies are not additive. For two noninteracting fragments A and B, the locally excited states of A+B are the union of the excited states of A and of B, so that and two identical fragments give exactly degenerate pairs of states. A multistate readout that sums atomic contributions, as used in many excited-state MLIPs [14, 17], produces excitation energies that grow with system size. Mean pooling is intensive but averages the fragments. Extremal pooling recovers the lowest excitation [20], and set-based readouts [18] and fragment corrections [19] have been proposed, but none of these readouts reproduces the union of the spectra for the whole low-lying spectrum.
Hamiltonian readouts have this property by construction, because the spectrum of a block-diagonal matrix is the union of the spectra of its blocks. Learned Hamiltonians have been used to predict orbital energies from a fixed-size pseudo-Hamiltonian [21], excitation energies of hydrocarbons from an atom-centred minimal-basis one-electron Hamiltonian followed by a simplified Tamm–Dancoff calculation [22], diabatic two-state Hamiltonians for photoswitches [11], exciton Hamiltonians of molecular aggregates built from physically defined local and charge-transfer states [23], and latent Hamiltonians for the excited states of individual molecules [24]. Whether a latent excitation Hamiltonian, assembled from local atom-centred blocks and trained only on adiabatic excitation energies and forces, transfers across chemical space and to larger molecules has not been tested.
The second problem is data. Large excited-state databases such as QM8 [25], QCDGE [26], and the data used for OMNI-P2x [19] cover hundreds of thousands of molecules, but mostly at equilibrium geometries and without excited-state forces. Datasets for excited-state dynamics, such as SHNITSEL [27] and WS22 [28], contain many geometries of a few molecules. Neither kind of dataset is suited to train and test a model that transfers across chemical space and remains accurate away from equilibrium, where excited-state dynamics takes place.
Here we address both problems. We introduce LEH-MACE, a local excitation Hamiltonian (LEH) predicted by a MACE [29] encoder: onsite blocks for every atom and coupling blocks for atom pairs within a cutoff form a symmetric matrix whose sorted eigenvalues are the singlet excitation energies (Figure 1). The predicted spectrum of separated molecules is therefore exactly the union of their spectra. We also present a TD-B97X-D/6-31G* dataset of 303,061 geometries of 106,933 molecules with ground- and excited-state energies and forces at equilibrium, thermal, and excited-state-sampled geometries, together with test sets of larger molecules, conjugated dyes, and ladder molecules with up to 30 heavy atoms. Of the readouts we compared (direct multistate readouts on the same encoder, X-MACE [18], and OMNI-P2x [19]), only LEH is exactly size consistent. Its advantage over the other readouts grew with the distance of the test molecules from the training data, from larger molecules to conjugated dyes and ladder molecules, and it was largest for pairs of molecules, where only LEH kept its single-molecule accuracy. For molecules similar to the training molecules, the direct readouts were slightly more accurate, and for the lowest excited state, the direct softmin readout remained competitive also beyond them. An active-learning loop along model-driven dynamics added 68,119 geometries in twisted and low-gap regions, on which we trained production models for excited-state dynamics.
Figure 1. Local excitation Hamiltonian and data. a, A MACE encoder computes atom features within the cutoff Å. Onsite blocks and pair blocks form a symmetric matrix , whose sorted eigenvalues are the excitation energies b, Two molecules A and B further apart than . is block diagonal, and its spectrum is the union of the spectra of A and B. c, Generation of the training data. QCDGE-derived molecules at equilibrium and thermal geometries and geometries from model-driven and dynamics are labelled with TD-B97X-D/6-31G* energies and forces (–, ten singlet excitation energies). An active-learning loop repeats the excited-state sampling for the production models.
Results
A local excitation Hamiltonian
LEH replaces the multistate readout of an MLIP by a learned symmetric matrix (Figure 1a). A MACE encoder computes features for every atom. An onsite network maps them to one diagonal element per atom, and a pair network gives a coupling for every pair of atoms closer than Å. The sorted eigenvalues of this matrix are the singlet excitation energies, the ground-state energy is a sum of atomic energies, and the excited-state forces follow from the derivatives of the eigenvalues (Methods). The matrix grows with the molecule, so a larger system has more low-lying states. For two molecules further apart than , the matrix is block diagonal, and its spectrum is exactly the union of the two spectra (Figure 1b). Readouts that pool over atoms, by summing, averaging, or taking a soft minimum, and models built on per-state atomic sums do not have this property.
An off-equilibrium excited-state dataset
Excited-state dynamics leaves the equilibrium geometry within femtoseconds, so a transferable model needs excited-state forces at distorted geometries of many molecules. We computed TD-B97X-D/6-31G* ground- and excited-state energies and forces (–, ten singlets) for 303,061 geometries of 106,933 molecules, 203,040 of them with excited-state forces (Figure 1c and Figure 2, Supplementary Note 1, Supplementary Fig. 1). Most molecules are derived from QCDGE [26]. Each of them contributes its QCDGE minimum and thermal geometries from ground-state xTB dynamics at 300–1000 K. Another 85,950 geometries come from model-driven and dynamics, from one round of sampling for the benchmark models and three active-learning rounds for the production models (Methods). The thermal geometries reach several eV above the minimum in but rarely small excitation energies, which the excited-state-sampled geometries cover (Figure 2b,Figure 2c). We trained the benchmark models of the following sections on the first version of the dataset, 168,587 geometries of 76,621 molecules with up to ten heavy atoms. The test sets contain random and unseen-scaffold molecules of the same size, 500 larger molecules with 11–20 heavy atoms, 5,018 conjugated dyes, and 1,393 ladder molecules with up to 30 heavy atoms (Figure 2a, Supplementary Fig. 2), none of which the benchmark models saw in training.
Figure 2. Dataset. a, Number of molecules versus the number of heavy atoms for the QCDGE-derived training molecules, the larger test molecules, the dyes (training and test scaffolds, the training scaffolds are part of the production training set only), and the ladder molecules (test only). b, Distribution of the excitation energy of the production training geometries by geometry type (number of geometries in brackets). c, Distribution of the energy above the minimum of the same molecule. d, Mean absolute error of the sampling models of active-learning rounds 2–4 on the geometries they selected, before labelling, versus the largest double-bond twist of the geometry.
Accuracy on held-out molecules
For test molecules similar to the training molecules, LEH, the direct softmin readout on the same encoder, and X-MACE [18] reached similar accuracy with the full training set, with – errors of 0.18 eV, 0.17 eV, and 0.17 eV for random molecules at thermal geometries (Figure 3a, Table 1). With fewer training data, LEH was the most accurate model, and for larger molecules it had the lowest error from 10k training molecules on (Figure 3b). With the full training set, the direct readouts had slightly lower errors for unseen scaffolds and -sampled geometries.
Figure 3. Learning curves. Mean absolute error of the – excitation energies (a, b) and of the forces (c, d) versus the number of training geometries, for random test molecules with up to 10 heavy atoms at thermal geometries (a, c) and larger test molecules with 11–20 heavy atoms (b, d). The training sets contain 1k–76.6k molecules with about 2.2 geometries per molecule. Markers show seed means and bars seed standard deviations. LEH, : encoder with higher angular resolution, full training set only. OMNI-P2x arch.: the OMNI-P2x architecture trained from scratch on our data. MS-ANI and the pretrained and fine-tuned OMNI-P2x are shown in Supplementary Figure 3. All models use the same splits and test sets.
| Model | – MAE (meV), random | scaffold | -sampl. | larger | forces (meV/), random | larger | (meV/atom), random | larger |
| LEH | ||||||||
| LEH, | 177 | 207 | 300 | 153 | 204 | 164 | 4.6 | 3.9 |
| Direct-softmin | ||||||||
| X-MACE | 168 | 186 | 268 | 208 | 272 | 353 | 4.0 | 12.1 |
| OMNI-P2x arch. | 447 | 462 | 421 | 774 | 448 | 356 | 10.7 | 8.2 |
Table 1. Errors of the models trained on the full training set. Mean absolute errors of the – excitation energies for random, unseen-scaffold, -sampled, and larger test molecules, of the forces, and of the energy. Mean standard deviation over seeds where available. Bold: lowest error per column. All targets with and the decomposition of the energy error into and excitation-energy errors are listed in Supplementary Table 2.
An encoder with higher angular resolution (, one seed) reduced the error of LEH for larger molecules to 0.15 eV, the lowest of all models, and for -sampled geometries to 0.30 eV (Table 1). The OMNI-P2x architecture trained from scratch on the same data reached 0.45 eV for random molecules, and MS-ANI, the same architecture without the level-of-theory input, 0.41 eV. The pretrained OMNI-P2x model fine-tuned with the protocol of its authors was less accurate for these molecules (0.60 eV) but more accurate than training from scratch for larger and conjugated molecules (Supplementary Note 3). The forces differed more between the models than the energies. LEH and the direct softmin readout predicted them with errors of 0.21 eV/ and 0.21 eV/, and X-MACE with 0.27 eV/ (Figure 3c,Figure 3d). The ground-state energies of LEH and Direct-softmin had errors of 5.1 meV per atom and 4.9 meV per atom for random molecules and 3.9 meV per atom and 3.7 meV per atom for larger molecules. X-MACE had the lowest ground-state error for random molecules (4.0 meV per atom) but three times the error of LEH for larger molecules (12.1 meV per atom, Table 1). At the equilibrium geometries, where the training data contain no excited-state forces, LEH predicted the forces, which set the initial direction of an excited-state relaxation, with errors of 0.17 eV/ and 0.13 eV/ for random and larger molecules (, X-MACE: 0.23 eV/ and 0.24 eV/, Supplementary Note 10). For LEH, Direct-softmin, and X-MACE, all errors decreased with the number of training geometries and did not reach a plateau with the full training set, whereas the errors of the OMNI-P2x architecture changed little beyond 10k training molecules. At a fixed number of training geometries, the accuracy did not depend on whether they came from more molecules or from more geometries per molecule (Supplementary Note 8).
Transfer beyond the training distribution
LEH and the direct readouts differed most when the test molecules differed from the training molecules (Figure 4a). The direct softmin readout and X-MACE were slightly more accurate than LEH for molecules similar to the training molecules, but their errors grew faster with the distance from the training distribution. For larger molecules, LEH reached 0.17 eV (Direct-softmin 0.19 eV, X-MACE 0.21 eV), for the 3,875 sulfur-free dyes, none of which these models saw in training, 0.29 eV (0.35 eV and 0.96 eV), and for ladder molecules 0.36 eV (0.61 eV and 0.77 eV). The difference increased with molecule size (Figure 4b,Figure 4c). The error of X-MACE increased most among these three models, because X-MACE decodes all excited-state energies from one sum of atomic features over the whole molecule, and this sum leaves the range seen in training for larger molecules (Methods, Supplementary Note 9). The OMNI-P2x architecture, which predicts each state as a sum of atomic energies, had the strongest size dependence. Fine-tuning the pretrained OMNI-P2x model reduced its errors for larger molecules, dyes, and ladder molecules to 0.62 eV, 0.88 eV, and 0.77 eV, which are still two to four times those of LEH.
Figure 4. Transferability. a, Mean absolute error of the – excitation energies of the models trained on the full training set, for test sets ordered by their distance from the training distribution. Shaded: test molecules similar to the training molecules. Dyes: all 3,875 sulfur-free dyes at their equilibrium geometries, none of which these models saw in training. Ladders: all sulfur-free ladder molecules at their equilibrium geometries. Dimers: two nonbonded test molecules at minimum distances of 2.5–15 Å, compared with TD-B97X-D dimer spectra (Figure 5). b, c, Mean absolute error of the – excitation energies versus the number of heavy atoms for dyes (b) and ladder molecules (c) at their equilibrium geometries, in bins of three heavy atoms. Errors relative to LEH for – and for alone are shown in Supplementary Figure 6, and MS-ANI, the pretrained OMNI-P2x, and the adapted fine-tuning protocol in Supplementary Figure 4. OMNI-P2x arch.: OMNI-P2x architecture trained from scratch on our data. OMNI-P2x, fine-tuned: pretrained OMNI-P2x fine-tuned on our data with the protocol of its authors.

Supplementary Figure 6. Mean absolute errors of Direct-softmin, X-MACE, and the OMNI-P2x architecture (trained from scratch) relative to LEH on the test sets of main-text Figure 4a, ordered by their distance from the training distribution. Values above 1 indicate a lower error of LEH. Shaded: test molecules similar to the training molecules. a, – excitation energies. b, excitation energy. Models trained on the full training set.
