WorksPhysics and Astronomy
Orbital-weighted learning for transition metal complexes
Abstract
Frontier orbital energies are key to the description of molecular light-matter interaction and pose a challenge both for electronic structure calculations and for machine learning approaches, especially for transition metal complexes (TMC). These are highly relevant since the gap between the highest occupied molecular orbital (HOMO) and the lowest unoccupied molecular orbital (LUMO) is in the range of visible light. For TMCs, gap position and magnitude are determined by the metal center, its coordination and ligand identity. Since the systems exhibit a wide range of charge and spin states, frontier orbital energies can only be predicted with machine learning models that explicitly incorporate spin and charge. In this work, we show that a semi-empirical orbital population as a weight rather than a feature can increase efficiency for both physics-based and neural-network-based atomic representations. Because orbital populations are highly sensitive to both charge and spin, they offer a general pathway to render existing geometry-based machine learning representations useful for orbital energies. On TMC datasets, our approach reduces the required training data by up to 95 %, and improves physics-based representations to almost match the accuracy of pre-trained neural network representations at 12% of the inference compute cost. This greatly reduces the data and compute requirements to accurately model frontier orbitals in TMCs
Introduction
The highest occupied molecular orbital (HOMO) and the lowest unoccupied molecular orbital (LUMO) together are the frontier orbitals. Both their absolute energy value as well as the difference, the so-called gap, are of fundamental interest to interactions between molecules and light, even as approximation[1, 2]. The HOMO energy is approximately the negative of the ionization potential, and as such describes approximately the amount of energy required to remove one electron from a molecule. The negative of the LUMO energy, again, is a proxy for the electron affinity, a measure of how easily a molecule would accept an electron. The orbital energies and the gap are relevant for applications such as OLEDs[3] or organic solar cells[4], but in theoretical modeling also finds use in reactivity analysis[5, 6].
Even though many quantum chemistry methods yield values for frontier orbital energies, the accuracy thereof varies widely, for example for transition metal complexes[7, 8]. With higher levels of theory being able to recover accurate values at higher cost, many researchers have investigated machine learning routes to enable accurate predictions at affordable cost, building upon density functional theory (DFT) data at GGA level in most cases[9], as only few other data points exist[7, 10]. Unlike energies which can be approximated well as sum of local energy contributions, frontier orbitals are sensitive to the overall molecular structure, so they pose a harder problem even though their effective dimensionality is lower than the one of the energy[11]. Moreover, applications regarding frontier orbitals often include charged molecules or a specific spin state for which many models have no native support[12].
One successful approach for dealing with highly diverse and clustered data sets is splitting data set into groups such as charge and spin[13, 14, 15, 16], which is independent of the particular learning approach. Similarly, learning the difference between a computationally inexpensive and a more accurate higher level result, Δ-ML[17], and its generalization multi-fidelity learning has been applied to frontier orbitals and related properties such as excitation energies[18].
Kernel or linear models require a representation or a descriptor of the system. Over the years, many such representations have been developed[19, 20]. Frontier orbitals or gaps have been learned using a wide range of geometry-based representations such as BAML[21], MTM[22], many-body expansions[23], MBTR[24, 25], cMBDF[26], SPAᴴM and its variants[27, 28], SOAP[29]. For some application classes such as transition metal complexes (TMC), using the atomic representations of the metal center only[12] proved to be an effective simplification. Some models extend existing representations such as SLATM with DFTB-derived properties[30] or built upon geometry free features from element embeddings[31], autocorrelation functions on the molecular graph[32, 33] or general graph features[34, 35, 36, 37, 38, 39, 40, 41, 42] including fragmentation approaches[43]. Other approaches in similar spirit combine cheminformatics descriptors with electronic structure features through stacking[44], attention[45], or concatenation[24]. Fundamentally different is the idea to build representations from molecule-specific electronic structure calculations, as done in MOB-ML using Fock, Coulomb and Exchange matrix from a Hartree-Fock calculation which proved particularly helpful for post-SCF correlation energies[46, 47, 48] which has been similarly developed for general learning[49]. This includes DFT-derived descriptors of metal centers[50] or the full system[51], and learning a Löwdin-symmetrized effective Hamiltonian to reproduces MO energies[52] and low-fidelity pre-training[53].
For neural networks, several approaches have been considered: using semi-empirical Fock, density, Hamiltonian and overlap matrices as feature in OrbNet-Equi[54], environments of top- atoms for multiple frontier orbitals in one representation (ESIN)[55], using a physics-inspired representation as in put to a neural network[56], using a ZINDO-derived HOMO-weighted radial distribution function (MO-RDF) for TD-DFT in AttentiveFP[57], learning the a weighted average pooling function for a MPNN with a second jointly-trained neural network[58], or transfer learning through augmentation of a feature with a low-fidelity label[59]. In a more general approach, learning the full wave function (e.g. in SchNOrb[60]) or orbitals (e.g. PaiNN[29]) gives direct access to the frontier orbital energies and allows further analysis of the effective many-body-character[61].
Other approaches not including electronic structure-based features are described in comprehensive reviews[62, 63].
Methods
Many molecular representations exist in two variants: global and local. In the global version, the information of a molecule is collected in a single feature vector, while the local version contains a feature vector for every atom. The atomic feature vectors then form a matrix with the shape for every molecule. While a local representation usually yields better performance for Kernel-Ridge-Regression (KRR) models for geometry-based properties, it comes with higher computational cost and hardware requirements, such as increased memory usage which may require batching[64].
