WorksPhysics and Astronomy
Learning the self-energy through equivariant neural networks
Abstract
The electronic self-energy determines the single-particle Green's function and, through it, quasiparticle energies and spectral functions, but computing it from first principles is expensive and it is a difficult machine-learning target: complex-valued, matrix-valued in a local-orbital basis, and dependent on a continuous frequency. We predict the dynamical correlation part of the imaginary-frequency G_0W_0 self-energy matrix of molecules from atomic species and coordinates with an E(3)-equivariant message-passing network, the rotational covariance of the orbital-space matrix being imposed by the architecture itself and the static exchange part taken from the density-functional reference. Frequency enters either as an explicit network input or through frequency-specific output heads, and the predicted matrices are continued to the real axis in a representation in which causality and a non-negative spectral density hold by construction. The frequency-conditioned model reproduces matrix elements to below 1 meV on the imaginary axis for the systems studied here, and that accuracy survives the continuation: for water clusters, including a tetramer absent from the training set, and for chemically diverse molecules from QM9, the reconstructed spectral functions follow the G_0W_0 reference peak by peak and the frontier quasiparticle levels are recovered to a few tens of meV. Learning the self-energy on the imaginary axis is therefore a practical route to real-frequency spectra: once the model is trained, and given the density-functional reference it is built on, a new geometry costs no explicit G_0W_0 calculation.
Introduction
First-principles prediction of charged excitations — quasiparticle energies, band gaps and spectral functions — underpins the interpretation of photoemission experiments and the design of optoelectronic materials.[1] Kohn–Sham density-functional theory (KS-DFT)[2, 3] is the workhorse of electronic-structure theory, but its eigenvalues are not quasiparticle energies: local and semilocal density-functional approximations systematically underestimate the band gaps of semiconductors and insulators,[4] and the formalism gives no direct access to lifetimes or satellite structures. Generalized Kohn–Sham schemes such as hybrid functionals improve the predicted gaps[5] at the price of adjustable parameters. The approximation[6] addresses the problem at its origin, replacing the static exchange-correlation potential by the electronic self-energy — complex-valued, frequency-dependent and spatially nonlocal — from which quasiparticle energies, lifetimes and spectral functions all follow.[7, 8] and it describes charged excitations in molecules and in weakly to moderately correlated materials with broad success.[9, 10, 11, 12, 13]
is by now a standard tool, but it has not become part of high-throughput materials workflows in the way Kohn–Sham DFT has. The reason is cost. The self-energy is built from the dynamically screened Coulomb interaction, a nonlocal and frequency-dependent object
that has to be constructed and then convolved with the Green’s function at every point of a frequency grid; even with the low-scaling algorithms developed over the past decade, a calculation therefore remains considerably more expensive than the ground-state calculation it starts from.
For ground-state and mean-field quantities, machine learning has substantially alleviated the corresponding bottleneck.[14, 15] Equivariant neural networks build translational, rotational and permutational symmetries into the architecture, improving data efficiency, physical consistency and transferability.[16, 17, 18, 19, 20, 21, 22, 23] Exploiting the nearsightedness of electronic matter,[24] the Kohn–Sham Hamiltonian matrix itself can be learned in a local-orbital representation directly from the atomic geometry,[25, 26, 27, 28, 29, 30, 31] bypassing the self-consistent-field step altogether.[32] A single -equivariant network suffices for transferable orbital-space Hamiltonians in molecules and solids, hybrid-functional Hamiltonians included,[33, 34] and universal models now cover broad classes of materials, with and without spin-orbit coupling.[35, 36]
Learning dynamical many-body quantities has accordingly begun. For correlated model systems, machine learning has been used to represent the Green’s function of the Anderson impurity model,[37] to predict impurity spectral functions,[38] and to build a data-driven dynamical mean-field theory solver.[39] For molecules and materials, machine-learned Green’s functions have yielded molecular excitation spectra,[40] an -equivariant network has predicted self-energy matrices in the Lehmann representation, from which finite-temperature Green’s functions are reconstructed,[41] and a unified framework predicts the many-body self-energy at the and coupled-cluster levels.[42] The two closest to the present work differ from it in ways best stated as trade-offs. Reference 41 shares the equivariant architecture and the choice of the self-energy as target, but parametrizes it by a set of poles common to all matrix elements, so that the analytic structure is imposed by construction at the price of fixing the spectral resolution in advance; its targets are self-consistent second-order and self-energies on the finite-temperature Matsubara axis for a small set of species. Reference 42 learns the same object in a symmetry-adapted basis obtained by block-diagonalizing the mean-field Fock matrix, in which the matrix elements are rotationally invariant; its graph is built on orbitals rather than atoms, and its features are the converged mean-field matrices themselves. The network therefore emits scalars and needs no equivariant tensor algebra, which makes it computationally lighter, but it takes a converged mean-field solution as input rather than the atomic structure alone.