Most of the advantage of LEH lies in the states above . For and , LEH had errors of 0.24 eV for the ladder molecules and 0.24 eV for the held-out dyes, compared with 0.58 eV and 0.35 eV for the softmin readout. For alone, the two readouts had similar errors (ladder molecules: 0.60 eV and 0.66 eV, held-out dyes: 0.41 eV and 0.39 eV, Table 2, Supplementary Figure 6). This is consistent with a soft minimum over atomic contributions, which targets the lowest excitation but not the states above it. The production models, which we trained additionally on the active-learning data (Methods), showed the same pattern for the ladder molecules, with – errors of 0.18 eV versus 0.30 eV and errors of 0.43 eV versus 0.32 eV (LEH versus softmin readout, committee predictions, Table 2).
| Training data | Model | held-out dye scaffolds – MAE (meV) | held-out dye scaffolds MAE (meV) | held-out dye scaffolds | ladder molecules – MAE (meV) | ladder molecules MAE (meV) | ladder molecules |
| benchmark | LEH | 292±17 | 406±35 | 0.17±0.02 | 357±12 | 596±10 | 0.48±0.01 |
| Direct-softmin | 365±38 | 392±42 | 0.04±0.33 | 606±11 | 660±6 | 0.31±0.01 | |
| X-MACE | 1125 | 1543 | -8.71 | 772 | 891 | -0.54 | |
| OMNI-P2x arch. | 1821 | 1336 | -6.51 | 1182 | 1199 | -1.62 | |
| OMNI-P2x, fine-tuned | 976 | 657 | -1.16 | 767 | 679 | 0.21 | |
| + training dyes | LEH | 171±15 | 207±36 | 0.79±0.07 | 324±76 | 532±163 | 0.51±0.25 |
| Direct-softmin | 171±7 | 188±5 | 0.82±<0.01 | 492±14 | 492±<1 | 0.57±0.01 | |
| production: + training dyes + active learning | LEH | 171±7 | 192±6 | 0.81±0.02 | 283±58 | 442±141 | 0.64±0.18 |
| Direct-softmin | 160±11 | 173±13 | 0.86±0.02 | 313±28 | 325±26 | 0.81±0.02 | |
| LEH, committee | 152 | 172 | 0.85 | 266 | 434 | 0.67 | |
| Direct-softmin, committee | 148 | 164 | 0.88 | 307 | 321 | 0.82 |
Table 2. Conjugated molecules. Mean absolute errors of the – and excitation energies and of for held-out dye scaffolds and ladder molecules (sulfur-free) at their equilibrium geometries. The benchmark models were trained on molecules with up to ten heavy atoms, the dye-augmented models additionally on the training dye scaffolds, and the production models on the data of the dye-augmented models and all active-learning geometries with the weighted loss (Methods). No model was trained on ladder molecules. Mean ± standard deviation over seeds (two for the benchmark and dye-augmented models where available, four for the production models). Committee: seed-averaged prediction. OMNI-P2x, fine-tuned: average of the three fine-tuned ensemble members. Bold: lowest error or highest per column within each training-data group (for the production models, seed means and committees separately). Thermal geometries are listed in Supplementary Tables 7 and 11.
Figure 6. Excited-state dynamics. a, Mean absolute error of the excitation energy and b, of the change of the energy along the TD-B97X-D trajectories of seven validation molecules, for the sampling models after each active-learning round (small markers: the two models, line: mean, round 1: benchmark LEH) and for the production models (small markers: four seeds, large markers: mean). The dashed line marks the 0.2 eV threshold of the stopping criterion. c, Mean absolute error along the five test trajectories for the benchmark LEH before active learning and for the production LEH and Direct-softmin models (small markers: seeds, large outlined markers: committee). The structures of the test molecules are shown above d–h. d–h, excitation energy and i–m, energy relative to along the test trajectories, evaluated at the TD-B97X-D geometries up to the frame where TD-DFT breaks down (acrolein and uracil: first 300 fs). Production models: committee prediction (line) and range over the four seeds (band).
The benchmark models captured little of the variation of between the held-out dye scaffolds ( of 0.17 for LEH and 0.04 for the softmin readout, negative for X-MACE and both OMNI-P2x models, Table 2). Adding the training dye scaffolds to the training data (dye-augmented models) reduced the errors for held-out dye scaffolds to 0.17 eV for both LEH and the softmin readout and raised the of to 0.79 and 0.82, without changing the accuracy for small molecules. For the ladder molecules, which remained unseen, LEH was more accurate (0.32 eV versus 0.49 eV, Table 2).
Size consistency and noninteracting molecules
Only LEH kept its single-molecule accuracy when a second molecule was added (Figure 4a). For two molecules 10 apart, the levels that LEH predicted for the pair agreed within eV with the levels it predicted for the two molecules separately, and for two identical molecules every level appeared twice, as in TD-B97X-D (Figure 5a,Figure 5b). The pooled readouts shifted and spread the levels of the pair, and the per-state atomic sums of the OMNI-P2x architecture moved them up by several eV.
Figure 5. Separability. a, Excitation energies (–) of two different molecules A and B at a minimum interatomic distance of 10 Å, beyond the cutoff Å. For TD-B97X-D and each model, the left column (A, B) shows the levels of A and B computed separately, coloured by molecule, and the right column (A+B) the levels of the pair computed together. Solid lines connect each level of the pair with the matching level of the separate molecules (TD-B97X-D and LEH, dimer levels coloured by the molecule they belong to). Dashed lines connect the -th levels of the two columns (pooled readouts, dimer levels in the model colour). X-MACE predicts – only. Arrows mark levels above the plotted range. b, The same for two identical molecules (A + A, with A from panel a). Degenerate levels are drawn side by side. The pairs in a and b have the median and a near-median (fifth of ten) Direct-softmin error at 10 Å. LEH and Direct-softmin: seed 0. c, Mean absolute error of the dimer – excitation energies with respect to TD-B97X-D versus the minimum interatomic distance (180 dimers of ten different and ten identical pairs). The dashed line marks . d, Mean absolute error of the dimer excitation energies for , –, and –, for the pair computed together (filled) and for the union of the predictions for the two molecules computed separately (open). In c and d, models trained on the full training set, seed mean where two seeds exist. Readouts on the same encoder are compared in Supplementary Figure 7.
At short distances, the two molecules interact. TD-B97X-D shifts the dimer excitation energies relative to the union of the monomer spectra by 31 meV and 65 meV at 2.5 (different and identical molecules, the latter by excitonic splitting of degenerate states) and by only 4–10 meV at 5 . LEH reproduced the order of magnitude of these short-range shifts and is exactly separable beyond its cutoff. Its error for the 180 dimers (0.15 eV and 0.17 eV for different and identical pairs) was therefore no larger than for the same molecules alone (0.18 eV, Figure 5c). The other models were about as accurate as LEH for the single molecules (Direct-softmin 0.19 eV, X-MACE 0.17 eV) but not for the dimers (0.55 eV and 0.64 eV), at every distance up to 15 .
Per-state atomic sums predicted the excitation energy of a dimer approximately as the sum of those of its monomers (OMNI-P2x architecture: 5.98 eV, fine-tuned OMNI-P2x: 6.01 eV).
For Direct-softmin, the dimer error came mainly from the states above . Its dimer error was moderately larger than that of LEH (0.20 eV versus 0.13 eV), consistent with a soft minimum, which selects the lower of the two monomer excitations. Its – and – errors were several times larger than those of LEH (0.73 eV versus 0.17 eV and 0.93 eV versus 0.24 eV, Figure 5d). The union of its own predictions for the two molecules computed separately would have been as accurate as LEH (0.15 eV), so the error comes from the readout and not from the encoder. Readouts that pool by summing, averaging, or a soft minimum on the same encoder deviated from the union of their monomer spectra by 0.4–1.3 eV at all distances (Supplementary Figure 7).

Supplementary Figure 7. Mean absolute deviation of the predicted dimer excitation energies (–) from the union of the predicted monomer spectra versus the minimum interatomic distance, for pairs of different (a, c) and identical (b, d) test molecules. a, b, Readouts on the same encoder architecture. c, d, LEH and the baselines. All models were trained on the same 10k molecules (seed 0). Lines show the mean and bands the interquartile range over 60 pairs. Values below eV are shown at eV. Stars: deviation of the TD-B97X-D dimer spectrum from the union of the monomer spectra (ten pairs). Dotted vertical line: cutoff .
Public benchmark
On the public QCDGE dataset, we trained on equilibrium geometries of molecules with up to eight heavy atoms and tested on molecules with nine and ten heavy atoms (Table 3). LEH and the softmin readout were equally accurate within the training size range (0.090 eV and 0.092 eV) and for nine heavy atoms. For ten heavy atoms, LEH was slightly more accurate for the three lowest states (0.256 eV versus 0.265 eV) and by a larger margin for all ten states (0.204 eV versus 0.229 eV). For ten-heavy-atom molecules with scaffolds absent from training, the two readouts were equally accurate (0.291 eV and 0.285 eV). Sum and mean pooling were less accurate. The error of the OMNI-P2x architecture increased from 0.178 eV to 0.511 eV, and that of X-MACE, which decodes all states from one sum over the atoms, from 0.182 eV to 0.849 eV ( = 0.23). The step from nine to ten heavy atoms is not only a step in size. In QCDGE, the molecules with up to nine heavy atoms are derived from QM9 and those with ten heavy atoms from GDB-11, so the ten-heavy-atom molecules also differ in their chemistry. LEH trained on a random split of the same size, in which 83% of the training molecules have nine heavy atoms, reached 0.233 eV on the ten-heavy-atom test molecules, close to the 0.256 eV of the size split. Most of the error increase from nine to ten heavy atoms therefore comes from this change in chemistry and not from the larger size (Supplementary Note 13).
| – MAE (meV) | – MAE (meV) | – | ||||
| Heavy atoms | ≤ 8 (val.) | 9 | 10 | 10, new scaffolds | 10 | 10 |
| LEH | 90±1 | 115±1 | 256±3 | 291±4 | 204±1 | 0.90±<0.01 |
| Direct-softmin | 92±1 | 117±3 | 265±5 | 285±7 | 229±6 | 0.89±<0.01 |
| Direct-sum | 103±1 | 139±1 | 293±3 | 297±17 | 268±10 | 0.88±<0.01 |
| Direct-mean | 102±<1 | 134±2 | 280±6 | 301±13 | 244±4 | 0.89±<0.01 |
| X-MACE | 182 | 260 | 849 | 682 | 903 | 0.23 |
| OMNI-P2x arch. | 178 | 240 | 511 | 494 | 517 | 0.64 |
Table 3. Table 3: QCDGE size extrapolation. Mean absolute errors and of the excitation energies of models trained on QCDGE equilibrium geometries of molecules with up to eight heavy atoms (19,849 molecules, ten supervised excitation energies, no forces), for 5,000 test molecules each with nine and ten heavy atoms and for the ten-heavy-atom molecules with scaffolds absent from training. Mean ± standard deviation over two or three seeds (single runs without). per state, averaged over –. Bold: lowest error or highest per column. Random-split controls, errors, and the transition-dipole variant are listed in Supplementary Table 8.
Excited-state dynamics
Excited-state dynamics drives molecules into twisted geometries with small excitation energies, which thermal sampling and one round of excited-state sampling cover poorly. On the geometries selected in active-learning round 2, the error of the benchmark model increased from 0.37 eV below a double-bond twist of to 1.11 eV beyond (Figure 2d). Three active-learning rounds [30] along model-driven dynamics added 68,119 geometries of 16,381 molecules (Methods). The error of the sampling models on the geometries they selected decreased from 0.53 eV to 0.40 eV (Figure 2d). Their mean error along the TD-ωB97X-D trajectories of seven validation molecules, which never entered training, decreased from 0.53 eV to 0.27 eV (Figure 6a, Supplementary Note 14), and the error of the change of their energy from 0.20 eV to 0.11 eV (Figure 6b). Models trained from scratch on all data reached 0.17 eV on these trajectories, below the 0.2 eV threshold at which the loop stops. For the production models, we used a loss that gives more weight to geometries with small excitation energies, which reduced the overestimation of near / crossings (Methods). With four seeds each, the production LEH and Direct-softmin models reached 0.19±0.05 eV and 0.15±0.01 eV on the validation trajectories (Figure 6a) and errors of 0.09 eV and 0.07 eV for the change of the energy (Figure 6b). Their own dynamics from the validation minima was stable, with a total-energy drift of at most 7 meV over 1 ps and no bond changes.
We did not use the five test trajectories for any decision. Active learning reduced the mean error of LEH along these trajectories from 0.42 eV for the benchmark model to 0.29 eV (Direct-softmin: 0.15 eV, Figure 6c). Along an trajectory, the potential energy is the sum of the energy and the excitation energy. The excitation energy determines when a molecule reaches the crossing with (Figure 6d–Figure 6h), and the energy rises as the molecule moves away from the ground-state minimum (Figure 6i–Figure 6m). Both production models reproduced the excitation and energies of acrolein and uracil within 0.09 eV. For the three larger chromophores, trans-stilbene, HBDI, and DMABN, both models reproduced the increase of the energy by 1.8–3.7 eV within 0.08–0.25 eV. As the excitation energy decreased during relaxation and twisting, both models predicted it too high, on average by 0.11–0.22 eV (Direct-softmin) and 0.35–0.54 eV (LEH) for the committee predictions. All four LEH seeds overestimated it in the same way, so their disagreement does not indicate this error. At the same time, both models underestimated the rise of the energy, and the two errors largely cancelled (correlation coefficients between −0.97 and −0.99 for trans-stilbene and HBDI). The potential energy, which determines the forces in dynamics, therefore had smaller errors than the excitation energy (Direct-softmin: 0.08–0.15 eV versus 0.18–0.22 eV, LEH: 0.24–0.30 eV versus 0.36–0.54 eV), but both models kept and too far apart near the twisted geometries. For dynamics of single molecules larger than the training molecules, Direct-softmin is therefore the more accurate of the two production models.