Since frontier orbitals in TMCs are typically localized around a metal center, their character depends on the surrounding ligands, making them quite challenging to handle. As most of the physical representations scale with the number of unique elements in a dataset[20, 65] and the average number of atoms in TMCs is high (see Table I), the usage of local representations becomes impractical and for some long representations infeasible: some representations such as SLATM reach over a million features on a typical TMC dataset[12]. To build a model that can predict the energy of frontier orbitals in large, chemically diverse systems, it is therefore advantageous to use global representations, which can be seen as a pooling of local representations[58, 66]. For frontier orbitals, we propose a weighted sum of local representations where the weights are orbital coefficients from a semiempirical calculation. After the weighted sum, we employ a standard KRR model across different datasets and compare them with other state-of-the-art models such as pre-trained MACE variants, see the schematic in Figure 1.
Figure 1. Illustration of the proposed method (orbital-weighted learning, OWL) for a transition metal complex. Traditional approaches either use all atomic environments (local) or the atomic environment of the metal center only (red circle). OWL uses weighted sums of the local environments of atoms with orbital population (green area)
| Dataset | tmPHOTO | TM-GSspin | QM9 |
| Training points | 4,268 | 2,260 | 133,885 |
| Unique elements | 25 | 18 | 5 |
| Charges | -1 to 1 | -5 to 4 | none |
| Spin multiplicity | 1 to 6 | singlet | singlet |
| Average atom count | 73 | 54 | 18 |
Table I. Descriptive statistics of the datasets used in this work.
Unlike augmenting existing geometric representations with orbital features, the weighting offers a more direct way of grouping training points in feature space for this application, since small variations in geometry can yield a large discontinuous response in orbital shape, e.g. through crossing of orbital energies. Using geometry information alone would induce an erroneous similarity.
Orbital Weighting
When sum-pooling a local representation into a global version, the features of all atoms would contribute equally to the global vector. Since the frontier orbital is localized for most TMCs, the naïve sum will dilute the signal of the relevant atomic environments, leading to an effect similar to noise in the representation. To identify the atoms that carry a higher contribution to the formation of the frontier orbitals, we perform a Mulliken population analysis of one specific orbital only. By selecting a specific molecular orbital (MO) , we can obtain an expression for the weight of atom A to this MO as
with standard coefficient matrix C and overlap matrix S. The sum over runs over all atomic orbital (AO) functions centered on atom A, while runs over all AO in the molecule. With those weights, we obtain a global representation for each MO from all generic atomic representations :
This can be seen as a general expression to obtain a global representation from a local version. For sum-pooling, all weights are one, with every atom weighted uniformly. Using the metal center only would mean to have all weights zeros, except for one entry corresponding to the metal which would be of unit weight. To obtain the coefficients and the overlap matrix in (1), we calculate the orbital energies with the semi-empirical method GFN2-xTB and, in the case of open-shell systems, the spin-polarized version thereof [8]. While higher levels of theory would yield a different weight vector that might be more accurate, we chose xTB to balance computational cost with physically reasonable results. Nothing in our approach is restricted to this particular level of theory.
Charge and spin detrending
Detrending and centering data is a common practice in machine learning to remove a systematic offset. When a dataset contains molecules with different charges, the frontier-orbital distribution shifts and becomes multimodal, since molecular orbitals are sensitive to charge. The multimodal distribution tends to have higher variance, and representations without a charge and spin embedding will not capture this offset effectively. The effect of the spin might not be as strong as the charges, but for TMCs both properties are often coupled, and both will have an effect on the frontier orbitals. To detrend these properties, we collect the charges and spin states for every molecule in the dataset via one-hot encoding in a design matrix . With this matrix, we then perform a least-squares fit on the training data
This fitted contribution is then subtracted to obtain the detrended residuals
We then center the residuals around zero by subtracting the overall mean. The residuals then show almost no charge- and spin-dependent effects and reduced variance.
Kernel ridge regression
To test our method, we employed Kernel ridge regression (KRR), where a kernel function measures the similarity of data points. We use the Gaussian kernel. In this work, this refers to the similarity between representations . To predict the label of a query point we use the form
where the model coefficients are obtained via direct inversion
with kernel matrix element and regularization strength . Hyperparameter optimization and cross-validation follow previous work[64].
Datasets and Representations
To test our modifications on representations, we evaluate them on FCHL19[67], cMBDF[26], and MACE[68, 69]. MACE is built as a neural network potential, but a pre-trained version of the model can also be used as a (local) representation[69]. In this work, we use pre-trained versions of MACE: MACE-OFF, trained on small organic molecules[70]; MACE-MP0, trained on the MPtrj dataset[71]; and MACE-OMol, trained on the Omol25 dataset[72]. For MACE-OFF and MACE-MP0, we use the medium subvariant. The MACE-OMol version also features charge and spin embedding.
We tested on the tmPHOTO, TM-GSspin[12], and QM9 datasets. The tmPHOTO dataset is a subset of the tmQM[73] database and consists of photoactivity-associated complexes with a charge from to in a singlet spin state. The TM-GSspin dataset consists of TMCs with charges from to and spin multiplicity up to 6. The QM9 dataset is a well-established benchmark set for ML models and consists of organic molecules with up to nine heavy atoms. Further details of the dataset are shown in Table I.
Results
Photoactive transition metal complexes
Frontier orbital energies are important for the photophysical properties of complexes, such as absorption and emission. A suitable dataset is the tmPHOTO dataset, a subset of tmQM. This subset was derived from the original dataset by filtering for complexes that appeared in manuscripts related to photophysical activity[73]. When testing our method on the tmPHOTO dataset, we applied the same element filtering as in literature[12]. The remaining dataset contains 4,268 complexes with 25 unique elements.