The route taken here is chosen not so much to remedy these differences as to match, on both sides, the representation in which the calculation is actually carried out. The numeric atom-centered orbital (NAO) basis is where equivariant message-passing models of the Kohn–Sham Hamiltonian are most at home: it is compact, localized and sparse, and the orbital-space matrix is precisely the object such networks are constructed to emit. It is also the basis in which a mature ab initio route operates, in which the self-energy is evaluated on an imaginary-frequency grid and the real-frequency spectral functions and quasiparticle energies are obtained from it by analytic continuation.[43, 44, 45, 46, 47] The grids, the matrix representation and the continuation step are all established practice there. The question the present work asks is whether a machine-learned self-energy can reproduce that route end to end. A model that predicts on the same grid replaces its single most expensive step while leaving everything downstream — the Dyson equation, the continuation and the spectral analysis — unchanged, and can be trained on, and validated against, the data these codes already produce. The imaginary axis is favourable in its own right, being smooth there and free of the pole structure of the real axis.
We accordingly develop an -equivariant message-passing framework for the imaginary-frequency self-energy of molecules, conditioned only on the atomic species, the coordinates and . The learning target is the dynamical correlation part , the static exchange part being taken from the density-functional reference. We target the self-energy rather than the Green’s function because is known exactly from the preceding DFT calculation, so that the Dyson equation recovers the non-interacting limit and the analytic structure of . The target is nonetheless harder than a Kohn–Sham Hamiltonian: it is complex- and matrix-valued, it depends on a continuous frequency, it is sensitive to nonlocal screening, and its errors must survive analytic continuation. The reference calculations reported here are all-electron.
We compare two treatments of the frequency dependence. In the fixed-frequency formulation a shared equivariant backbone feeds one frequency-specific output head for each of the 32 sampled points; in the frequency-conditioned formulation is an explicit network input, so that a single model represents as a continuous and differentiable function that can be queried away from the training grid. The two have parameter counts of the same order. The predicted matrices are continued to the real axis and inserted into the Dyson equation to obtain the Green’s function, the quasiparticle levels and the spectral function; because the continuation is ill-conditioned and can turn small matrix-element errors into spurious spectral features, we perform it in a representation in which causality and a non-negative spectral density hold by construction, described in Sec. IIC.
The results show that a symmetry-preserving architecture can learn this dynamical, matrix-valued many-body quantity accurately enough for downstream spectroscopy, and that the imaginary-axis accuracy survives the continuation. The model is validated on water monomers, dimers and clusters, including a tetramer absent from the training set, and on the chemically more diverse QM9 set.[48] On QM9 the frequency-conditioned workflow is the more accurate of the two for the HOMO, the LUMO and the gap, for the representative spectral functions and for both imaginary-axis matrix diagnostics, the margin ranging from a factor of three on the LUMO to a factor of eight on the element-wise self-energy error, while the fixed-frequency workflow remains a useful reference at comparable parameter count. Once either model is trained, obtaining for a new geometry requires no explicit calculation.