Discussion
A latent excitation Hamiltonian assembled from local blocks gives an excited-state MLIP a property that pooled multistate readouts lack: its spectrum for noninteracting molecules is exactly the union of the spectra of the molecules. For pairs of molecules, LEH was three to four times more accurate than every direct readout. For single molecules that are larger or more extended than the training molecules, LEH transferred better than the direct readouts, with lower errors mainly for the states above . This advantage grew with the distance from the training distribution, where a model for photochemistry across chemical space is needed most. With up to 10k training molecules, LEH was the most accurate model, and its excited-state forces were as accurate as those of the best direct readout. With the full training set, the direct readouts had 0.01–0.06 eV lower errors for molecules similar to the training molecules. The remaining errors of 0.15–0.2 eV for the excitation energies of molecules similar to the training molecules are comparable to the error of TD-DFT with range-separated hybrids relative to high-level references, about 0.2–0.3 eV [31, 32, 33].
For the lowest excited state of a single molecule, the direct softmin readout is the stronger model. A soft minimum over atomic contributions targets the lowest excitation [20], which is the state that governs dynamics. The production Direct-softmin model was more accurate than LEH for along the test trajectories of the larger chromophores and for of long ladder molecules. This difference does not come from the weighted loss, which made LEH less accurate along the trajectories, because the models trained with the uniform loss showed the same order (0.28 eV versus 0.20 eV along the test trajectories). In both cases, LEH predicted too high an excitation energy where extended conjugation or twisting lowers it. Our results presented here do not yet indicate why the eigenvalue readout learns this lowering less well. More latent channels per atom, longer-range couplings, and active learning on larger chromophores are possible approaches for improvement. The two readouts thus have different use-cases: LEH is suited on the path towards a universal MLIP, where the spectrum above or more than one molecule matters, for example for absorption spectra, photophysical processes that involve higher-lying excited states [34], or chromophores in explicit solvent. Direct-softmin is suited for dynamics of single molecules.
Limitations and outlook. The latent Hamiltonian is size consistent only within the sector it represents, the union of local excitations of separated fragments. The short-range blocks do not represent charge-transfer states between fragments, double excitations, or long-range excitonic couplings. For the closed-shell organic molecules of this work, TD-B97X-D shifts the dimer spectra by less than 10 meV beyond 5 Å, well below the model error, but this will not hold for donor–acceptor complexes or molecular aggregates. A dipole–dipole coupling between distant channels would describe excitonic interactions in aggregates. The eigenvalues determine the Hamiltonian only up to orthogonal transformations. The learned couplings and eigenvectors therefore have no diabatic meaning, and the model provides no nonadiabatic couplings. Unlike a pooled readout, LEH can nevertheless be trained with information on its eigenvectors. With a loss on the TD-DFT excitation populations of – (10k training molecules), the weight shared by the eigenvector and the TD-DFT population on validation molecules increased from 0.67 to 0.78, and the – error of random test molecules decreased from 0.227 eV to 0.215 eV. For the ladder molecules, neither this agreement nor the error changed (0.62 eV and 0.62 eV, Supplementary Note 7). Transition dipoles built from the eigenvectors give oscillator strengths, which were accurate within but not beyond the training size range in our tests (Supplementary Note 13). The ground state is predicted by a separate extensive readout, to which the excitation energies are added. The topology of conical intersections is therefore not represented, whereas intersections between excited states arise from the eigenvalues of a smooth matrix. Linear-response TD-DFT itself describes intersections incorrectly [35], so an improvement would require other reference methods. A coupling of each eigenmode to the ground state, as in latent Hamiltonians that include the ground state [24], would represent intersections without giving up separability, and spin-flip or multireference data would allow training it. With one latent channel per atom, the model has as many states as atoms, which restricts diatomic and triatomic molecules to two or three states. More channels per atom remove this restriction. LEH and the direct readouts predicted negative excitation energies for a few highly strained cage molecules of the scaffold test set. Soft bounds on the LEH blocks limit these errors but do not remove them (production models: lowest of eV for LEH and eV for Direct-softmin), and the disagreement between seeds identifies them. Systematic errors, such as the too high excitation energies of the larger chromophores, were shared by all seeds of both architectures, so the disagreement between seeds is not a sufficient error estimate far from the training data.
Our results outline the steps toward a universal excited-state MLIP, comparable to the foundation models for the electronic ground state [36, 37]. The first step is a readout that is size consistent by construction. A ground-state MLIP predicts one extensive energy, whereas an excited-state MLIP predicts a set of states whose number, order, and character change with the size of the system. The latent Hamiltonian defines what a model should predict for local excitations of separated fragments, the union of their spectra, and a matrix with local blocks can still have eigenvectors that extend over a whole conjugated molecule. This readout transferred better to larger, conjugated, and noninteracting molecules, and fine-tuning a pretrained multistate model did not make up for a readout without this property. Extending the representation to charge-transfer states, states delocalized beyond the cutoff, and couplings between states with a physical meaning is the next step on the model side. The second step is data: The learning curves did not saturate, and power laws fitted to them reach an – error of 0.1 eV for random test molecules at 0.6 to 4 million training molecules, depending on the model, 8 to 49 times our training set. In our fixed-budget comparison, the number of labelled geometries limited the accuracy and not their distribution over molecules. Additional data are therefore most useful at geometries where the model is uncertain or where the dynamics goes, as in the excited-state sampling of this work, together with error estimates that also detect the errors shared by all seeds. Beyond the closed-shell molecules with C, H, N, O, and F and the singlet states of our dataset, a universal model needs other elements, charged and open-shell molecules, triplet states for intersystem crossing, and larger conjugated systems, for which our models had their largest errors. Because the model errors are already comparable to those of TD-DFT, the step beyond TD-DFT accuracy needs higher-level reference data, for example through transfer learning from many TD-DFT labels to fewer high-level labels, as for ground-state MLIPs [38]. We release the dataset, the code, and all trained models of this work as building blocks for these steps.
Methods
Local excitation Hamiltonian
For a molecule with atomic numbers and positions we attach latent channels to every atom ( in all models of this work), giving a matrix of size . A shared MACE encoder [29] produces invariant node features , obtained by concatenating the scalar () channels after each interaction layer. The internal features enter the many-body messages. The model assembles a real symmetric matrix from local blocks,
where and are multilayer perceptrons that output blocks, is a Bessel radial basis, is a smooth polynomial cutoff with radius , and is the mean supervised excitation energy of the training set. Its ascending eigenvalues are the vertical singlet excitation energies, and excited-state energies and forces follow as
where is predicted by an extensive atomic-energy readout on the same encoder. Away from degeneracies, (Hellmann–Feynman), which automatic differentiation of the symmetric eigendecomposition provides exactly. The blocks are rotation-invariant scalars, so for any rotation , and permuting atoms permutes rows and columns of , leaving the spectrum unchanged. is an effective excitation Hamiltonian. It is not a Kohn–Sham or many-electron Hamiltonian, and its eigenvectors carry no guaranteed diabatic meaning. Variable-size matrices in a batch are zero-padded to the largest and the padding rows receive distinct large diagonal values, so they never mix with or fall below physical eigenvalues. A batched eigendecomposition is then exact for each molecule.
Size consistency. For two fragments A and B separated by more than the encoder receptive field and , the features of A do not depend on B, and all pair blocks between A and B vanish. Hence
where is the multiset union. This holds exactly and gives and exact degeneracies for identical fragments, while stays additive. This is size consistency within the sector of local single excitations that the model represents. Charge-transfer states between distant fragments, double excitations, and long-range excitonic couplings are outside this sector. Pooled multistate readouts do not share this property: sum pooling is extensive, mean pooling averages the fragments, and per-state extremal pooling gives the correct but not the higher roots of the union.
Identifiability and bounded blocks. Eigenvalues determine only up to orthogonal transformations, so couplings and eigenvectors are not identified. This does not affect adiabatic energies and forces. Only the lowest of the eigenvalues are supervised, and the sorted assignment of to keeps unsupervised eigenvalues above at training geometries. Because nothing bounds the block elements, an out-of-distribution environment can produce a large coupling and drive eigenvalues far below zero. This happened for 4 of 87,396 test predictions of the two full-data LEH models (pair couplings of 20–32 eV in the affected geometries, compared with at most 2.4 eV in all other test geometries). Such failures can be detected from negative excitation energies and from disagreement between seeds, but they dominate when they occur. The production models therefore use soft bounds, with eV for onsite and eV for pair elements, which leave ordinary blocks practically unchanged. They limit such failures but do not remove them, because many bounded couplings can still add up. The direct readouts also predict negative excitation energies for the affected molecules, which are highly strained cases (Supplementary Note 11).
Encoder, readouts and training
All models of this work share a MACE encoder [29] with two interaction layers, 128 channels with hidden irreducible representations, spherical harmonics up to , correlation order 3, eight Bessel functions, and a cutoff of 5 Å, and differ only in the excitation readout. Besides LEH, we used three direct multistate readouts on the same encoder. An MLP maps each to an -vector, which is summed over atoms (Direct-sum, the readout of a conventional multistate MLLP), averaged (Direct-mean), or pooled per state with a learned soft minimum, with (Direct-softmin, in the spirit of extremal pooling [20]). For the direct readouts, the -th lowest output is the excitation energy. The loss combines the mean squared errors of the ten lowest excitation energies (sorted assignment, weight , ), of the per-atom energy (weight 100, , per-element reference energies from a least-squares fit to the training set), of the forces (weight 10, ), and of the – forces (weight 10), with excited-state forces from Eq. (3). Missing labels are masked. We trained in double precision with AdamW, batch size 16, and a learning rate of with 1000 warm-up updates and cosine decay to 1% over a fixed budget of 100k (up to 10k training molecules), 150k (30k), or 300k updates (full set), and used an exponential moving average of the weights (decay 0.99). We kept the checkpoint with the lowest validation loss. We chose these settings in single-seed LEH runs on 10k and 30k training molecules of the benchmark dataset (batch size, learning rate and schedule, 32 to 128 channels, or 2, and weight 10, 100, or 300), based on the errors and outliers on the test sets of these subsets, and report the encoder as a variant. The direct readouts used the same settings, and X-MACE, MS-ANI, and OMNI-P2x the settings recommended by their authors. Unless stated otherwise, errors of LEH and Direct-softmin are averaged over two (full set) or three seeds.
Production models. We trained the production models from scratch on the final dataset including all active-learning geometries (228,594 training geometries) with the settings above, 300k updates, soft bounds on the LEH blocks, and a weighted excitation-energy loss. With uniform weights, the production models overestimated on validation geometries with a TD-DFT below 1 eV by 0.73 eV on average, and similarly on training geometries. The uniform loss gives little weight to the geometries near crossings, which make up 2% of the training set (7% below 2 eV). The weighted loss multiplies the squared error of state of a geometry by , where is the TD-DFT excitation energy, for –, and for higher states, and normalises by the sum of the weights. The floor at 0.5 eV avoids emphasising geometries at which TD-DFT itself becomes unreliable. The weighted loss was motivated by the overestimation of near crossings by the uniform-loss models, and the decision itself used validation data only. We fixed the decision rule before the comparison runs finished and applied it to the mean over both architectures (two seeds each). The mean signed error of validation geometries below 2 eV had to decrease by at least 30%, and neither the validation – error nor the error along the validation trajectories was allowed to increase by more than 5%. The weighted loss met this rule, with changes of , , and . For Direct-softmin, all three criteria improved. For LEH, the low-gap error decreased by 38%, but the error along the validation trajectories increased by , mainly because of one of the two seeds. With four seeds, the weighted loss lowered the mean signed error of the validation geometries below 2 eV from 0.39 eV to 0.25 eV for LEH and from 0.43 eV to 0.22 eV for Direct-softmin and reduced the errors of both architectures on the active-learning test geometries. For LEH, the errors along the validation and test trajectories and for the ladder molecules were larger on average than with the uniform loss and varied more between seeds (Supplementary Table 10). We kept the weighted loss for both architectures as decided by the rule, because a choice per architecture after these results would have used test data. We trained four seeds of each production architecture and report the mean standard deviation over seeds and the committee prediction, the mean over the four seeds. The committee was 2–12% more accurate than the average seed on the test sets of individual geometries, but gained little along the test trajectories, where the seeds share their errors (Supplementary Table 11). We did not select seeds for reporting. The released default checkpoint of each architecture is the seed with the lowest validation – error.
Dataset
Electronic structure. We computed all labels with TURBOMOLE 7.5.1 [39] at the TD-B97X-D/6-31G* level [40, 41], the level of QCDGE [26], using full linear response (RPA), RI-J, grid m4, SCF convergence , density convergence , and response convergence . For each geometry, we computed and , the ten lowest singlet excitation energies with oscillator strengths, and analytic gradients of , , and [42], each in a separate egrad run with energy-ordered roots. If the Davidson solver did not converge, we repeated the run once with 200 iterations. We masked a force label if its root deviated by more than from the ten-root calculation or if the energy in the gradient file was inconsistent with . At QCDGE geometries, RPA reproduced the QCDGE excitation energies with a mean absolute deviation of 0.005 eV (20 molecules), whereas total energies differed by a nearly constant eV. We therefore reused the QCDGE excitation energies at QCDGE minima but recomputed and there.
Molecules and geometries. We drew molecules from all four QCDGE subsets (QM9-, GDB-11-, and PubChemQC-derived, up to ten heavy atoms, closed-shell molecules with C, H, N, O, and F) in a fixed, source-balanced random order, so the dataset grew in tranches without changing earlier entries. Every molecule contributes its QCDGE minimum and up to two thermal geometries from GFN2-xTB [43] NVT dynamics started at the minimum with a temperature drawn from 300–1000 K (one frame between 0.5 ps and 1.5 ps). We removed geometries whose bonding changed during the dynamics or whose energy lay below that of the minimum of the same molecule (a reaction). In addition, 12,869 training geometries and the corresponding test geometries come from one round of model-driven excited-state sampling (see below). We used this first version of the dataset, 222,172 geometries of 100,522 molecules, for all benchmark models. Three active-learning rounds (see below) added 68,119 geometries for the production models. Together with the conjugated test sets, the released dataset contains 303,061 geometries of 106,933 molecules (Supplementary Table 1).