With our orbital-weighted learning (OWL), every property (HOMO, LUMO and gap) is learned more efficiently, see Figure 2. For HOMO, using orbital-weighting makes both physics-based representations FCHL19 and cMBDF more accurate with 95 % fewer data (128 data points instead of 3,414, i.e. 80 % of the dataset). Large pre-trained models are only slightly better than physics-based representations for HOMO energies in both the direct and learning case. While this indicates that the pre-training on extensive datasets does not yield a material improvement, using KRR as additional head/readout function during MACE training may further improve the models. The improvement for frontier orbitals is a consequence of OWL being highly sensitive to the molecular charges. Since the dataset contains three different molecular charges, the raw frontier-orbital distribution is multi-modal, which constitutes a harder learning target. This applies particularly to kernel ridge regression which reverts to predicting the mean label for query points far away from any training data. This is in line with earlier observations[12] that using MACE-QS (a MACE variant which includes charge and spin embeddings) performs significantly better than MACE alone.
Figure 2. Learning curves for the tmPHOTO dataset[73] after inclusion of the proposed orbital weightings, both for direct (filled markers) and -learning (empty markers) with an xTB baseline. Data shown for HOMO (A), LUMO (B) and gap (C) up to 80% of the dataset (3,414 training points). The target accuracy of 0.1 eV is shaded in grey. The reference values (grey crosses) show models[12] without the proposed method.
For the LUMO in Figure 2B, we observe the same pronounced improvement in data efficiency and OWL-FCHL19 and OWL-MACE-MP are close in performance. Only OWL-MACE-OMol shows higher data efficiency requiring about 25 % of the training data that other representations need. All representations for all properties benefit from our orbital weighting. As one would expect from learning the difference of two separate functions, the learning curves for the gap in Figure 2C follow the behavior of the quantity that is harder to learn in the first place, the LUMO energy. Since we have to choose either the HOMO or LUMO weights when learning the gap, neither of these options on its own is solely representative of the gap character. We find that for direct learning LUMO yields better results and for learning, the HOMO weights perform better. Remarkably, the prediction accuracy improves by using the weights even for predicting the gap.
Since our method requires an xTB single-point calculation, a natural extension to direct learning is a -learning approach between the target’s higher level of theory and the xTB orbital energies. For the HOMO energy, we obtained an MAE similar to MACE-QS, the best-performing model on this dataset, when applying our method to the physics-inspired representation FCHL19. We achieved the overall lowest MAE with the data-driven representation MACE-OMol in our KRR setting. For all data-driven pre-trained representations, the effective training data available to the pre-trained models is much larger: the OMol dataset contains 140 million data points[72]. Consequently, their use constitutes transfer learning. For physics-inspired representation FCHL19, all improvement is achieved without any pre-training.
In the -learning approach, three of the four representations were able to achieve an MAE below 0.1 eV between 2,048 and 3,414 datapoints. Only cMBDF did gain little across all tests, since it is built to exhibit a constant length, independent of the number of unique elements in the dataset and the element identity only enters through the elemental kernel function. Our summation therefore suffers from information loss that impairs its learning behavior. This is evident in the flattening of the learning curves for larger training set sizes. The learning curves for the LUMO energy in direct learning are comparable to the curves for the HOMO. With the approach however, it shows a significantly lower MAE, with all the representations reaching the 0.1 eV threshold with less than 1,024 training points. To investigate the different behavior of the learning curves, we examined the correlation between the xTB energies and the target values at the TPSSh-D3(BJ)/def2-SVP level and observed a higher correlation for the LUMO than for the HOMO energy. Due to this difference in correlation, the LUMO- is the easier-to-learn property compared to the HOMO-.
The multi-modal character of the frontier orbital energy as consequence of the charge state is greatly reduced for the HOMO-LUMO gap, leading to better relative performance of the representations without charge embedding in Figure 2C. While cMBDF yields a higher MAE with our method due to information loss in the summation, OWL improves the performance of FCHL19 by reducing the data requirements by 75 %. When using the weights corresponding to the LUMO, FCHL19 shows a 15 % lower MAE, while the usage of the weights corresponding to the HOMO yields a 5 % lower MAE. For every tested representation, using the LUMO weights in direct learning yields the lower MAE. Interestingly, for the -learning approach, this pattern reverses; here, every representation benefits from using the HOMO weights. While a strong offset effect from using learning with only flat learning afterwards is documented for the strict superset tmQM[74], we observe continuous learning for tmPHОTO even after reducing the offset. When both weights are applied subsequently to the representation, the models perform worse than with a single weight representative of either HOMO or LUMO, so we always choose either MO. In Figure 2C, we show the best case for each curve. For MACE-MP0, we measured an improvement similar to FCHL19.
TMCs with higher spin multiplicity
Another important feature of TMCs is that they occur in various spin multiplicities. Many established representations have been developed for and benchmarked on small organic molecules and, hence, do not embed spin information. In OWL, the weighting naturally depends on the electronic state including spin and charge.
To test our method across different spin multiplicities, we chose the TM-GSspin dataset, an extension of the TM-GSspin dataset designed to build models related to spin-related properties[12, 75]. TM-GSspin yields results similar to those of tmPHОTO. While the MAE is higher for TM-GSspin, the improvement due to OWL is comparable. For both TMC datasets, the MAE of the frontier orbital prediction could be reduced significantly, except for MACE-OMol, where the performance improvement is only 15 % for the HOMO and 21 % for the LUMO energy. The TM-GSspin dataset shows greater diversity in spin multiplicity than the tmPHОTO dataset, but by using spin-polarized xTB[8], our method shows a consistent improvement with every representation we tested. The relative improvement on the spin dataset is slightly higher than on tmPHОTO, with MAE reductions up to 70 % for FCHL, 74 % for MACE-MP, and 59 % for cMBDF. MACE-OMol remained rather constant. The higher MAE shown in Figure 3 for TM-GS compared to the other datasets is a combination of the higher complexity due to spin multiplicities and the smaller training set size: our largest training set size of 80 % of the dataset corresponds to 3,414 training points for tmPHOTO and 1,808 training points for GS.