The remainder of this paper is organized as follows. Section II introduces the self-energy on the imaginary-frequency axis, the equivariant network architecture and the two frequency treatments, and the causal analytic-continuation procedure. Section III describes the reference calculations, the datasets and the training protocol. Section IV presents benchmark results for water clusters, including an out-of-distribution tetramer, and for molecules drawn from QM9. Section V summarizes the main conclusions and discusses extensions to periodic systems.
Method
In this section, we present the methodological framework developed in this work. We first review the formalism for evaluating the electronic self-energy on the imaginary-frequency axis and discuss the properties that make this quantity well suited to machine learning from first-principles data. We then introduce the equivariant message-passing architecture for predicting structure-dependent matrices and describe two strategies for incorporating frequency dependence. Finally, we show how the predicted self-energies are used to reconstruct the Green’s function and spectral function, providing a downstream physical validation of the model. The complete frequency-conditioned workflow is summarized in Figure 1.
Figure 1. Frequency-conditioned E(3)-equivariant workflow for predicting the correlation self-energy . Atomic species, coordinates and the sampled imaginary frequency are encoded as equivariant features, which the graph neural network and the matrix-construction block map onto the orbital-space self-energy. The prediction is then analytically continued to the real axis and combined with the Kohn–Sham reference to construct , from which the quasiparticle energies and the spectral function follow.
Dynamical self-energy from the approach
In this work we use the widely adopted one-shot approximation, in which the noninteracting Green’s function and the screened Coulomb interaction are both constructed from a fixed DFT reference. We work throughout on the imaginary-frequency axis. Separating the screened interaction into its bare and induced parts, , splits the self-energy into a static exchange term and a dynamical correlation term,
where the exchange part is the Fock operator evaluated with the KS orbitals,
and the correlation part is the convolution of with the induced interaction ,
Here is the noninteracting Green’s function,
where and are the wave function and orbital energy of the KS state , and is the chemical potential, which for the closed-shell molecules considered here is placed at the centre of the HOMO–LUMO gap.
Only the correlation part is taken as the learning target in this work. Two of its properties make it a well-conditioned quantity to learn: it decays to zero at large , whereas the full self-energy approaches the constant , so the target has a known asymptotic limit; and its magnitude is not dominated by the large and essentially structure-independent exchange contribution.
This choice is one of scope rather than of principle. The static part is equally amenable to machine learning: predicting static mean-field Hamiltonian matrices in a local-orbital representation is precisely what the Hamiltonian-learning literature does,[27, 32, 33] and has the same matrix structure and obeys the same rotational covariance as those targets. The closest precedent is the deep-learning treatment of hybrid functionals,[34] where the learned Hamiltonian already contains a nonlocal Fock exchange operator of exactly this form. The complete effective Hamiltonian could therefore in principle be supplied by two networks, one for the static part and one for . We do not pursue that here because the static part is not the computational bottleneck: it follows from the occupied KS orbitals of the same DFT reference at considerably lower cost, whereas the dynamical, screening-dependent part is what makes a calculation expensive. The reference self-energies are evaluated with NAO basis sets combined with the resolution-of-identity (RI) technique.[43, 44] Within RI, products of NAOs are expanded in an auxiliary basis,
where are the RI expansion coefficients, label localized NAOs, and label auxiliary basis functions . This expansion allows the polarizability and the screened interaction to be represented in the auxiliary basis. All calculations reported here use the localized RI variant (RI-LVL),[49, 50] which restricts the auxiliary functions to the atoms associated with each orbital pair and thereby reduces the computational cost while retaining high accuracy. In this representation, the orbital-space correlation self-energy becomes
where is the RI coefficient with one index contracted into the KS space through the KS eigenvector matrix . The matrices , expressed in the NAO basis, are the quantities predicted by the neural network; their real and imaginary parts are treated as two separate real, equivariant targets.