Excited-state sampling. Thermal ground-state sampling rarely reaches the twisted or stretched geometries with small excitation energies that excited-state dynamics visits. For 10% of the molecules (hash-selected, all splits) and all larger test molecules, we ran NVE dynamics on (80%) or (20%) of a preliminary LEH model for up to 500 fs from the minimum with Maxwell–Boltzmann velocities (300–600 K). A trajectory stopped at a bond change, an excitation energy below 0.5 eV, or a non-finite energy. We labelled two frames per trajectory with TD-DFT, a random frame and the frame with the largest disagreement between two independently trained models.
Test sets and splits. We assigned molecules to training (85%), validation (5%), and test (10%) sets by a hash of the molecule identifier, with all geometries of a molecule in the same set. An additional balanced Bemis–Murcko scaffold hold-out [44] removes 10,003 molecules of small scaffold families from training (unseen-scaffold test set). Nested training subsets of 1k, 3k, 10k, 30k, and all 76,621 molecules (168,587 geometries) share one validation subset of 1,500 molecules and identical test sets. The larger test molecules are 500 neutral closed-shell molecules with C, H, N, O, and F and 11–20 heavy atoms from the ZINC250k subset [45] of ZINC [46] (50 per heavy-atom count), which we embedded with RDKit [47] and optimised with GFN2-xTB. Each has its minimum, one thermal geometry, and two -sampled geometries. The conjugated test molecules are 5,018 dyes from two experimental databases of organic chromophores [48, 49] and 1,393 ladder molecules (polyenes, polyacenes, oligophenylenes, and related series) with up to 30 heavy atoms, each at its minimum and one thermal geometry (12,770 geometries). We evaluated the sulfur-containing ones separately, because the main models never saw sulfur. For the dye-augmented models, we split the dye scaffolds into training (80%), validation (10%), and test (10%) families. We removed the molecules of all reference trajectories from the training and validation sets of the active-learning rounds and the production models by comparing covalent graphs. The benchmark training set contains three validation molecules (styrene, benzaldehyde, and cinnamaldehyde, at equilibrium and thermal geometries). The training set of the dye-augmented models, from which we fine-tuned the sampling models of rounds 3 and 4, additionally contains azobenzene, HBDI, and DMABN among the training dyes.
Active learning for the production models
Active learning, in which the models select new geometries for labelling, for example by committee disagreement, is an established way to extend the training data of MLIPs [8, 30]. Three further rounds of model-driven excited-state sampling (rounds 2–4) extended our dataset to the regions that excited-state dynamics reaches. In each round, two sampling models ran 1 ps of NVE dynamics on (80 %) or (20 %) of the first model, with Maxwell–Boltzmann velocities at 300–600 K. The trajectories started from the minima of all molecules with a photoisomerisable motif (an acyclic , , or bond conjugated with an aromatic ring or another double bond, 8,808 molecules), of all larger test molecules, and, from round 3 on, of a new hash-selected 5 % of all molecules. A trajectory stopped at a predicted excitation energy below 0.2 eV, a bond change, or a non-finite energy. We labelled up to two geometries per trajectory with TD-DFT, the one with the largest double-bond twist and the one with the largest disagreement between the two sampling models. For molecules without a motif, we took the most -relaxed geometry instead of the most twisted one. Geometries of validation and test molecules formed new test sets, so the earlier test sets remained unchanged. In round 2, we sampled with the two benchmark LEH models. We fine-tuned the sampling models of rounds 3 and 4 for 60k updates (learning rate , cosine decay, bounded blocks) on all data up to the previous round, starting from the dye-augmented LEH model and from the sampling models of the previous round, respectively.
After each round, we evaluated models trained on all data up to that round on seven validation molecules that never entered training or validation (styrene, azobenzene, cyclohex-2-enone, (E)-hexatriene, cytosine, benzaldehyde, and cinnamaldehyde). The loop stops when two conditions are met. (i) The dynamics of the models themselves from the validation minima (five trajectories of 1 ps at 300 K per molecule) is stable, with finite energies, no bond change, and a total-energy drift below 0.1 eV. (ii) The mean errors of the excitation energy and of the change of the energy along the TD-DFT trajectories of these molecules are at most 0.2 eV or improve by less than 10 % over the previous round, at the latest after round 4. Production models trained from scratch on all data after round 4 met both conditions, and the loop stopped (Supplementary Note 14).
Baselines
We trained and evaluated all baselines on the same splits, nested training subsets, and test sets as our models. Their predictions are put in adiabatic order, that is, the excited states are sorted by energy and the forces of each state are permuted with its energy. We computed the errors of all models with the same code.
X-MACE. We used the implementation of the authors [18] (AutoencoderExcitedMACE) with the settings of its documentation: 128 channels with , two interaction layers, correlation order 3, cutoff 5 Å, energy and force weights 100, learning rate , batch size 10, an exponential moving average of the weights (0.99), single precision, and 150 epochs. In this model, the energies of all states, including , are the sorted eigenvalues of a tridiagonal matrix, which a multilayer perceptron decodes from a 16-dimensional invariant summed over all atoms (a DeepSets autoencoder). The per-atom energy readouts of the encoder are not used. The eigendecomposition creates the cusps of adiabatic surfaces at state crossings, so the network itself only has to fit smooth quantities. Because the decoder acts on a molecule-level sum, the energies are not size extensive. For molecules larger than the training molecules, the predicted forces kept their direction (correlation 0.96–0.97 with the reference) but were systematically too small (least-squares slope 0.77 for 15–19 and 0.66 for 20 heavy atoms, compared with 0.99 for LEH, Supplementary Note 9). X-MACE masks forces per geometry rather than per state. We therefore trained it on –, the states with force labels, and it received no forces on the equilibrium geometries (45 % of the training set), whose excited-state forces are not part of the dataset. To test whether these label differences explain the comparison, we trained LEH on exactly the labels that X-MACE sees (three supervised states, the same force mask) with 30k training molecules. Its errors were equal to or lower than those of LEH trained on all labels (– MAE of 0.177 eV and 0.182 eV for random and larger test molecules, compared with 0.205 eV and 0.184 eV, and X-MACE 0.199 eV and 0.246 eV).
MS-ANI and OMNI-P2x. MS-ANI [17] extends ANI by appending the state index to the atomic environment vectors and sums per-atom energies for every state. OMNI-P2x [19] uses the same architecture with an additional one-hot input for the level of theory of the training labels and is pretrained on data from several levels of theory. We trained both with MLatom [50] (version 3.25.5) and its defaults, which are identical for the two models: atomic environment vectors with radial and angular cutoffs of 5.2 Å and 3.5 Å, element networks with 160, 128, and 96 neurons and CELU activations, a force weight of 0.1, and a weight of 0.01 on the gaps between adjacent states. When trained from scratch on labels from a single level of theory, the level-of-theory input is constant, so the OMNI-P2x architecture and MS-ANI are the same model up to initialisation and training noise. Their errors were similar for random test molecules (– MAE of 0.45 eV and 0.41 eV with all training data), and both failed for larger molecules (0.77 eV and 1.10 eV, Supplementary Figure 3). With the full training set, the MLatom training of both models stopped at the 46 h wall-time limit of our cluster (OMNI-P2x architecture: 393 epochs, compared with 934 and 1,495 epochs with 30k and 10k training molecules), and we used the checkpoint with the lowest validation loss at that point. In the main figures, we show the OMNI-P2x architecture trained from scratch, referred to as OMNI-P2x arch., and the pretrained OMNI-P2x fine-tuned with the protocol of its authors, referred to as OMNI-P2x, fine-tuned. We compared the pretrained OMNI-P2x model and two fine-tuning protocols in Supplementary Note 3. Without training on our data and with its excitation energies recalibrated linearly on our validation set, the pretrained model reached an – MAE of 0.63 eV for random test molecules, and its forces deviated by 1.2 eV/Å. Fine-tuning the three ensemble members with the protocol of the authors (100 epochs, first and third layer of every element network frozen, gap weight 0.01) gave 0.60 eV with all training data. In this network, the state index enters only the frozen first layer, so the state-dependent offset between the pretraining level of theory and ours cannot adapt. At 30k training molecules, the fine-tuned energies were 0.78 eV too low on the validation set. With all layers trainable and a gap weight of 1, the offsets disappeared and the error decreased to 0.45 eV, the accuracy of training from scratch, but the size transferability of the pretrained layers was lost (0.98 eV for larger molecules, compared with 0.63 eV with the protocol of the authors). For molecules similar to the training molecules, training from scratch therefore gave the most accurate OMNI-P2x model, and for larger and conjugated molecules, fine-tuning with the protocol of the authors.

Supplementary Figure 3. Learning curves as in main-text Figure 3, including MS-ANI, the pretrained OMNI-P2x (dotted, energies recalibrated on the validation set), and OMNI-P2x fine-tuned with the protocol of the authors and with the adapted protocol. Mean absolute error of the – excitation energies (a, b) and of the forces (c, d) for random test molecules (a, c) and larger test molecules (b, d).
Size-consistency test
We placed sixty pairs of different and sixty pairs of identical equilibrium test molecules at minimum interatomic distances of 2.5–20 Å, with molecule B randomly rotated and shifted. For each model, we compared the predicted dimer spectrum with the union of the predicted monomer spectra (–). For ten pairs of each kind, we computed TD-B97X-D dimer spectra (twelve roots) at 2.5–15 Å with the dataset protocol. These spectra contain the physical interaction and serve as the reference for the accuracy of the dimer predictions.
QCDGE benchmark
As a public benchmark without forces, we trained on QCDGE molecules with up to eight heavy atoms (19,849 molecules, validation 2,205) and tested on 5,000 molecules with nine and 5,000 with ten heavy atoms, including subsets with scaffolds not present in training. A random split of the same training-set size served as the in-distribution control. All models used the settings above with ten supervised excitation energies and no forces.
Excited-state dynamics
We computed reference adiabatic trajectories with TURBOMOLE at the level of the dataset, using velocity Verlet with a 0.5 fs step from the GFN2-xTB minimum with Maxwell–Boltzmann velocities at 300 K, six roots, and the gradient at every step. The trajectories ran for 500–1000 fs or until TD-DFT broke down near an crossing, which we detected as an excitation energy below 0.5 eV or a total-energy jump above 0.1 eV. The test trajectories, which we did not use for any decision, are acrolein (1 ps), uracil, trans-stilbene, the green-fluorescent-protein chromophore (Z)-HBDI, and 4-(dimethylamino)benzonitrile (DMABN, 500 fs each). The validation trajectories, which we used for the active-learning criterion and the choice of the loss only, are styrene, azobenzene, cyclohex-2-enone, (E)-hexatriene, cytosine, benzaldehyde, and cinnamaldehyde. We evaluated the models on every reference frame before TD-DFT breaks down and started model trajectories from the same initial conditions.
Excited-state forces at equilibrium geometries
The training data contain excited-state forces only at distorted geometries. To test the gradient at the Franck–Condon point, we computed – forces at the equilibrium geometries of 2,000 random and all 500 larger test molecules (test only). The recomputed energies agreed with the dataset to 0.002 meV.