Figure 3. MAE with and without orbital-weighted learning (OWL) for all direct learning cases of the representations (colors), datasets (filling) and properties (shape). The data is shown for 80% of the dataset for tmPHOTO and TM-GSSpin⁺ and 4,096 training points for QM9. Points below the diagonal show an improvement, while points above show a decline in performance.
molecules with an average size of 73 atoms per molecule for tmPHOTO and 54 atoms per molecule for TM-GS, and a single metal center. Therefore, the frontier orbitals are relatively localized in TMCs. Our method exploits this localization; QM9, by contrast, contains smaller molecules with 18 atoms per molecule. Because of the size and structural differences between the organic molecules and the TMCs, frontier-orbital localization is less pronounced, which decreases the effectiveness of our method on datasets such as QM9.
Small organic molecules
Most physics-based representations were developed on datasets consisting of relatively small molecules with only a handful of unique elements. These representations perform well e.g. on QM9 and similar datasets. We do not expect OWL to be particularly effective for small systems, since the frontier orbitals are rarely localized only on small parts of these systems. To explore the limits of OWL, we tested the prediction accuracy on QM9. In this case, we replaced MACE-MP0 with MACE-OFF, as this variant was pre-trained on organic molecules similar to QM9 and performs better than the MP0 variant on such datasets. In our tests, most representations performed worse on this dataset when OWL is included. MACE-OMol again had the lowest MAE of all representations, and the effect of our method was minor but still slightly negative, with a 1 % improvement in the HOMO, a 7 % decrease for the LUMO, and a 16 % decrease for the gap. For cMBDF, the MAE for the HOMO increased by 23 % with our method; the gap and LUMO predictions showed a slight decrease in MAE of 3 % and 8 %, respectively. For FCHL19, we saw a noticeable improvement in the LUMO (33 %), a moderate improvement in the gap (19 %), and a minor improvement in the HOMO (8 %). The biggest outlier is MACE-OFF; here we observed a significant increase in the MAE for every property when applying our method. The MAE for the HOMO and gap predictions doubled, while the MAE for the LUMO increased by 47 %. This aligns with earlier work showing that scaling and compressing features for MACE as a representation is ineffective[76] on QM9. However, the decrease in efficiency for our method stems from the dataset structure itself. The TMC datasets contain large
Overall effects
For a systematic comparison, we collect models for all representations on the same data with and without using our method OWL, see Figure 3. In general, we observe OWL to improve data efficiency for large molecules and orbital energies. Interestingly, MACE-MP benefits substantially from OWL, more than halving the error. Since that model is pre-trained on materials and not molecules, OWL supports its transfer onto molecular systems without additional training, thus increasing its applicability domain. This is likely due to the strong influence of the metal-center on the orbital character, where the metal center is closer to the training data of MACE-MP.
For physics-inspired representations, OWL typically halves the error, in some cases reduces it to a quarter as shown in Figure 3. This can be understood since the many-body interactions for large systems involve atoms which do not influence the orbitals strongly. This behaves like an invariance: along changes of molecular structure or atomic environment far away from the orbital in question, neither orbital shape nor energy are subject to substantive change. This invariance then needs to be discovered by the models from the data alone, which makes those models less data efficient. If those of the many-body contributions which are less relevant are suppressed by the weighting, this inefficiency is removed.
As shown in Figure 2, the learning curves for HOMO, LUMO and gap are all very similar, although shifted. That is the because the label variance after categorical detrending by charge state in tmPHOTO is ordered with HOMOLUMOgap.
The data efficiency (i.e. the learning curve slope) is comparable for all models, so only the offset originating from the label variance is different. This trend is reproduced for OWL in Figure 3, where HOMO and LUMO benefit either both or none from the orbital weighting, indicating that the relevance selection induced by this method yields consistent results for relevant orbitals.
The weights in OWL are approximate, since they are obtained at low computational cost at semi-empirical level only. This assumes a strong similarity of the orbital characters in both methods which is validated by the success of the method only. In Figure 3, this seems to be the case even for transition metal complexes where semi-empirical methods cannot offer a reliable quantitative assessment of different spin states. It is to be expected that the support from the semi-empirical weights then helps inaccurate models to improve but offers little improvement in the limit of already accurate models. This is observed in Figure 3, where OWL seems to improve the models to about eV. In the limit of infinite training data, the benefit from OWL vanishes the same way as the one of -learning does.
Timings
The time required to use our method is the sum of the time required to first calculate the geometry-based representation and then the time to perform a semi-empirical calculation. The time to apply the weight vector to the representation is negligible. Figure 4 shows the timings for a random subset of 500 molecules each from the tmPHOTO and QM9 datasets. The xTB calculations were performed with the tblite Python package (version 0.6.0) using spin-polarized GFN-2 xTB. We performed the time measurements on a single CPU core to enable a fair comparison of the total compute effort. The neural networks would benefit from GPU acceleration while FCHL and cMBDF would benefit from multicore CPU parallelization.
Figure 4. Time required to calculate the different representations for 500 molecules on tmPHOTO (A) and QM9 (B) on a single core.