The network therefore supplies and nothing else. Every other ingredient of the downstream analysis — the exchange self-energy , the KS eigenvalues and eigenvectors, the overlap matrix, and the exchange-correlation potential — is taken directly from the same DFT reference from which and are constructed. A ground-state DFT calculation is thus still performed for each structure; what the model replaces is the step. These static ingredients enter the quasiparticle Hamiltonian only through the combination , which is itself an established target for equivariant Hamiltonian learning, so the division is one of scope rather than of principle; we return to it in Sec. V. To make the rotational covariance of these matrices explicit, we resolve the orbital indices as and , where denote atomic centers, at positions , radial channels, and the NAOs are products of a radial function and a real spherical harmonic, . Let act on the molecular geometry, so that the atom at moves to while retaining its label, and let denote the correlation self-energy matrix of the rotated structure. Then each orbital-pair block with fixed satisfies
where is the real Wigner rotation matrix acting on the degree- real spherical harmonics, so that . The blocks are in addition invariant under translations of the molecule and under permutations of identical atoms. Equation (7) is the exact constraint that the architecture of Sec. II B enforces by construction. The continuation of the predicted self-energies to the real-frequency axis and the subsequent reconstruction of spectral quantities are described in Sec. II C.
Equivariant neural-network architecture and workflow-dependent training
At each imaginary-frequency point the learning target is the orbital-resolved self-energy matrix , which we predict from the molecular geometry with an -equivariant message-passing network, the middle block of Figure 1. Each molecule is treated as an effectively fully connected graph whose nodes are the atoms , and whose edges carry the pairwise geometric relations, the localized orbitals keeping the lowercase indices , of Sec. II. The equivariant backbone is taken over unchanged from the HamGNN construction for graph-based prediction of orbital-space matrices, [23, 33, 51] and is summarized below only as far as the rest of this section requires. What the present target adds to it is of three kinds. The matrix is complex rather than real, so that its real and imaginary parts are carried by separate branches throughout. It depends on a continuous frequency argument instead of being static, which is what the frequency encoding introduced below adds to the node features. And it is an intermediate quantity rather than an end point: the predicted matrix is passed to the analytic continuation of Sec. II C, so that the accuracy which ultimately matters is that of the continued spectral quantities and not of the matrix alone.
Under a rotation a feature of angular momentum transforms with the Wigner -matrix, , and such features are combined by Clebsch–Gordan tensor products, following the standard tensor-field construction for three-dimensional equivariant networks,[18, 23]
Here , , label feature channels within a given angular-momentum sector, and are trainable channel-mixing maps at fixed and , and combines the coupled channels into the output representation. Each atom accordingly carries a set of equivariant features that are irreducible representations of , labelled by angular momentum and parity with channels each; Eq. (8) is the only operation by which they are combined, and it is retained only when and .
The remaining ingredients are those of Ref. 33 and are stated here only to fix notation. The features are initialized from a learnable embedding of the atomic species, which populates the even-parity scalar channels and leaves all others empty; every higher irrep is generated subsequently by equivariant message passing, , in which is the coupling of Eq. (8) and the edge features combine a learnable radial embedding of the interatomic distance with the spherical harmonics of the bond direction. Node features are then updated by residual blocks with a gated nonlinearity, which acts on the scalar channels and gates the higher- channels by a learned scalar, so that the update is nonlinear without breaking equivariance. On-site and off-site atomic-pair features and are formed from the updated node features by further tensor-product paths, the off-site path additionally coupling in the edge feature .
The parity label is not redundant, and it is what the present target makes essential. A spherical harmonic evaluated on an interatomic direction obeys , so the edge features enter with ; coupling two of them, however, already generates the opposite assignment, the elementary case being the cross product of two polar vectors, which carries with . Angular momentum alone therefore does not fix the behaviour of a feature under inversion, and the two labels must be carried separately.