Metrics
Excitation-energy errors are mean absolute errors (MAEs) over states and geometries in eV, reported for – (supervised by all models) unless stated otherwise. In the tables, energy errors are given in meV. Force errors are component-wise MAEs in ( in the tables), and energy errors are per atom. All predictions are in adiabatic order. We used the same evaluation code and test sets for all models, and Supplementary Table 2 lists all values with . Uncertainties are standard deviations over seeds. The sampling uncertainty of the test sets is of similar size or smaller. The 95 % intervals of the error of a single model from a bootstrap over molecules, with all geometries of a molecule resampled together, are ±0.002–0.01 eV for the test sets of thermal and -sampled geometries. We resampled the ladder molecules by homologous family and the trajectories as a whole, because the errors along a trajectory are correlated over 10–160 fs.
| Test set | Model | MAE (meV) | – MAE (meV) | – | MAE (meV) | (meV/atom) | (meV/Å) | (meV/Å) | |
| random, thermal | LEH | ||||||||
| LEH, | 191 | 0.96 | 177 | 0.97 | 172 | 4.6 | 73 | 204 | |
| Direct-softmin | |||||||||
| X-MACE | 167 | 0.97 | 168 | 0.97 | 140 | 4.0 | 103 | 272 | |
| OMNI-P2x arch. | 603 | 0.70 | 447 | 0.82 | 494 | 10.7 | 194 | 448 | |
| MS-ANI | 542 | 0.75 | 412 | 0.84 | 444 | 11.3 | 211 | 458 | |
| random, eq. | LEH | – | |||||||
| LEH, | 155 | 0.97 | 154 | 0.98 | 171 | 4.3 | 44 | – | |
| Direct-softmin | – | ||||||||
| X-MACE | 130 | 0.98 | 144 | 0.98 | 113 | 3.0 | 74 | – | |
| OMNI-P2x arch. | 599 | 0.76 | 438 | 0.85 | 509 | 9.0 | 114 | – | |
| MS-ANI | 548 | 0.79 | 404 | 0.87 | 478 | 9.5 | 125 | – | |
| scaffold, thermal | LEH | ||||||||
| LEH, | 235 | 0.90 | 207 | 0.93 | 212 | 5.6 | 87 | 245 | |
| Direct-softmin | |||||||||
| X-MACE | 195 | 0.96 | 186 | 0.97 | 161 | 4.7 | 120 | 301 | |
| OMNI-P2x arch. | 630 | 0.68 | 462 | 0.81 | 509 | 11.7 | 222 | 461 | |
| MS-ANI | 557 | 0.74 | 420 | 0.84 | 458 | 13.0 | 241 | 470 | |
| random, -sampled | LEH | ||||||||
| LEH, | 312 | 0.64 | 300 | 0.76 | 243 | 7.7 | 110 | 202 | |
| Direct-softmin | |||||||||
| X-MACE | 270 | 0.91 | 268 | 0.94 | 205 | 7.7 | 161 | 230 | |
| OMNI-P2x arch. | 449 | 0.80 | 421 | 0.86 | 359 | 13.5 | 257 | 352 | |
| MS-ANI | 469 | 0.79 | 429 | 0.86 | 374 | 13.6 | 265 | 360 | |
| larger, thermal | LEH | ||||||||
| LEH, | 139 | 0.96 | 153 | 0.97 | 153 | 3.9 | 69 | 164 | |
| Direct-softmin | |||||||||
| X-MACE | 214 | 0.91 | 208 | 0.93 | 371 | 12.1 | 272 | 353 | |
| OMNI-P2x arch. | 530 | 0.53 | 774 | 0.19 | 435 | 8.2 | 187 | 356 | |
| MS-ANI | 699 | 0.18 | 1096 | -0.53 | 584 | 8.5 | 204 | 368 | |
| larger, eq. | LEH | ||||||||
| LEH, | 151 | 0.96 | 158 | 0.97 | 164 | 4.8 | 46 | 134 | |
| Direct-softmin | |||||||||
| X-MACE | 326 | 0.80 | 325 | 0.85 | 984 | 35.2 | 158 | 241 | |
| OMNI-P2x arch. | 610 | 0.42 | 912 | 0.04 | 467 | 7.9 | 118 | 309 | |
| MS-ANI | 828 | -0.02 | 1286 | -0.80 | 670 | 8.1 | 131 | 314 | |
| larger, -sampled | LEH | ||||||||
| LEH, | 195 | 0.93 | 199 | 0.94 | 191 | 4.1 | 75 | 143 | |
| Direct-softmin | |||||||||
| X-MACE | 311 | 0.85 | 314 | 0.85 | 721 | 23.0 | 266 | 299 | |
| OMNI-P2x arch. | 773 | 0.24 | 1020 | -0.23 | 512 | 12.1 | 187 | 307 | |
| MS-ANI | 982 | -0.16 | 1368 | -1.11 | 700 | 11.9 | 198 | 315 |
Supplementary Table 2. Errors of the models trained on the full training set on all test sets. Mean absolute errors in meV, meV/Å, or meV per atom, and per state, averaged over the states. Mean ± standard deviation over seeds where available.
Use of large language models
Code, workflow scripts, analyses, and a first draft of the manuscript were written with the assistance of Claude Code by Anthropic under the direction of the author, who checked the results and revised the text.
Data availability
The dataset (geometries, energies and forces, ten singlet excitation energies with oscillator strengths, – forces, molecule identifiers, splits and all test sets, including all activelearning geometries) and the TURBOMOLE input and output files (without molecular-orbital files) of the 200,445 calculations made after we began archiving them, including all excited-state-sampled geometries, are available on Zenodo under a CC BY 4.0 licence (https://doi.org/10.5281/zenodo.24042143). The TD-B97X-D reference trajectories of the validation and test molecules are included with their initial conditions and a flag marking the frames after TD-DFT breaks down, together with the TD-DFT spectra of the molecule pairs and the excitation populations of Supplementary Note 7. QCDGE is available from its original publication [26].
Code availability
The LEH-MACE implementation, the checkpoints of all trained models (all seeds, with the seed with the lowest validation – error of each production architecture marked as the default), training configurations, analysis and figure scripts are available at https://github.com/aimat-lab/LEH under the MIT licence and archived on Zenodo (https://doi.org/10.5281/zenodo.24044613).
Acknowledgements
P.F. gratefully acknowledges the computing time made available on the high-performance computer HoreKa 2 at the National High-Performance Computing Center NHR@KIT. This center is jointly supported by the Federal Ministry of Research, Technology and Space and the state governments participating in the National High-Performance Computing (NHR) joint funding program (https://go-nhr.de/our-partners). P.F. acknowledges support by the state of Baden-Württemberg through bwHPC. P.F. acknowledges the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy for funding via the Excellence Cluster “3D Matter Made to Order” (EXC-2082/2-390761711). P.F. acknowledges funding from the Klaus Tschira Stiftung gGmbH (SIMPLAIX 2).
Author contributions
P.F. conceived the study, performed the experiments, analyzed the data, and wrote the manuscript.
Competing interests
The authors declare no competing interests.
Dataset
The released dataset contains 303,061 geometries of 106,933 molecules, 203,040 of them with – forces (Supplementary Table 1). Its first version (v1), on which we trained all benchmark models, contains 222,172 geometries of 100,522 molecules: 103,820 thermal geometries, 17,831 geometries from one round of excited-state sampling, the QCDGE minima, and the minima of the larger test molecules. The conjugated test sets add 12,770 geometries. Three active-learning rounds added 68,119 geometries of 16,381 molecules for the production models (Supplementary Note 14). Geometries of validation and test molecules from these rounds form separate test sets, so the v1 test sets are unchanged. The splits are by molecule. The training set contains 168,587 geometries of 76,621 molecules (12,869 of them -sampled), and the validation set contains 1,500 molecules. The test sets are random molecules (9,182 thermal geometries), molecules with Bemis–Murcko scaffolds [1] absent from training (10,003 molecules), 1,449 -sampled geometries of random test molecules, 500 larger molecules with 11–20 heavy atoms, 5,018 dyes, and 1,393 ladder molecules.
The QCDGE-derived molecules come from all four QCDGE subsets (QM9-, GDB-11-, and PubChemQC-derived, closed-shell molecules with C, H, N, O, and F and up to ten heavy atoms), and the larger test molecules from the ZINC250k subset of ZINC (11–20 heavy atoms, 50 per heavy-atom count). The dyes are taken from two databases of experimentally characterised organic chromophores [2, 3]. We selected neutral, closed-shell, single-component molecules with 11–30 heavy atoms, the elements H, C, N, O, F, and S, and a conjugated system of at least eight atoms. The ladder molecules form 13 homologous series (polyenes, diphenylpolyenes, oligophenylenes, phenylenevinylenes, oligofurans, oligopyrroles, oligothiophenes, azobenzenes, acenes, phenacenes, aza-aromatics, quinones, and classic dye cores) with up to 34 combinations of donor and acceptor end groups and up to 30 heavy atoms. Of all 106,933 molecules, 73,465 contain nitrogen, 80,687 oxygen, 12,625 fluorine, and 1,307 sulfur (dyes and ladder molecules only). Examples of the molecules of each set are drawn in Supplementary Figure S1 and Figure S2.

Figure S1. Random selections of the training molecules. a, QCDGE-derived molecules with up to eight heavy atoms, b, with nine heavy atoms, and c, with ten heavy atoms. d, Training molecules with - or -sampled geometries from excited-state sampling and active learning, drawn from their minimum geometries. e, Molecules of the seven validation (styrene to cinnamaldehyde) and five test (acrolein to DMABN) reference trajectories of dynamics. Hydrogen atoms are implicit. Molecules are sorted by the number of heavy atoms within each panel and drawn with the same bond length within a panel, except for molecules too large for their cell, which are drawn smaller.

Figure S2. Test and extension molecules. a, Random selection of molecules with Bemis–Murcko scaffolds absent from training. b, Random selection of larger molecules with 11–20 heavy atoms. c, Random selection of dyes (training and test scaffolds). d, One member of each of the 13 ladder series with hydrogen end groups, with three repeat units (two for the azobenzenes and phenylenevinylenes) or, for the acenes, phenacenes, quinones, aza-aromatics, and classic dye cores, the third member of the series. Hydrogen atoms are implicit. Molecules are sorted by the number of heavy atoms within each panel and drawn with the same bond length within a panel, except for molecules too large for their cell, which are drawn smaller.
| Geometry type | Geometries | Molecules | with – forces |
| QCDGE minima | 100,021 | 100,021 | 0 |
| thermal (GFN2-xTB, 300–1000 K) | 103,820 | 98,908 | 103,820 |
| excited-state sampling, round 1 | 17,831 | 9,256 | 17,831 |
| active learning, round 2 | 17,915 | 8,979 | 17,915 |
| active learning, round 3 | 25,122 | 12,765 | 25,122 |
| active learning, round 4 | 25,082 | 12,731 | 25,082 |
| minima of the larger test molecules | 500 | 500 | 500 |
| dyes and ladders, minima | 6,407 | 6,407 | 6,407 |
| dyes and ladders, thermal | 6,363 | 6,363 | 6,363 |
| total | 303,061 | 106,933 | 203,040 |
Supplementary Table 1. Composition of the released dataset by geometry type. The excitation energies at the QCDGE minima are taken from QCDGE (same level of theory). The energies and forces are recomputed for all geometries.
All targets of the models trained on the full training set
Supplementary Table 2 lists the mean absolute errors and of all targets, including the energy and forces and the total energy . The comparison of the error with the and excitation-energy errors decomposes the total error. For all models, the excitation energy contributes most of it.
Baselines: MS-ANI, pretrained and fine-tuned OMNI-P2x
MS-ANI and the OMNI-P2x architecture differ only by the one-hot level-of-theory input of OMNI-P2x, which is constant when the model is trained on a single level of theory (Methods).
Supplementary Figure 3–Figure 5 show both models together with the pretrained OMNI-P2x and two fine-tuning protocols.
Pretrained (zero-shot). We evaluated the three ensemble members of OMNI-P2x directly and averaged them. In the learning curves, the excitation energies are recalibrated linearly per state on our validation set, choosing the lowest error of no, constant-shift, and linear recalibration per test set. The forces and all other panels show raw predictions.
Fine-tuned, protocol of the authors. We fine-tuned each ensemble member separately on our training subset and validation set with the defaults of the OMNI-P2x fine-tuning in MLatom: 100 epochs, the first and third layer of every element network frozen, gap weight 0.01, force weight 0.1, and the energy shifter refitted to our level of theory. The state index enters only the frozen first layer. The remaining state-dependent offset ( too low by 0.78 eV, by 0.50 eV, and by 0.35 eV on the validation set at 30k training molecules) causes the irregular, size-independent errors for random test molecules in Supplementary Figure 3. A constant per-state shift fitted on the validation set removes most of it (from 0.67 eV to 0.38 eV for random test molecules at 30k) but does not transfer to
larger molecules.
Fine-tuned, adapted protocol. The same with all layers trainable and a gap weight of 1. The errors for random test molecules decrease smoothly with the training-set size and match training from scratch, whereas the errors for larger molecules and dyes increase (Supplementary Figure 3 and Supplementary Figure 4).

Supplementary Figure 4. Mean absolute error of the – excitation energies versus the number of heavy atoms, including MS–ANI, both fine-tuning protocols, and the pretrained OMNI-P2x (raw predictions). a, Random and larger test molecules at thermal geometries (shaded: training size range). b, Sulfur-free dyes at equilibrium geometries. c, Sulfur-free ladder molecules at equilibrium geometries. In b and c, bins of three heavy atoms.
The lowest excited state and the states above it
Supplementary Fig. 6 shows the errors of the baselines relative to LEH on the test sets of main-text Fig. 4a, for – and for alone. Beyond the training distribution, LEH was more accurate than Direct-softmin for – on every test set. For alone, the errors of the two models differed by a few per cent, except for ladder molecules and dimers, where LEH was more accurate (main-text Fig. 5d). X-MACE and the OMNI-P2x architecture were less accurate than LEH beyond the training distribution for both – and .
Size consistency of pooled readouts
Supplementary Figure 7 compares, for the same pairs as in main-text Figure 5, the dimer spectrum with the union of the monomer spectra predicted by the same model. All models in this figure were trained on the same 10k molecules. Panels a and b compare readouts on the same encoder architecture and settings, so that the readout is the only difference. TD-B97X-D shifts the spectrum by 31 meV and 65 meV at 2.5 Å (different and identical molecules), by 4 meV and 10 meV at 5 Å, and by 1 meV and 2 meV at 10 Å. LEH reproduced the size of the short-range shifts, and its deviation vanished beyond . Sum, mean, and softmin pooling deviated by 0.4–1.3 eV independent of the distance. Models built on per-state atomic sums predicted .
Ablations
Supplementary Figure 8 compares LEH variants and pooled readouts at 10k training molecules. Without pair blocks, that is, with a diagonal matrix of atomic excitation energies, the error increased on all test sets, so the coupling between atoms is necessary. Supervising six instead of ten states did not change the – error. Without -sampled training geometries, the error on -sampled test geometries increased from 0.38 eV to 0.50 eV, whereas the errors on thermal geometries did not change. Sum, mean, and softmin pooling were as accurate as LEH for molecules similar to the training molecules but less accurate for larger molecules, sum pooling by almost a factor of two.