The data-driven representations require significantly more time to calculate, with MACE-OMol taking the most time due to its parameter count. While this representation on average performed better than the geometry-based representations for direct learning tasks, its performance for -learning is comparable to other representations, as shown in Figure 2. On the tmPHOTO dataset, 500 molecules required 2,163 seconds for MACE-OMol. FCHL19 takes 14 seconds instead, and the xTB single-point calculations took 406 additional seconds. The compute cost does not increase for multiple frontier orbitals, since all weights for all MOs are obtained from the same calculation. In our KRR workflow, inference time is dominated by calculating the representations rather than by model inference[64]. Using orbital weighting benefits from the fast calculation time of geometry-based representations while keeping the extra computational time to a minimum by using xTB for the orbital weights.
Conclusions
In this work, we demonstrated that the applicability of geometry-based representations can be extended to large and diverse transition metal complexes (TMCs) by including orbital information. With only geometry information, spin and charge of the TMCs cannot be resolved by the representation. Conceptually, ligand parts which do not contribute to the orbital in question form an invariance in feature space, since many modifications of that part of the molecule would not change the orbital shape or energy significantly. With our orbital weighting, we scope the feature space to only the relevant parts of the molecule. Interestingly, both physics-based and neural-network representations benefit.
With orbital weighting learning (OWL), TMCs of different spin and charge states can be trained jointly even if the original geometry-based representation is not adapted for charge and spin. Orbital weighting performs best if large molecules with partially localized frontier orbitals are considered, such as TMCs. The improvement over existing approaches is most prominent for cases with large element diversity and applications with many charge and spin combinations. For small molecules where frontier orbitals are distributed over essentially the full molecule, OWL offers no advantage.
OWL inference requires an approximate molecular orbital expressed in per-atom coefficients. Although we use semi-empirical methods in this work, this is not a requirement, as orbital populations are commonly available. Geometry-based representations with OWL are much faster than neural-network models but regularly achieve comparable performance. This reduces inference time, but leaves the semi-empirical method as the current bottleneck.1
References
References
- [1]P. Kimber and F. Plasser, Energy Component Analysis for Electronically Excited States of Molecules: Why the Lowest Excited State Is Not Always the HOMO/LUMO Transition, Journal of Chemical Theory and Computation 19, 2340 (2023).DOI
- [2]L. Kronik, T. Stein, S. Refaely-Abramson, and R. Baer, Excitation Gaps of Finite-SizedSystems from OptimallyTuned Range-Separated Hybrid Functionals, Journal of Chemical Theory and Computation 8, 1515 (2012).DOI
- [3]A. Aarti, B. P. Veettil, A. Rodger, and K. Venkatesan, Recent Progress and Challenges in Molecular Design for Hyperfluorescent Based Organic Light Emitting Diodes (OLEDs), Advanced Optical Materials 13, e01648 (2025).DOI
- [4]J. Yi, G. Zhang, H. Yu, and H. Yan, Advantages, challenges and molecular design of different material types used in organic solar cells, Nature Reviews Materials 9, 46 (2023).
- [5]R. A. Miranda-Quintana, F. Heidar-Zadeh, S. Fias, A. E. A. Chapman, S. Liu, C. Morell, T. Gómez, C. Cárdenas, and P. W. Ayers, Molecular Interactions From the Density Functional Theory for Chemical Reactivity: The Interaction Energy Between Two-Reagents, Frontiers in Chemistry 10, 906674 (2022).DOI
- [6]R. Pucci and G. G. N. Angilella, Density functional theory, chemical reactivity, and the Fukui functions, Foundations of Chemistry 24, 59 (2022).
- [7]S. Dohm, A. Hansen, M. Steinmetz, S. Grimme, and M. P. Checinski, Comprehensive Thermochemical Benchmark Set of Realistic Closed-Shell Metal Organic Reactions, Journal of Chemical Theory and Computation 14, 2596 (2018).DOI
- [8]H. Neugebauer, B. Bädorf, S. Ehlert, A. Hansen, and S. Grimme, High-throughput screening of spin states for transition metal complexes with spin-polarized extended tight-binding methods, Journal of Computational Chemistry 44, 2120 (2023).DOI
- [9]D. Balcells and B. B. Skjelstad, tmQM Dataset—Quantum Geometries and Properties of 86k Transition Metal Complexes, Journal of Chemical Information and Modeling 60, 6135 (2020).DOI
- [10]A. G. Garrison, J. W. Toney, T. Nikolaeva, R. G. S. Michel, C. J. Stein, and H. J. Kulik, The BOS-TMC Dataset: DFT Properties of 159k Experimentally Characterized Transition Metal Complexes Spanning Multiple Charge and Spin States (2022).
- [11]A. Banjafar and G. F. Von Rudorff, Intrinsic dimensionality of molecular properties, The Journal of Chemical Physics 163, 174301 (2025).
- [12]Y. Cho, K. R. Briling, Y. Calvino Alonso, R. Laplaza, and C. Corminboeuf, Benchmarking physics-inspired machine learning models for transition metal complexes with diverse charge and spin states, Digital Discovery 5, 2103 (2026).DOI
- [13]B. Mazouin, A. A. Schöpfer, and O. A. Von Lilienfeld, Selected machine learning of HOMO–LUMO gaps with improved data-efficiency, Materials Advances 3, 8306 (2022).
- [14]C. Zeni, A. Anelli, A. Gliemo, S. De Gironcoli, and K. Rossi, Divide-and-conquer potentials enable scalable and accurate predictions of forces and energies in atomistic systems, Digital Discovery 3, 113 (2024).