Finally, an inverse tensor product converts each coupled feature into an orbital-space matrix block
Here is the degree- part of the pair feature and the operator performs a Clebsch–Gordan-based projection from it onto the magnetic components of orbital pairs with angular momenta and , and is the inverse of the coupling in Eq. (8). This construction guarantees the rotational covariance of the resulting matrix blocks and provides the symmetry-preserving map from molecular geometry to . The parities are fixed by the same construction rather than chosen: the NAO of angular momentum is a radial function times a real spherical harmonic and therefore carries parity , so the block connecting orbitals of angular momenta and transforms with parity , and every irreducible component extracted from it in Eq. (9) must carry that parity. An – block thus consists of a single component and a – block of , and components. The target therefore dictates which irreducible representations the output layer may emit, and a network that tracked angular momentum alone could produce blocks that do not transform as a self-energy matrix at all. [23, 33, 41]
The off-site pair features are directional and are not symmetric under exchange of and , whereas the self-energy matrix is. The blocks delivered by Eq. (9) are therefore symmetrized,
which for an on-site block reduces to averaging it with its own transpose. The symmetrization is part of the output layer, so holds for the predicted matrix by construction, and it acts on the real and imaginary parts alike.
The contrast with the symmetry-adapted, block-diagonal strategy of Ref. 42, outlined in § 1, is made precise here. There the rotational degrees of freedom are removed from the target by the basis transformation itself, so the network learns rotation-invariant blocks; here the matrix is predicted directly in the NAO basis and the transformation law of Eq. (7) is imposed at every stage, each operation being a Clebsch–Gordan coupling of irreducible features and the final projection returning blocks that transform with the Wigner -matrices of the two orbital angular momenta. Equivariance is thus a property of the architecture rather than of the representation, and applies in the same way to on-site and off-site blocks.
The blocks are assembled into the complete orbital-space self-energy matrix used for the Green’s-function reconstruction. In the fixed-frequency workflow the shared backbone feeds 32 frequency-specific output heads, one per , with separate branches for the real and imaginary parts; the reported model shares all backbone layers, so that the frequency dependence enters only at the final projection. In the frequency-conditioned workflow the imaginary-frequency coordinate is supplied to a single network as an additional scalar input. It is encoded on a logarithmic frequency scale,
where and are the smallest and the largest sampled frequencies of the imaginary-frequency grid specified below in Eq. (24), held fixed rather than recomputed per batch, and the offset keeps the mapped coordinate away from the nodes of the sine basis, at which the feature vector would vanish identically. The logarithmic coordinate is dictated by the sampling grid rather than by the self-energy: that grid spans six orders of magnitude and is approximately uniform in , so distributes the sampled points evenly over the interval on which the sine basis varies, and the encoding both separates neighbouring frequencies and stays smooth between them. The resulting vector is mapped by a learnable linear layer onto an even-parity scalar feature,
which has the same number of channels as the elemental embedding and is added to it. The elemental initialization is thereby replaced by
the components with still vanishing at the input. Since is an even-parity scalar, this addition leaves the transformation law of the node features untouched, and the frequency dependence is carried into every subsequent operation by the message passing. A single network therefore represents as a continuous function of its frequency argument.
The two frequency treatments differ not only in how frequency is supplied to the network but also in their training objectives. Both workflows include direct supervision of the self-energy in the NAO basis and an auxiliary loss evaluated after transformation to the KS basis. The frequency-conditioned model additionally includes a frequency-derivative loss that promotes a smooth and accurate dependence on . The two objectives are
The individual loss terms are defined as
where runs over all matrix elements, with the real and imaginary parts accumulated separately. All energies are in eV, and the weights and are fixed empirically and absorb the differing units of the three terms. is obtained by transforming the predicted or reference NAO-basis self-energy into the KS-orbital basis,
where holds the KS eigenvectors in the NAO basis, , and satisfies with the NAO overlap matrix. The KS-basis loss helps control the propagation of self-energy errors into downstream quasiparticle and spectral quantities and is therefore included in both workflows. By contrast, is used only in the frequency-conditioned workflow, because only this model represents as a differentiable function of a continuous frequency input. In the calculations reported below, we set eV and .