Supplementary Figure 8. Mean absolute error of the – excitation energies of LEH variants and direct readouts on the same encoder architecture (10k training molecules, three seeds). a, Random test molecules, thermal geometries. b, Unseen scaffolds, thermal geometries. c, Random test molecules, -sampled geometries. d, Larger test molecules, thermal geometries. e, Larger test molecules, -sampled geometries. Small markers: seeds. Large markers: mean. Vertical line: mean of LEH.
Supervising the eigenvectors with TD-DFT excitation populations
A loss on the eigenvalues alone leaves the eigenvectors of undetermined, but unlike a pooled readout, LEH can be trained with information on its eigenvectors. We tested in the 10k setting of the ablations (Supplementary Note 6) whether such information is learned, whether it transfers to other molecules, and whether it changes the accuracy. For all 25,349 training and validation geometries of this setting, we computed the TD-B97X-D/6-31G* response vectors of the ten lowest singlet states with TURBOMOLE escf at the settings of the dataset (Methods). From the transition density matrix of state in the Löwdin-orthogonalised atomic-orbital basis, , we obtained the charge-transfer numbers [4], the hole and electron populations and , and the excitation population of atom A, each normalised to one over the atoms. In LEH, atom A carries the weight in eigenstate , which
also sums to one. For both quantities, we added the weight of each hydrogen atom to that of its nearest heavy atom. The population loss
runs over the pairs of a molecule in the batch and a state whose TD-DFT excitation energy is separated from those of the neighbouring roots by at least 0.1 eV. We added to the training loss with weight 1 (three seeds) and, with one seed each, with weight 0.1 and with a target that leaves the choice between hole and electron to the model, . The eigenvector gradients use a Lorentzian-broadened inverse of the eigenvalue differences (width eV), which is exact away from degeneracies and finite at them. Checkpoint selection used the validation loss without , and all other settings were those of the ablations. As the measure of agreement, we used the shared weight , which is 1 for identical distributions, for states whose LEH eigenvalue is also separated from its neighbours by at least 0.1 eV. Its chance level is the shared weight after permuting the LEH weights among the heavy atoms of the same element.
Without the population loss, the eigenvector shared 0.67 of its weight with the TD-DFT excitation population on the validation geometries (chance level 0.48). It followed the hole more closely than the electron (shared weights of 0.72 and 0.59). With the population loss, the shared weight on these geometries, which were not used for training, increased to 0.78 for , and it also increased for and (Supplementary Table 3). The – error of random test molecules decreased from 0.227 eV to 0.215 eV and the energy error from 7.6 meV to 7.1 meV per atom, and every seed trained with the population loss was more accurate on this test set than every seed trained without it. The errors on the other test sets of molecules similar in size to the training molecules decreased slightly as well. The agreement did not transfer to conjugated molecules. For the ladder molecules, the shared weight was 0.71 without and 0.70 with the population loss, and the eigenvectors of the ladder molecules with at least 22 heavy atoms remained more localised than the TD-DFT excitation populations. The error of the ladder molecules did not change (0.62 eV and 0.62 eV), and the change for the held-out dyes was within the spread between seeds. The weaker loss changed little, and the target that leaves the choice between hole and electron to the model gave an agreement between those of the baseline and the full loss. Two models trained with the population loss each predicted a large negative excitation energy for one test geometry (one of these geometries also for a baseline seed, Supplementary Note 11), which lowers but not the MAE of the affected test sets.
| Baseline | Population loss | Weight 0.1 | Hole or electron | |
| shared weight with the TD-DFT excitation population | ||||
| validation, | ||||
| validation, , chance level | ||||
| validation, | ||||
| validation, | ||||
| ladder molecules, | ||||
| dyes, | ||||
| errors (meV unless stated otherwise) | ||||
| random, thermal, – MAE | ||||
| random, thermal, – | ||||
| random, -sampled, – MAE | ||||
| unseen scaffolds, thermal, – MAE | ||||
| larger, thermal, – MAE | ||||
| random, thermal, energy MAE (meV per atom) | ||||
| random, thermal, force MAE (meV/Å) | ||||
| ladder molecules, equilibrium, MAE | ||||
| ladder molecules, equilibrium, | ||||
| held-out dyes, equilibrium, MAE | ||||
| held-out dyes, equilibrium, |
Supplementary Table 3. Population loss on the LEH eigenvectors (10k training molecules). Shared weight of the LEH eigenvector and the TD-DFT excitation population (median over geometries) on the validation geometries and on the 160 sulfur-free ladder molecules and 60 sulfur-free held-out dyes at equilibrium for which we computed TD-DFT populations. Errors on the test sets of the benchmark and on all sulfur-free ladder molecules and held-out dyes at equilibrium. Baseline and population loss (weight 1): mean standard deviation over three seeds. Weight 0.1 and hole-or-electron target: one seed.
Molecules or geometries?
At a fixed number of training geometries ( molecules, geometries per molecule), the error did not depend on . For random test molecules, the – errors were 0.25 eV, 0.25 eV, and 0.25 eV for 10k geometries with and 0.23 eV and 0.22 eV for 20k geometries with (Supplementary Fig. 9). The same holds for unseen scaffolds, larger molecules, and forces. Additional thermal geometries of known molecules are thus as valuable as new molecules.
X-MACE: force scaling with molecule size
For each heavy-atom bin of the test sets of thermal geometries, we fitted the predicted to the reference force components with a line through the origin. For X-MACE trained on all data, the slope of the forces is 0.97, 0.77, and 0.66 for larger molecules with 10–14, 15–19, and 20 heavy atoms (correlation 0.96–0.97), and 0.99 for LEH in all bins. A slope below one that decreases with size, at high correlation, is not noise. It is the expected behaviour of a nonlinear decoder that acts on a molecule-level sum outside its training range (Supplementary Table 4).
Label-matched control. X-MACE is trained on – only and uses forces only on geometries with all – forces. LEH trained on exactly these labels (30k molecules, two seeds) reached – MAEs of 0.177 eV, 0.182 eV, and 0.206 eV (random, larger, and scaffold test molecules at thermal geometries), force MAEs of 0.090 eV/Å and 0.084 eV/Å, and force MAEs of 0.230 eV/Å and 0.172 eV/Å (random and larger). The additional labels therefore do not explain the difference to X-MACE (Supplementary Table 5).
Excited-state forces at equilibrium geometries
The training set contains excited-state forces only at thermal and -sampled geometries. We computed TD-ωB97X-D – forces at the equilibrium geometries of 1,999 random and all 500 larger test molecules. LEH predicted the forces with MAEs of 0.17 eV/Å and 0.13 eV/Å (, seed 0), Direct-softmin with 0.17 eV/Å (random), and X-MACE with 0.23 eV/Å and 0.24 eV/Å (Supplementary Table 6).
Outliers and bounded blocks
Of 87,396 test predictions of the two full-data LEH models, 4 have an below −5 eV, all for different molecules and seeds (−32, −54, and −35 eV for two scaffold-test molecules with seed 0, −71 eV for one -sampled geometry with seed 1). In each of these geometries, single pair couplings reach 20–32 eV, compared with at most 2.4 eV in all other test geometries of the same models, and drive one eigenvalue far below zero. The other seed predicts finite but also too low energies for these geometries (by 1.2–1.6 eV), and the disagreement between seeds identifies them. The outliers hardly change the MAEs but dominate the of the affected test sets.
The affected scaffold-test molecules are highly strained cage molecules, two of them with a trivalent oxygen atom (for example CC1=C2[C(OH)]3C(=C[CH−1])[O+]23 in the bond orders perceived from the geometry), and they are out of distribution for every model. The direct softmin readout also predicts their too low, down to −1.3 eV for the benchmark models. Soft bounds on the blocks, with eV (onsite) and 5 eV (pair), which we used in the production models, limit the size of these errors but do not remove them, because the eigenvalues of a matrix with many bounded couplings can still lie far below zero. Of all 271,488 validation and test predictions of the production LEH models, 7 have eV (lowest −13 eV, compared with −71 eV without bounds), all for the same three molecules. The production Direct-softmin models have 3 such predictions (lowest −21 eV).
Supplementary Table 4: Slope (correlation) of predicted versus reference force components per heavy-atom bin for test molecules at thermal geometries (9 heavy atoms: random test molecules, 10–20: larger test molecules), for models trained on the full training set (seed 0).
| Forces | Model | 9 | 10–14 | 15–19 | 20 |
| LEH | 0.99 (1.00) | 0.99 (1.00) | 0.99 (1.00) | 0.99 (1.00) | |
| Direct-softmin | 0.99 (1.00) | 0.99 (1.00) | 0.99 (1.00) | 1.00 (1.00) | |
| X-MACE | 0.99 (0.99) | 0.97 (0.99) | 0.77 (0.97) | 0.66 (0.95) | |
| LEH | 0.96 (0.98) | 0.96 (0.97) | 0.96 (0.98) | 0.96 (0.96) | |
| Direct-softmin | 0.96 (0.98) | 0.96 (0.98) | 0.96 (0.98) | 0.97 (0.97) | |
| X-MACE | 0.93 (0.97) | 0.91 (0.96) | 0.76 (0.95) | 0.66 (0.92) |
Table 4.
Supplementary Table 5: Label-matched control with 30k training molecules, test molecules at thermal geometries. LEH trained on all labels, LEH trained on exactly the labels that X-MACE uses (– supervised, forces only on geometries with – force labels), and X-MACE. Mean ± standard deviation over two seeds where available.
| Model (30k molecules) | – MAE (meV) random | – MAE (meV) larger | – MAE (meV) scaffold | forces (meV/Å) random | forces (meV/Å) larger | forces (meV/Å) random | forces (meV/Å) larger |
| LEH, all labels | 205±1 | 184±3 | 239±2 | 98±<1 | 92±<1 | 236±1 | 181±2 |
| LEH, X-MACE’s labels | 177±2 | 182±1 | 206±2 | 90±<1 | 84±<1 | 230±<1 | 172±<1 |
| X-MACE | 199 | 246 | 221 | 129 | 284 | 312 | 372 |
Table 5.
Training on conjugated molecules
The dye-augmented models add the training dye scaffolds, that is, all dye molecules except the held-out scaffolds, to the benchmark training set. The ladder molecules remain unseen. On held-out dye scaffolds, the LEH error decreased from 0.29 eV to 0.17 eV and the softmin error from 0.37 eV to 0.17 eV. On ladder molecules, LEH reached 0.32 eV and softmin 0.49 eV. The errors on small molecules did not change (LEH, random test molecules at thermal geometries: 0.18 eV, larger molecules: 0.14 eV). Main-text Table 2 and Supplementary Table 7 list all values.
QCDGE benchmark: all variants
Supplementary Table 8 adds random-split controls, the errors, and the transition-dipole variant to main-text Table 3. With a random split of the same size, in which 83% of the training molecules have nine heavy atoms and none has ten, LEH reached 0.233 eV on the ten-heavy-atom test molecules (size split: 0.256 eV). In QCDGE, the molecules with up to nine heavy atoms are derived from QM9 and those with ten heavy atoms from GDB-11. Most of the error increase from nine to ten heavy atoms therefore reflects this change in chemistry and not the larger size. In the transition-dipole variant, each latent channel on atom carries a transition-dipole vector built from the bond vectors to its neighbours, , which is rotation equivariant. With the eigenvectors of , eigenstate has the transition dipole and the oscillator strength (atomic units). Trained additionally on oscillator strengths, this variant had the same excitation-energy errors as LEH. Its oscillator strengths reached in the training size range and 0.69 for nine heavy atoms but only 0.02 for ten heavy atoms (bright-state recall 0.42, 0.37, and 0.16). All main models are therefore trained on energies and forces only.
Supplementary Table 6: Excited-state forces at equilibrium geometries for models trained on the full training set. Mean absolute errors in meV/Å and over force components. Number of molecules in brackets. Mean ± standard deviation over seeds where available.
| Model | random (1,999) MAE | random (1,999) | larger (500) MAE | larger (500) | random MAE | random MAE |
| LEH | 168±1 | 0.92±<0.01 | 135±1 | 0.87±<0.01 | 273±<1 | 354±1 |
| LEH, | 161 | 0.92 | 134 | 0.87 | 265 | 349 |
| Direct-softmin | 171±2 | 0.92±<0.01 | 146±3 | 0.87±<0.01 | 293±1 | 364±1 |
| X-MACE | 227 | 0.87 | 241 | 0.69 | 333 | 390 |
| OMNI-P2x arch. | 413 | 0.49 | 309 | 0.53 | 444 | 470 |
| MS-ANI | 421 | 0.47 | 314 | 0.52 | 453 | 475 |
Table 6.