- [15]D. Lemm, G. Falk von Rudorff, and O. Anatole von Lilienfeld, Improved decision making with similarity based machine learning: applications in chemistry, Machine Learning: Science and Technology 4, 045043 (2023).DOI
- [16]Š. Sršen, O. A. Von Lilienfeld, and P. Slavíček, Fast and accurate excited states predictions: machine learning and diabatization, Physical Chemistry Chemical Physics 26, 4306 (2024).DOI
- [17]R. Ramakrishnan, P. O. Dral, M. Rupp, and O. A. von Lilienfeld, Big data meets quantum chemistry approximations: The Δ-machine learning approach, Journal of Chemical Theory and Computation 11, 2087 (2015).DOI
- [18]D. Lyu, V. Vinod, M. Holzenkamp, Y. M. Holtkamp, S. Maity, C. R. Salazar, U. Kleinekathöfer, and P. Zaspel, Excitation Energy Transfer between Porphyrin Dyes on a Clay Surface: A Study Employing Multifidelity Machine Learning, Advanced Theory and Simulations 8, e00271 (2025).
- [19]M. F. Langer, A. Goebmann, and M. Rupp, Representations of molecules and materials for interpolation of quantum-mechanical simulations via machine learning, npj Computational Materials 8, 41 (2022).DOI
- [20]F. Musil, A. Grisafi, A. P. Bartók, C. Ortner, G. Csányi, and M. Ceriotti, Physics-Inspired Structural Representations for Molecules and Materials, Chemical Reviews 121, 9759 (2021).DOI
- [21]E. A. Faber, A. S. Christensen, B. Huang, and O. A. von Lilienfeld, Alchemical and structural distribution based representation for universal quantum machine learning, The Journal of Chemical Physics 148, 241717 (2018).DOI
- [22]K. Gubaev, E. V. Podryabinkin, and A. V. Shapeev, Machine learning of molecular properties: Locality and active learning, The Journal of Chemical Physics 148, 241727 (2018).
- [23]W. Pronobis, A. Tkatchenko, and K.-R. Müller, Many-Body Descriptors for Predicting Molecular Properties with Machine Learning: Analysis of Pairwise and Three-Body Interactions in Molecules, Journal of Chemical Theory and Computation 14, 2991 (2018).DOI
- [24]O. Rahaman and A. Gagliardi, Deep Learning Total Energies and Orbital Energies of Large Organic Molecules Using Hybridization of Molecular Fingerprints, Journal of Chemical Information and Modeling 60, 5971 (2020).
- [25]A. Stuke, M. Todorović, M. Rupp, C. Kunkel, K. Ghosh, L. Himanen, and P. Rinke, Chemical diversity in molecular orbital energy predictions with kernel ridge regression, The Journal of Chemical Physics 150, 10.1063/1.5086105 (2019).DOI
- [26]D. Khan and O. A. von Lilienfeld, Generalized convolutional many-body distribution functional representations, Proceedings of the National Academy of Sciences 122, e2415662122 (2025).
- [27]K. R. Briling, Y. Calvino Alonso, A. Fabrizio, and C. Corminboeuf, SPAʰ M(a,b): Encoding the Density Information from Guess Hamiltonian in Quantum Machine Learning Representations, Journal of Chemical Theory and Computation 20, 1108 (2024).DOI
- [28]A. Fabrizio, K. R. Briling, and C. Corminboeuf, SPAʰ M: the spectrum of approximated Hamiltonian matrices representations, Digital Discovery 1, 286 (2022).DOI
- [29]T. Kirschbaum, B. Von Seeggern, J. Dzub iella, A. Bande, and F. Noé, Machine Learning Frontier Orbital Energies of Nanodiamonds, Journal of Chemical Theory and Computation 19, 4461 (2023).
- [30]A. Hinostroza Caldas, A. Kokorin, A. Tkatchenko, and L. Medrano Sandonas, Assessing the performance of quantum-mechanical descriptors in physicochemical and biological property prediction, Digital Discovery 5, 803 (2026).DOI
- [31]J. P. Janet and H. J. Kulik, Predicting electronic structure properties of transition metal complexes with neural networks, Chemical Science 8, 5137 (2017).
- [32]J. P. Janet and H. J. Kulik, Resolving Transition Metal Chemical Space: Feature Selection for Machine Learning and Structure–Property Relationships, The Journal of Physical Chemistry A 121, 8939 (2017).
- [33]A. Nandy, C. Duan, J. P. Janet, S. Gugler, and H. J. Kulik, Strategies and Software for Machine Learning Accelerated Discovery in Transition Metal Chemistry, Industrial & Engineering Chemistry Research 57, 13973 (2018).DOI
- [34]F. Zheng, Z. Zhu, J. Lu, Y. Yan, H. Jiang, and Q. Sun, Predicting the HOMO-LUMO gap of benzenoid polycyclic hydrocarbons via interpretable machine learning, Chemical Physics Letters 814, 140358 (2023).DOI
- [35]S. Thinius, High throughput tight binding calculation of electronic HOMO–LUMO gaps and its prediction for natural compounds, Digital Discovery 5, 203 (2026).DOI
- [36]M. M. Hasan, O. T.arkhen, S. D. Bungay, R. A. Poirier, and S. M. Islam, Predicting HOMO–LUMO Gaps Using Hartree–Fock Calculated Data and Machine Learning Models, Journal of Chemical Information and Modeling 65, 9497 (2025).DOI
- [37]F. A. Faber, L. Hutchison, B. Huang, J. Gilmer, S. S. Schoenholz, G. E. Dahl, O. Vinyals, S. Kearnes, P. F. Riley, and O. A. von Lilienfeld, Prediction Errors of Molecular Machine Learning Models Lower than Hybrid DFT Error, Journal of Chemical Theory and Computation 13, 5255 (2017).DOI
- [38]G. A. Pinheiro, J. Mucelini, M. D. Soares, R. C. Prati, J. L. F. Da Silva, and M. G. Quiles, Machine Learning Prediction of Nine Molecular Properties Based on the SMILES Representation of the QM9 Quantum-Chemistry Dataset, The Journal of Physical Chemistry A 124, 9854 (2020).DOI
- [39]M. Orsi and A. Frei, ELECTRUM: an electron configuration-based universal metal fingerprint for transition metal compounds, Digital Discovery 4, 3567 (2025).