Causal analytic continuation and spectral reconstruction
Analytic continuation connects the correlation self-energy learned on the imaginary-frequency axis to real-frequency spectral quantities. Throughout this subsection denotes ; the exchange part is static and enters only through the Dyson equation below.
For self-energies obtained directly from a calculation this is a well-controlled step. The imaginary-axis data are deterministic and carry no statistical noise, and the rational, or Padé, approximants introduced by Vidberg and Serene[52] reproduce valence quasiparticle energies robustly; this is established practice in numerical-atom-centered-orbital implementations[43, 44] and underlies the molecular benchmarks of the GW100 test set.[10, 53] The Bayesian techniques developed for the stochastic imaginary-time data of quantum Monte Carlo, of which the maximum-entropy method[54] is the best known, address a different problem and are not required here.
What changes in the present context is that the self-energy is predicted rather than computed. Analytic continuation is ill-conditioned with respect to perturbations of the input: the coefficients of a high-order rational approximant depend very sensitively on the sampled values,[55] so that matrix-elements errors small enough to be immaterial on the imaginary axis can be amplified into spurious poles and unphysical structure on the real axis. The prediction error of the network therefore propagates into the spectra in a way that has no counterpart when the reference data are continued directly, and it is this, rather than any deficiency of the Padé construction itself, that motivates the construction introduced below.
Strategies for stabilizing the fit against such perturbations include averaging over approximants of different order,[56] enforcing causality by construction through a Nevanlinna representation,[57] and robust pole estimation.[58] We follow the second of these: rather than fitting a rational function whose poles are free to move, we impose the spectral representation of the self-energy on a fixed grid of real frequencies and solve for non-negative weights, so that causality and a non-negative spectral density hold by construction rather than being recovered by penalization. We also tested the AAA algorithm,[59] the standard modern approach to rational approximation, and found no improvement for the present data. The continuation is carried out in the Kohn–Sham basis: the predicted matrix is transformed with Eq. (16) and its diagonal elements are continued independently. Each is represented by its spectral, or Lehmann, form
which is discretized on a grid of real frequencies that is fixed in advance and identical for every matrix element, every structure and both workflows,
We use nodes spanning eV with a uniform spacing eV; this choice is discussed together with the regularization strength below.
The residues are determined from the sampled imaginary-frequency values. With the kernel , the real and the imaginary parts of the data are fitted to the same residue vector,
with , , and the second-difference operator divided by , which penalizes rapid variation of the residues from node to node; we use . (19) is a non-negative least-squares problem and is solved as such.
The spacing and the penalty strength are not independent, and were fixed together by two criteria. The first is independent of the neural network: the reference self-energies are continued with a given pair and the resulting quasiparticle energies are compared with those returned directly by the reference implementation of § 2, which isolates the error contributed by the continuation itself. The second compares the continued target with the continued prediction, and therefore measures how faithfully a given imaginary-axis matrix error is transmitted to the real axis. Both criteria are optimized by the values used throughout, eV and . Coarser spacings leave the quasiparticle energies and the principal spectral peaks unchanged but degrade the weak satellite features, whose amplitudes lie an order of magnitude below the quasiparticle peaks, whereas appreciably smaller penalties admit numerous narrow, unphysical structures in the continued self-energy. The accuracy floor set by the first criterion is quantified in § 2.
Three properties follow by construction rather than by tuning. First, because the node positions are fixed, Eq. (19) is a convex quadratic program with a unique solution: there are no local minima, and the instability of rational fits with free denominators—in which a denominator zero can approach the real axis—cannot occur.
Second, a single analytic expression supplies both the real and the imaginary part, so the two satisfy the corresponding Kramers–Kronig relation automatically. Third, evaluating Eq. (18) at with gives
for any , so that each continued matrix element is causal and carries non-negative spectral weight whatever the prediction error. We use eV. Equation (20) is a statement about the continued diagonal elements; since the off-diagonal elements are not continued, it does not amount to positive semidefiniteness of the full matrix spectral density. No explicit high-frequency constant is included in Eq. (18).