Supplementary Table 7: Benchmark and dye-augmented models on the conjugated test sets (sulfur-free held-out dye scaffolds and ladder molecules) at equilibrium and thermal geometries. Mean ± standard deviation over two seeds where available. OMNI-P2x, fine-tuned: average of the three fine-tuned ensemble members.
| Training data | Model | – MAE (meV) dyes eq. | – MAE (meV) dyes therm. | – MAE (meV) ladders eq. | – MAE (meV) ladders therm. | MAE (meV) dyes eq. | MAE (meV) dyes therm. | MAE (meV) ladders eq. | MAE (meV) ladders therm. | dyes eq. | dyes therm. | ladders eq. | ladders therm. |
| benchmark | LEH | 292±17 | 280±2 | 357±12 | 365±17 | 406±35 | 387±6 | 596±10 | 582±28 | 0.17±0.02 | 0.29±0.06 | 0.48±0.01 | 0.43±0.05 |
| benchmark | Direct-softmin | 365±38 | 389±46 | 606±11 | 553±15 | 392±42 | 408±37 | 660±6 | 622±12 | 0.04±0.33 | -0.19±0.46 | 0.31±0.01 | 0.29±0.01 |
| benchmark | X-MACE | 1125 | 755 | 772 | 453 | 1543 | 1005 | 891 | 503 | -8.71 | -3.71 | -0.54 | 0.43 |
| benchmark | OMNI-P2x arch. | 1821 | 1551 | 1182 | 1220 | 1336 | 1123 | 1199 | 1203 | -6.51 | -4.38 | -1.62 | -1.84 |
| benchmark | OMNI-P2x, fine-tuned | 976 | 718 | 767 | 909 | 657 | 536 | 679 | 926 | -1.16 | -0.38 | 0.21 | -0.52 |
| + training dyes | LEH | 171±15 | 178±13 | 324±76 | 332±68 | 207±36 | 214±35 | 532±163 | 545±141 | 0.79±0.07 | 0.76±0.07 | 0.51±0.25 | 0.44±0.24 |
| + training dyes | Direct-softmin | 171±7 | 179±5 | 492±14 | 461±24 | 188±5 | 195±2 | 492±<1 | 489±6 | 0.82±<0.01 | 0.80±<0.01 | 0.57±0.01 | 0.52±0.01 |
Table 7.
Supplementary Table 8: QCDGE size extrapolation, all variants. : oscillator strength of (transition-dipole variant only). Mean ± standard deviation over seeds where available.
| Heavy atoms | – MAE (meV) ≤ 8 (val.) | – MAE (meV) 9 | – MAE (meV) 10 | – MAE (meV) 10, new scaf. | – 10 | – 10 | MAE 10 | 10 | ≤ 8 | 10 |
| LEH | 90±1 | 115±1 | 256±3 | 291±4 | 204±1 | 0.90±<0.01 | 264±4 | 0.90±<0.01 | – | – |
| Direct-softmin | 92±1 | 117±3 | 265±5 | 285±7 | 229±6 | 0.89±<0.01 | 277±5 | 0.89±<0.01 | – | – |
| Direct-sum | 103±1 | 139±1 | 293±3 | 297±17 | 268±10 | 0.88±<0.01 | 319±1 | 0.88±<0.01 | – | – |
| Direct-mean | 102±<1 | 134±2 | 280±6 | 301±13 | 244±4 | 0.89±<0.01 | 305±5 | 0.88±<0.01 | – | – |
| X-MACE | 182 | 260 | 849 | 682 | 903 | 0.23 | 640 | 0.61 | – | – |
| OMNI-P2x arch. | 178 | 240 | 511 | 494 | 517 | 0.64 | 554 | 0.66 | – | – |
| MS-ANI | 168 | 234 | 529 | 560 | 508 | 0.61 | 569 | 0.63 | – | – |
| LEH + transition dipoles | 92±1 | 118±3 | 257±1 | 289±7 | 205±1 | 0.89±<0.01 | 273±3 | 0.89±<0.01 | 0.78±0.01 | 0.02±0.01 |
| LEH, random split | 106±1 | 112±<1 | 233±2 | – | 190±1 | 0.92±<0.01 | 243±7 | 0.92±<0.01 | – | – |
| Direct-softmin, random split | 112±1 | 117±2 | 248±7 | – | 206±2 | 0.91±<0.01 | 257±11 | 0.91±0.01 | – | – |
Table 8.
Active learning
Supplementary Table 9 summarises the excited-state sampling rounds (protocol in Methods). Round 1 is part of the benchmark dataset. In round 2, we labelled 18,411 geometries from model-driven dynamics of molecules with photoisomerisable double bonds, selected by double-bond twist and by committee disagreement. Before labelling, the benchmark model predicted their energy with an MAE of 0.53 eV and systematically too high. The error increased from 0.37 eV for twist angles below 30° to 1.11 eV beyond 75°, where TD-B97X-D places at 1–2 eV. These twisted geometries, which are rare in the benchmark training data, are the geometries along which photoisomerisation proceeds. In rounds 3 and 4, the error of the sampling models on the geometries they selected decreased to 0.42 eV and 0.40 eV (beyond 75°: 0.95 eV and 0.91 eV, main-text Figure 2d). The mean error along the seven validation trajectories of the models trained after each round decreased from 0.53 eV (benchmark) to 0.37 eV, 0.32 eV, and 0.27 eV. All sampling models were stable in their own dynamics (35 trajectories each, total-energy drift at most 7 meV, no bond changes). The benchmark models contain three of the seven validation molecules (styrene, benzaldehyde, and cinnamaldehyde) in their training set, and the sampling models of rounds 3 and 4, fine-tuned from the dye-augmented LEH model, additionally azobenzene (Methods), so the criterion values of these models are slightly optimistic. The production models contain none of the trajectory molecules. The sampling models did not reach the 0.2 eV threshold, and the improvement per round did not fall below 10%. The production models trained from scratch on all data after round 4 reached 0.17 eV, and the loop stopped. The two lineages of sampling models differed consistently across rounds (values in brackets in Supplementary Table 9), so the differences between single models along single trajectories are large compared with the changes from round to round.
| Round | Sampling models | Geometries | Molecules | MAE | signed | twist > 75° | MAE | -change MAE |
| 1 | preliminary LEH | 17,831 | 9,256 | – | – | – | 533 (533) | 196 (196) |
| 2 | benchmark LEH | 17,915 | 8,979 | 535 | +389 | 1112 | 372 (265, 478) | 141 (133, 150) |
| 3 | warm-started from dye-augmented LEH | 25,122 | 12,765 | 416 | +205 | 950 | 315 (219, 411) | 120 (119, 121) |
| 4 | warm-started from round 3 | 25,082 | 12,731 | 397 | +176 | 906 | 273 (185, 360) | 106 (107, 105) |
| production, uniform loss | trained from scratch | 167 | 78 |
Supplementary Table 9. Excited-state sampling rounds. Geometries and molecules: labelled geometries in the released dataset. Pre-label error: error of the sampling models on the geometries they selected, before labelling (signed: prediction minus TD-DFT). Criterion: mean absolute errors of the excitation energy and of the change of the energy along the seven validation trajectories for the models trained on all data up to and including the round (mean over the two models, single models in brackets). Last row: production models trained from scratch with the uniform loss (two LEH and two Direct-softmin models).
Production models
We trained the production models (LEH and Direct-softmin, four seeds each) from scratch on all data after active-learning round 1 with the weighted loss (Methods). Supplementary Table 11 lists their errors on every test set and along the trajectories. The weighted loss lowered the mean signed error of validation geometries with a TD-DFT below 2 eV from 0.39 eV to 0.25 eV (LEH) and from 0.43 eV to 0.22 eV (Direct-softmin). We used the uniform-loss models of Supplementary Table 9 for the stopping decision of the active-learning loop and for the decision on the loss.
Supplementary Table 11: Production models (weighted loss) on the validation split and all test sets and along the TD-B97X-D trajectories (frames before TD-DFT breaks down). Seeds: mean standard deviation over the four seeds. Comm.: committee (seed-averaged prediction). Trajectories: mean absolute error. Signed: mean error of the committee (prediction minus TD-DFT).
| Test set | LEH – MAE (meV) seeds | LEH MAE comm. | LEH comm. | Direct-softmin – MAE (meV) seeds | Direct-softmin MAE comm. | Direct-softmin comm. | |
| validation split | 5,395 | 174 | 169 | 162 | 157 | ||
| random, thermal | 9,182 | 155 | 140 | 147 | 134 | ||
| unseen scaffolds, thermal | 10,401 | 179 | 172 | 164 | 162 | ||
| random, -sampled | 1,449 | 223 | 201 | 200 | 184 | ||
| unseen scaffolds, -sampled | 1,802 | 235 | 225 | 210 | 200 | ||
| larger, thermal | 498 | 122 | 107 | 141 | 129 | ||
| larger, -sampled | 963 | 156 | 141 | 166 | 138 | ||
| random, active-learning frames | 2,232 | 255 | 270 | 225 | 246 | ||
| unseen scaffolds, active-learning frames | 2,282 | 285 | 318 | 244 | 277 | ||
| larger, active-learning frames | 947 | 192 | 211 | 187 | 187 | ||
| held-out dyes, eq. | 373 | 152 | 172 | 148 | 164 | ||
| held-out dyes, thermal | 373 | 161 | 185 | 150 | 169 | ||
| ladders, eq. | 1,226 | 266 | 434 | 307 | 321 | ||
| ladders, thermal | 1,228 | 252 | 384 | 285 | 303 | ||
| MAE along the TD-B97X-D trajectories (meV) | |||||||
| test: acrolein | 89 | signed +89 | 67 | signed +67 | |||
| test: uracil | 43 | signed +37 | 41 | signed −27 | |||
| test: trans-stilbene | 543 | signed +543 | 218 | signed +157 | |||
| test: HBDI | 433 | signed +433 | 223 | signed +220 | |||
| test: DMABN | 357 | signed +352 | 176 | signed +108 | |||
| mean of the 5 test trajectories | 293 | 145 | |||||
| mean of the 7 validation trajectories | 179 | 141 |
Supplementary Table 11. Production models (weighted loss) on the validation split and all test sets and along the TD-B97X-D trajectories (frames before TD-DFT breaks down). Seeds: mean standard deviation over the four seeds. Comm.: committee (seed-averaged prediction). Trajectories: mean absolute error. Signed: mean error of the committee (prediction minus TD-DFT).
Supplementary Table 10 compares both losses for each architecture. We applied the decision rule (Methods) to the mean over both architectures, and the weighted loss met it. For Direct-softmin, the weighted loss improved all three criteria and nearly all test sets, including the test trajectories. For LEH, it reduced the low-gap error and the errors on the active-learning test geometries by similar amounts, but the errors along the validation and test trajectories and for the ladder molecules were larger on average and varied more between seeds than with the uniform loss. These differences come mostly from single seeds. Seed 0 of the weighted LEH was more accurate than seed 0 of the uniform LEH along both sets of trajectories and for the ladder molecules, and seed 1 was less accurate. With two uniform-loss seeds, we cannot say whether the weighted loss changes the accuracy of LEH along trajectories or only its seed-to-seed variation.
| Errors (meV) | LEH | Direct-softmin | ||
| uniform | weighted | uniform | weighted | |
| decision criteria (validation data) | ||||
| validation, – MAE | ||||
| validation, , signed error | ||||
| validation trajectories, MAE | ||||
| test sets (– MAE unless stated otherwise) | ||||
| random, thermal | ||||
| unseen scaffolds, thermal | ||||
| random, -sampled | ||||
| larger, thermal | ||||
| larger, -sampled | ||||
| random, active-learning geometries | ||||
| unseen scaffolds, active-learning geometries | ||||
| larger, active-learning geometries | ||||
| held-out dyes, equilibrium | ||||
| ladder molecules, equilibrium | ||||
| test trajectories, MAE |
Supplementary Table 10. Uniform and weighted excitation-energy loss for the production models.
We kept the weighted loss for both architectures, as decided by the rule. For LEH, the uniform loss would have given lower errors for the ladder molecules and along the test trajectories, so the LEH results in the main text are not selected in favour of LEH.
Supplementary Table 10: Uniform and weighted excitation-energy loss for the production models. Mean absolute errors (meV), except for the mean signed error of the validation geometries with , as mean standard deviation over seeds (uniform loss: two seeds, weighted loss: four seeds). The first three rows are the criteria of the decision rule.
As for the benchmark models, Direct-softmin was more accurate for molecules similar to the training molecules (random test molecules: committee 0.147 eV versus 0.155 eV), and LEH was more accurate for larger molecules (0.122 eV versus 0.141 eV) and for – of ladder molecules (0.266 eV versus 0.307 eV). For of ladder molecules, Direct-softmin was more accurate (0.32 eV versus 0.43 eV). Both models overestimated more with increasing number of heavy atoms, LEH by more than Direct-softmin. The spread over seeds was small on the test sets of individual geometries but large on the ladder molecules and along the trajectories, where single seeds differed by several tenths of an eV on the same trajectory. The committee averages out this scatter but not the bias that all seeds share. Along the test trajectories of trans-stilbene, HBDI, and DMABN, the committee energy was on average 0.35–0.54 eV too high for LEH and 0.11–0.22 eV too high for Direct-softmin, and this error increased during the relaxation as the TD-DFT energy decreased (main-text Figure 6).
Supplementary Note References

Figure S5. Predicted versus TD-DFT excitation energies for all models (hexagonal bins, grey scale: number of geometries per bin). a, Larger test molecules at thermal geometries. b, Sulfur-free dyes at equilibrium geometries. Each panel gives the mean absolute error, , and the fraction of points outside the plotted range.

Figure 9. Errors of LEH at a fixed number of training geometries (10k and 20k) versus the number of geometries per molecule. Mean absolute error of the – excitation energies for random test molecules (a), unseen scaffolds (b), and larger test molecules (c), and of the forces for random test molecules (d), at thermal geometries. Markers: mean over two seeds. Bars: standard deviation.