- [40]P. B. Jørgensen, M. Mesta, S. Shil, J. M. García Lastra, K. W. Jacobsen, K. S. Thygesen, and M. N. Schmidt, Machine learning-based screening of complex molecules for polymer solar cells, The Journal of Chemical Physics 148, 241735 (2018).DOI
- [41]F. Pereira, K. Xiao, D. A. R. S. Latino, C. Wu, Q. Zhang, and J. Aires-de Sousa, Machine Learning Methods to Predict Density Functional Theory B3LYP Energies of HOMO and LUMO Orbitals, Journal of Chemical Information and Modeling 57, 11 (2017).DOI
- [42]S.-H. Hung, Z.-R. Ye, C.-F. Cheng, A.-C. Yang, B. Chen, and M.-K. Tsai, Feature-Driven Prediction of HOMO–LUMO Gaps in Transition-Metal Complexes Using the SLEET Model: A SMILES-Based Transformer Framework, Journal of Chemical Theory and Computation 21, 6410 (2025).DOI
- [43]A. Mohamed, D. P. Visc o, K. Breim aier, and D. M. Bastidas, Effect of Molecular Structure on the B3LYP-Computed HOMO–LUMO Gap: A Structure–Property Relationship Using Atomic Signatures, ACS Omega 10, 2799 (2025).DOI
- [44]O. Mahmoudi and M.-h. Kim, Predicting electronic properties of molecules: a stacking ensemble model for HOMO and LUMO energy estimation, RSC Advances 16, 12444 (2026).DOI
- [45]S. Zhang and A. A. Lapkin, Enhancing molecular property prediction of transformer models with dual graph representation, Nature Communications 10.1038/s41467-026-75005-9 (2026).DOI
- [46]M. Welborn, L. Cheng, and T. F. Miller, Transferability in machine learning for electronic structure via the molecular orbital basis, Journal of Chemical Theory and Computation 14, 4772 (2018).DOI
- [47]L. Cheng, M. Welborn, A. S. Christensen, and T. F. Miller, A universal density matrix functional from molecular orbital-based machine learning: Transferability across organic molecules, The Journal of Chemical Physics 150, 131103 (2019).DOI
- [48]T. Husch, J. Sun, L. Cheng, S. J. R. Lee, and T. F. Miller, Improved accuracy and transferability of molecular-orbital-based machine learning: Organics, transition-metal complexes, non-covalent interactions, and transition states, The Journal of Chemical Physics 154, 064108 (2021).
- [49]K. Karandashev and O. A. Von Lilienfeld, An orbital-based representation for accurate quantum machine learning, The Journal of Chemical Physics 156, 114101 (2022).DOI
- [50]C. Duan, J. P. Janet, F. Liu, A. Nandy, and H. J. Kulik, Learning from Failure: Predicting Electronic Structure Calculation Outcomes with Machine Learning Models, Journal of Chemical Theory and Computation 15, 2331 (2019).DOI
- [51]O. Çaylak and B. Baumeier, Machine Learning of Quasiparticle Energies in Molecules and Clusters, Journal of Chemical Theory and Computation 17, 4891 (2021).DOI
- [52]E. Cignoni, D. Suman, J. Nigam, L. Cupellini, B. Mennucci, and M. Ceriotti, Electronic Excited States from Physically Constrained Machine Learning, ACS Central Science 10, 637 (2024).
- [53]T. Kirschbaum and A. Bande, Transfer learning for molecular property predictions from small datasets, AIP Advances 14, 105119 (2024).arxiv.org/abs/2404.13393
- [54]Z. Qiao, A. S. Christensen, M. Welborn, F. R. Manby, A. Anandkumar, and T. F. Miller, Informing geometric deep learning with electronic interactions to accelerate quantum chemistry, Proceedings of the National Academy of Sciences 119, e2205221119 (2022).DOI
- [55]Z. He, H. Bi, B. Liang, Z. Li, H. Zhang, and Y. Wang, Frontier molecular orbital weighted model based networks for revealing organic delayed fluorescence efficiency, Light: Science & Applications 14, 75 (2025).
- [56]F. Hou, Z. Wu, Z. Hu, X. Xiao, L. Wang, X. Zhang, and G. Li, Comparison Study on the Prediction of Multiple Molecular Properties by Various Neural Networks, The Journal of Physical Chemistry A 122, 9128 (2018).
- [57]A. Coxon, Ö. H. Omar, M. Del Cueto, and A. Troisi, Predicting S1 TDDFT Energies from ZINDO Calculations Using Message-Passing ΔML with Electronically Informed Descriptors, Journal of Chemical Theory and Computation 22, 1476 (2026).DOI
- [58]K. Chen, C. Kunkel, B. Cheng, K. Reuter, and J. T. Margraf, Physics-inspired machine learning of localized intensive properties, Chemical Science 14, 4913 (2023).DOI
- [59]D. Buterez, J. P. Janet, S. J. Kiddle, D. Oglic, and P. Lió, Transfer learning with graph neural networks for improved molecular property prediction in the multi-fidelity setting, Nature Communications 15, 1517 (2024).DOI
- [60]K. T. Schütt, M. Gastegger, A. Tkatchenko, K.-R. Müller, and R. J. Maurer, Unifying machine learning and quantum chemistry with a deep neural network for molecular wavefunctions, Nature Communications 10, 5024 (2019).