The number of nodes in Eq. (18) is a discretization parameter and not a measure of model flexibility. The positions are never adjusted, so refining the grid approaches the integral in Eq. (17) rather than adding fitting freedom, and it is this that keeps the problem convex; a representation with few but freely placed poles would be neither. What the sampled frequencies determine is correspondingly limited: each of them probes through a Lorentzian of width , the widest of these exceed that of and therefore measure essentially the same moment, and the narrowest are finer than . The data consequently fix only a small number of effective degrees of freedom, with the non-negativity constraint and the smoothness penalty determining the remainder. We therefore use Eqs. (17) and (18) as a stable and causal representation of the continued self-energy, and do not interpret the individual residues as a resolved spectral density.
Quasiparticle energies are obtained from the continued self-energy by solving
by fixed-point iteration. All quantities in what follows are expressed in the orthonormal KS representation, in which only the diagonal of is continued. The resulting self-energy is inserted into the Dyson equation
where removes the exchange-correlation contribution already included in the DFT reference. The static exchange self-energy of Eq. (2) is evaluated directly from the occupied KS orbitals of that reference rather than predicted by the network, and being frequency independent it requires no analytic continuation; the continuation is applied to alone. The Green’s function is evaluated slightly off the real axis, at , with an imaginary shift eV; Kohn–Sham states lying more than 20 eV above the chemical potential are omitted from the trace. The two small imaginary parts entering the calculation play different roles and are not to be confused: in Eq. (20) is the distance of the poles of the continued self-energy from the real axis, and therefore belongs to itself, whereas broadens the Green’s function built from it. The predicted NAO-basis matrix is already transpose symmetric by Eq. (10), so no symmetry has to be restored before the transformation to the Kohn–Sham basis and the continuation. In the orthonormal KS representation, the spectral function is defined as
Because Eq. (23) traces over the KS basis, is a total spectral function, that is, an interacting density of states, rather than a state-resolved quantity. The identical continuation procedure, with identical parameters, is applied to the reference and to the predicted self-energy of every structure, so that any difference between the resulting spectra originates in the self-energy itself and not in the continuation. The accuracy of that procedure can be quantified independently: continuing the reference self-energies and comparing the resulting quasiparticle energies with those produced directly by the reference implementation gives mean absolute deviations of 8.5, 7.8 and 15.5 meV for the HOMO, the LUMO and the gap over the QM9 evaluation set. These values are the accuracy floor of the continuation step, against which the prediction errors of Sec. IV should be read.
Computational details
Reference calculations
The reference correlation self-energies were generated with the periodic implementation in FHI-aims, an all-electron electronic-structure package built on NAOs, in which Hartree–Fock, hybrid-functional and many-body perturbation-theory methods are implemented through the resolution-of-identity technique. [44] [60] Each molecule or cluster was placed in a cubic cell with a side length of 21 Å, and the Brillouin zone was sampled at the point only. The Kohn–Sham starting point was obtained with the PBE exchange-correlation functional, [61] using scalar-relativistic atomic ZORA and Gaussian smearing of width 0.010 eV. The self-energy was evaluated on the modified Gauss–Legendre imaginary-frequency grid used by the FHI-aims implementation, [43]
where the are the nodes of the 32-point Gauss–Legendre quadrature on . The Mobius transformation in Eq. (24) maps onto , with the scale parameter fixing the midpoint ; half of the nodes therefore fall below Ha and the grid is much denser at low than at high frequencies. The resulting frequencies span to 365 Ha to eV). All 32 points are retained and used as training targets and as inputs to the analytic-continuation procedure. The NAO basis consisted of the minimal basis of occupied free-atom orbitals augmented by the standard first tier of hydrogen-like functions, which amounts to for H and for C and O. No higher tiers were included.