References
- [1]Friederich, P.; Häse, F.; Proppe, J.; Aspuru-Guzik, A. Machine-learned potentials for next-generation matter simulations. Nature Materials 2021, 20, 750–761, DOI: 10.1038/s41563-020-0777-6.DOI
- [2]Westermayr, J.; Marquetand, P. Machine Learning for Electronically Excited States of Molecules. Chemical Reviews 2021, 121, 9873–9926, DOI: 10.1021/acs.chemrev.0c00749.DOI
- [3]Dral, P. O.; Barbatti, M. Molecular excited states through a machine learning lens. Nature Reviews Chemistry 2021, 5, 388–405, DOI: 10.1038/s41570-021-00278-1.DOI
- [4]Li, J.; Lopez, S. A. A Look Inside the Black Box of Machine Learning Photodynamics Simulations. Accounts of Chemical Research 2022, 55, 1972–1984, DOI: 10.1021/acs.accounts.2c00288.DOI
- [5]Hu, D.; Xie, Y.; Li, X.; Li, L.; Lan, Z. Inclusion of Machine Learning Kernel Ridge Regression Potential Energy Surfaces in On-the-Fly Nonadiabatic Molecular Dynamics Simulation. The Journal of Physical Chemistry Letters 2018, 9, 2725–2732, DOI: 10.1021/acs.jpclett.8b00684.DOI
- [6]Dral, P. O.; Barbatti, M.; Thiel, W. Nonadiabatic Excited-State Dynamics with Machine Learning. The Journal of Physical Chemistry Letters 2018, 9, 5660–5663, DOI: 10.1021/acs.jpclett.8b02469.DOI
- [7]Chen, W.-K.; Liu, X.-Y.; Fang, W.-H.; Dral, P. O.; Cui, G. Deep Learning for Nonadiabatic Excited-State Dynamics. The Journal of Physical Chemistry Letters 2018, 9, 6702–6708, DOI: 10.1021/acs.jpclett.8b03026.DOI
- [8]Westermayr, J.; Gastegger, M.; Menger, M. F. S. J.; Mai, S.; González, L.; Marquetand, P. Machine learning enables long time scale molecular photodynamics simulations. Chemical Science 2019, 10, 8100–8107, DOI: 10.1039/c9sc01742a.DOI
- [9]Li, J.; Reiser, P.; Boswell, B. R.; Eberhard, A.; Burns, N. Z.; Friederich, P.; Lopez, S. A. Automatic discovery of photoisomerization mechanisms with nanosecond machine learning photodynamics simulations. Chemical Science 2021, 12, 5302–5314, DOI: 10.1039/d0sc05610c.DOI
- [10]Li, J.; Stein, R.; Adrion, D. M.; Lopez, S. A. Machine-Learning Photodynamics Simulations Uncover the Role of Substituent Effects on the Photochemical Formation of Cunanes. Journal of the American Chemical Society 2021, 143, 20166–20175, DOI: 10.1021/jacs.1c07725.DOI
- [11]Axelrod, S.; Shakhnovich, E.; Gómez-Bombarelli, R. Excited state non-adiabatic dynamics of large photoswitchable molecules using a chemically transferable machine learning potential. Nature Communications 2022, 13, 3440, DOI: 10.1038/s41467-022-30999-w.DOI
- [12]Axelrod, S.; Shakhnovich, E.; Gómez-Bombarelli, R. Thermal Half-Lives of Azobenzene Derivatives: Virtual Screening Based on Intersystem Crossing Using a Machine Learning Potential. ACS Central Science 2023, 9, 166–176, DOI: 10.1021/acscentsci.2c00897.DOI
- [13]Kumar, P.; Zhou, C.; Friederich, P.; Escudero, D. In silico study of degradation in host molecular materials for OLEDs. Advanced Theory and Simulations 2026, 9, e70584, DOI: 10.1002/adts.70584.DOI
- [14]Westermayr, J.; Gastegger, M.; Marquetand, P. Combining SchNet and SHARC: The SchNarc Machine Learning Approach for Excited-State Dynamics. The Journal of Physical Chemistry Letters 2020, 11, 3828–3834, DOI: 10.1021/acs.jpclett.0c00527.DOI
- [15]Mausenberger, S.; Müller, C.; Tkat chenko, A.; Marquetand, P.; González, L.; Westermayr, J. SPaiNN: equivariant message passing for excited-state nonadiabatic molecular dynamics. Chemical Science 2024, 15, 15880–15890, DOI: 10.1039/d4sc04164j.DOI
- [16]Westermayr, J.; Marquetand, P. Deep learning for UV absorption spectra with SchNarc: First steps toward transferability in chemical compound space. The Journal of Chemical Physics 2020, 153, 154112, DOI: 10.1063/5.0021915.DOI
- [17]Martyka, M.; Zhang, L.; Ge, F.; Hou, Y.-F.; Jankowska, J.; Barbatti, M.; Dral, P. O. Charting electronic-state manifolds across molecules with multi-state learning and gap-driven dynamics via efficient and robust active learning. npj Computational Materials 2025, 11, 132, DOI: 10.1038/s41524-025-01636-z.DOI
- [18]Barrett, R.; Ortner, C.; Westermayr, J. Transferable Machine Learning Potential X-MACE for Excited States using Integrated DeepSeets. 2025; arXiv:2502.12870.arxiv.org/abs/2502.12870
- [19]Martyka, M.; Tong, X.-Y.; Jankowska, J.; Dral, P. O. OMNI-P2x universal neural network potential for excited-state simulations. Nature Communications 2026, 17, 4949, DOI: 10.1038/s41467-026-71380-5.DOI
- [20]Malosso, C.; How, W. B.; Díaz Mirón, G.; Hassanali, A.; Ceriotti, M. Transferable machine learning of excited-state dynamics with extremal pooling. 2026; arXiv:2606.16859.arxiv.org/abs/2606.16859
- [21]Westermayr, J.; Maurer, R. J. Physically inspired deep learning of molecular excitations and photoemission spectra. Chemical Science 2021, 12, 10755–10764, DOI: 10.1039/d1sc01542g.DOI
- [22]Cignoni, E.; Suman, D.; Nigam, J.; Cupellini, L.; Mennucci, B.; Ceriotti, M. Electronic Excited States from Physically Constrained Machine Learning. ACS Cent. Sci. 2024, 10, 637–648, DOI: 10.1021/acscentsci.3c01480.DOI
- [23]Ren, F.; Chen, X.; Liu, F. Size-Transferable Prediction of Excited State Properties for Molecular Assemblies with a Machine Learning Exciton Model. The Journal of Physical Chemistry Letters 2025, 16, 2541–2552, DOI: 10.1021/acs.jpclett.4c03548.DOI
- [24]Juergens, D.; Stöhr, M.; Hillers-Bendtsen, A. E.; Fajen, O. J.; Martínez, T. J. Latent unified smooth Hamiltonians for excited state chemistry. 2026; arXiv:2609.01871.arxiv.org/abs/2609.01871
- [25]Ramakrishnan, R.; Hartmann, M.; Tapavicza, E.; von Lilienfeld, O. A. Electronic spectra from TDDFT and machine learning in chemical space. The Journal of Chemical Physics 2015, 143, 084111, DOI: 10.1063/1.4928757.DOI
- [26]Zhu, Y.; Li, M.; Xu, C.; Lan, Z. Quantum chemistry dataset with ground- and excited-state properties of 450 kilo molecules. Scientific Data 2024, 11, 948, DOI: 10.1038/s41597-024-03788-x.DOI
- [27]Curth, R.; Röhrkasten, T. E.; Müller, C.; Westermayr, J. Surface Hopping Nested Instances Training Set for Excited-state Learning. Scientific Data 2025, 12, 1300, DOI: 10.1038/s41597-025-05443-5.DOI
- [28]Pinheiro Jr, M.; Zhang, S.; Dral, P. O.; Barbatti, M. WS22 database, Wigner sampling and geometry interpolation for configurationally diverse molecular datasets. Scientific Data 2023, 10, 95, DOI: 10.1038/s41597-023-01998-3.DOI
- [29]Batatia, I.; Kovács, D. P.; Simm, G. N. C.; Ortner, C.; Csányi, G. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. Advances in Neural Information Processing Systems. 2022; pp 11423–11436, DOI: 10.48550/arXiv.2206.07697, arXiv:2206.07697.DOI
- [30]Zhou, C.; Neubert, M.; Koide, Y.; Zhang, Y.; Vuong, V.-Q.; Schlöder, T.; Dehnen, S.; Friederich, P. PAL – parallel active learning for machine-learned potentials. Digital Discovery 2025, 4, 1901–1911, DOI: 10.1039/d5dd00073d.DOI
- [31]Jacquemin, D.; Wathelet, V.; Perpète, E. A.; Adamo, C. Extensive TD-DFT benchmark: singlet-excited states of organic molecules. Journal of Chemical Theory and Computation 2009, 5, 2420–2435, DOI: 10.1021/ct900298e.DOI
- [32]Liang, J.; Feng, X.; Hait, D.; Head-Gordon, M. Revisiting the performance of time-dependent density functional theory for electronic excitations: assessment of 43 popular and recently developed functionals from rungs one to four. Journal of Chemical Theory and Computation 2022, 18, 3460–3473, DOI: 10.1021/acs.jctc.2c00160.DOI
- [33]Loos, P.-F.; Scemama, A.; Jacquemin, D. The quest for highly accurate excitation energies: a computational perspective. The Journal of Physical Chemistry Letters 2020, 11, 2374–2383, DOI: 10.1021/acs.jpclett.0c00014.DOI
- [34]He, Y.; Escudero, D. Beyond the three-state picture: when higher-lying excited states become quantitatively indispensable. Chemical Science 2026, 17, 10967–10981, DOI: 10.1039/d6sc01726f.DOI
- [35]Levine, B. G.; Ko, C.; Quenneville, J.; Martínez, T. J. Conical intersections and double excitations in time-dependent density functional theory. Molecular Physics 2006, 104, 1039–1051, DOI: 10.1080/00268970500417762.DOI
- [36]Batatia, I. et al. A foundation model for atomistic materials chemistry. The Journal of Chemical Physics 2025, 163, 184110, DOI: 10.1063/5.0297006.DOI
- [37]Levine, D. S. et al. The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models. 2025; arXiv:2505.08762.arxiv.org/abs/2505.08762
- [38]Smith, J. S.; Nebgen, B. T.; Zubatyuk, R.; Lubbers, N.; Devereux, C.; Barros, K.; Tretiak, S.; Isayev, O.; Roitberg, A. E. Approaching coupled cluster accuracy with a general-purpose neural network potential through transfer learning. Nature Communications 2019, 10, 2903, DOI: 10.1038/s41467-019-10827-4.DOI
- [39]Balasubramani, S. G. et al. TURBOMOLE: Modular program suite for ab initio quantum-chemical and condensed-matter simulations. The Journal of Chemical Physics 2020, 152, 184107, DOI: 10.1063/5.0004635.DOI
- [40]Chai, J.-D.; Head-Gordon, M. Long-range corrected hybrid density functionals with damped atom–atom dispersion corrections. Phys. Chem. Chem. Phys. 2008, 10, 6615–6620, DOI: 10.1039/B801898B.DOI
- [41]Hariharan, P. C.; Pople, J. A. The influence of polarization functions on molecular orbital hydrogenation energies. Theoretica Chimica Acta 1973, 28, 213–222, DOI: 10.1007/BF00533485.DOI
- [42]Furche, F.; Ahlrichs, R. Adiabatic time-dependent density functional methods for excited state properties. The Journal of Chemical Physics 2002, 117, 7433–7447, DOI: 10.1063/1.1508368.DOI
- [43]Bannwarth, C.; Ehlert, S.; Grimme, S. GFN2-xTB—An Accurate and Broadly Parametrized Self-Consistent Tight-Binding Quantum Chemical Method with Multipole Electrostatics and Density-Dependent Dispersion Contributions. Journal of Chemical Theory and Computation 2019, 15, 1652–1671, DOI: 10.1021/acs.jctc.8b01176.DOI
- [44]Bemis, G. W.; Murcko, M. A. The properties of known drugs. 1. Molecular frameworks. Journal of Medicinal Chemistry 1996, 39, 2887–2893, DOI: 10.1021/jm9602928.DOI
- [45]Gómez-Bombarelli, R.; Wei, J. N.; Duvenaud, D.; Hernández-Lobato, J. M.; Sánchez-Lengeling, B.; Sheberla, D.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; Aspuru-Guzik, A. Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules. ACS Cent. Sci. 2018, 4, 268–276, DOI: 10.1021/acscentsci.7b00572.
- [46]Irwin, J. J.; Shoichet, B. K. ZINC – A Free Database of Commercially Available Compounds for Virtual Screening. J. Chem. Inf. Model. 2005, 45, 177–182, DOI: 10.1021/ci049714+.DOI
- [47]RDKit contributors RDKit: Open-source cheminformatics. https://www.rdkit.org, 2026.
- [48]Beard, E. J.; Sivaraman, G.; Vázquez-Mayagoitia, A.; Vishwanath, V.; Cole, J. M. Comparative dataset of experimental and computational attributes of UV/vis absorption spectra. Scientific Data 2019, 6, 307, DOI: 10.1038/s41597-019-0306-0.DOI
- [49]Young, J. F.; Han, M.; Jeong, M.; Park, S. Experimental database of optical properties of organic compounds. Scientific Data 2020, 7, 295, DOI: 10.1038/s41597-020-00634-8.
- [50]Dral, P. O. et al. MLatom 3: A Platform for Machine Learning-Enhanced Computational Chemistry Simulations and Workflows. Journal of Chemical Theory and Computation 2024, 20, 1193–1213, DOI: 10.1021/acs.jctc.3c01203.DOI