- [61]M. Esders, T. Schnacke, J. Lederer, A. Kabylda, G. Montavon, A. Tkatchenko, and K.-R. Müller, Analyzing Atomic Interactions in Molecules as Learned by Neural Networks, Journal of Chemical Theory and Computation 21, 714 (2025).DOI
- [62]A. Nandy, C. Duan, M. G. Taylor, F. Liu, A. H. Steeves, and H. J. Kulik, Computational Discovery of Transition-metal Complexes: From High-throughput Screening to Machine Learning, Chemical Reviews 121, 9927 (2021).
- [63]L. J. Mantilla Santa Cruz, L. F. Faina, and J. H. De Souza Pereira, Machine learning for frontier orbital energetics: A review of HOMO–LUMO prediction methods, Carbon Trends 24, 100674 (2026).DOI
- [64]N. Grimblat, G. Klassen, and G. F. von Rudorff, Data-driven complete basis set limit estimates from a minimal auxiliary basis (2026).
- [65]J. P. Darby, J. R. Kermode, and G. Csányi, Compressing local atomic neighbourhood descriptors, npj Computational Materials 8, 166 (2022).DOI
- [66]S. De, A. P. Bartók, G. Csányi, and M. Ceriotti, Comparing molecules and solids across structural and alchemical space, Physical Chemistry Chemical Physics 18, 13754 (2016).DOI
- [67]A. S. Christensen, L. A. Bratholm, F. A. Faber, and O. Anatole Von Lilienfeld, FCHL revisited: Faster and more accurate quantum machine learning, The Journal of Chemical Physics 152, 044107 (2020).DOI
- [68]I. Batatia, D. P. Kovács, G. N. C. Simm, C. Ortner, and G. Csányi, MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields, NeurIPS 35, arXiv 2206.07697 (2022).arxiv.org/abs/2206.07697
- [69]T. Shiota, K. Ishihara, and W. Mizukami, Universal neural network potentials as descriptors: towards scalable chemical property prediction using quantum and classical computers, Digital Discovery 3, 1714 (2024).
- [70]D. P. Kovács, J. H. Moore, N. J. Browning, I. Batatia, J. T. Horton, Y. Pu, V. Kapil, W. C. Witt, I.-B. Magdău, D. J. Cole, and G. Csányi, MACE-OFF: Short-Range Transferable Machine Learning Force Fields for Organic Molecules, Journal of the American Chemical Society 147, 17598 (2025).
- [71]I. Batatia, P. Benner, Y. Chiang, A. M. Elena, D. P. Kovács, J. Riebesell, X. R. Advincula, M. Asta, M. Ayalon, W. J. Baldwin, F. Berger, N. Bernstein, A. Bhowmik, F. Bigi, S. M. Blau, V. Čaráre, M. Ceriotti, S. Chong, J. P. Darby, S. De, F. Della Pia, V. L. Deringer, R. Elijošius, Z. El-Machachi, F. Falcioni, E. Fako, A. C. Ferrari, J. L. A. Gardner, M. J. Gawkowski, A. Genreith-Schriever, J. George, R. E. A. Goodall, J. Grandel, C. P. Grey, P. Grigorev, S. Han, W. Handley, H. H. Heenen, K. Hermansson, C. Holm, C. H. Ho, S. Hofmann, J. Jaafar, K. Jakob, H. Jung, V. Kapil, A. D. Kaplan, N. Karimitari, J. R. Kermode, P. Kourtis, N. Kroupa, J. Kullgren, M. C. Kuner, D. Kuryla, G. Liepionuite, C. Lin, J. T. Margraf, I.-B. Magdău, A. Michaelides, J. H. Moore, A. A. Naik, S. P. Niblett, S. W. Norwood, N. O’Neill, C. Ortner, K. A. Persson, K. Reuter, A. S. Rosen, L. A. M. Rosset, L. L. Schaaf, C. Schran, B. X. Shi, E. Sivonxay, T. K. Stenczel, V. Svahn, C. Sutton, T. D. Swinburne, J. Tilly, C. van der Oord, S. Vargas, E. Varga-Umbrich, T. Vegge, M. Vondrák, Y. Wang, W. C. Witt, T. Wolf, F. Zills, and G. Csányi, A foundation model for atomistic materials chemistry (2024).
- [72]D. S. Levine, M. Shuaibi, E. W. C. Spotte-Smith, M. G. Taylor, M. R. Hasyim, K. Michel, I. Batatia, G. Csányi, M. Dzamba, P. Eastman, N. C. Frey, X. Fu, V. Gharakhanyan, A. S. Krishnapriyan, J. A. Rackers, S. Raja, A. Rizvi, A. S. Rosen, Z. Ulissi, S. Vargas, C. L. Zitnick, S. M. Blau, and B. M. Wood, The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models (2025).arxiv.org/abs/2505.08762
- [73]I. Kevlishvili, R. G. St. Michel, A. G. Garrison, J. W. Toney, H. Adamji, H. Jia, Y. Román-Leshkov, and H. J. Kulik, Leveraging natural language processing to curate the tmCAT, tm-PHOTO, tmBIO, and tmSCO datasets of functional transition metal complexes, Faraday Discussions 256, 275 (2025).
- [74]H. Kneiding and D. Balcells, D-Machine Learning for the Prediction of Metal Complex Properties, Chemistry – A European Journal , e71487 (2026).
- [75]Y. Cho, R. Laplaza, S. Vela, and C. Corminboeuf, Automated prediction of ground state spin for transition metal complexes, Digital Discovery 3, 1638 (2024).
- [76]A. Banjafar and G. F. von Rudorff, Property-Specific Molecular Representations via Feature-Space Transfer Compression (2026).