Dataset construction and partitioning
FHI-aims delivers in the KS-orbital basis; the NAO-basis learning targets were obtained by inverting Eq. (16), . The resulting matrices were decomposed into on-site and off-site blocks according to the atomic-orbital indexing, and each structure was represented as an atomistic graph whose nodes are the atoms and whose edges are the atom pairs carrying a nonzero self-energy block; matrix elements with absolute values below Ha were discarded. Each sample consists of one molecular geometry paired with one imaginary-frequency point, the network inputs being the atomic numbers, the Cartesian coordinates, the pairwise geometric information and . The water geometries were generated by first-principles molecular dynamics with ABACUS, a density-functional package that combines NAOs with pseudopotentials for large-scale simulations. [62, 63] Each system was propagated in a fixed cubic cell of side 20 Å in the canonical ensemble at 400 K, using a Nosé–Hoover chain with a single thermostat of oscillation frequency 0.025 fs and a time step of 1 fs. Each trajectory ran for steps (50 ps), of which the first 5 ps were discarded as equilibration. Snapshots were taken at least ten steps apart to reduce temporal correlation, and near-duplicate geometries were removed by a structural-distance criterion before the dataset was assembled. The same protocol was used for the monomer, the dimer and the clusters. The QM9 geometries, by contrast, were taken unchanged from the benchmark dataset; [48] 150 molecules were selected and used with their original compositions and equilibrium structures, without molecular dynamics or further perturbation. Reference self-energies for all selected structures were then computed with FHI-aims as described above.
Separate datasets were constructed for the water-monomer, water-dimer, water-cluster and QM9 benchmarks. The monomer and dimer datasets contained 200 geometries each, the cluster dataset 100 geometries for each of HO, (HO) and (HO), and the QM9 dataset 150 molecules; with 32 imaginary-frequency points per geometry these amount to 6400, 9600 and 4800 structure-frequency samples, respectively. The tetramer was held out entirely as an out-of-distribution test system. Unless otherwise stated the geometries, rather than the individual frequency samples, were divided at random into training, validation and test subsets in an 8:1:1 ratio, so that every structure carries all of its frequency points into a single subset; the QM9 test set is accordingly 15 complete molecules. The real and imaginary parts of were learned by separate branches throughout, so that the fixed-frequency workflow uses one shared backbone with 32 output heads per component and the frequency-conditioned workflow one network per component, with parameter counts of the same order.
Network architecture, training and post-processing
The network used three equivariant message-passing layers with feature angular momenta up to . The radial basis of the edge features consisted of 64 Bessel functions multiplied by a cosine cutoff envelope, and the radial embedding was a multilayer perceptron with two hidden layers of 64 neurons each. The atomic graph was built from atomic radii of 6.0 Bohr, scaled by a factor of 1.01, within an overall interaction cutoff Bohr ( Å); since this exceeds the diameter of every molecule and cluster studied here, every atom pair carries an edge, which is the sense in which the graph of Sec. II B is effectively fully connected. The feature multiplicities are not uniform across irreducible representations. Written in the irreducible-representation notation of e3nn,[20] in which denotes channels of angular momentum and parity , the hidden representation is
amounting to 263 channels and 877 components per atom. Almost half of the channels are scalars, and since the channel mixing at fixed scales as , about three quarters of those weights sit in the two scalar sectors, while the channels with — which serve only as intermediates in the message passing, the largest orbital block requiring — account for little more than one per cent. Each trained network contains approximately trainable parameters.
The frequency-conditioned model was optimized with and the fixed-frequency reference with . Optimization used AdamW[64] with AMSGrad, , , and no weight decay. The learning rate started at and was halved by a plateau scheduler after five epochs without improvement of the training loss, down to a floor of . Training ran for 200 epochs with an early-stopping patience of 30 epochs, with a batch size of 16 for the QM9 benchmark. Training was performed in single precision on a CPU using 48 threads. The predicted imaginary-frequency self-energy was analytically continued to the real axis with the procedure of Sec. II C, and the continued self-energy was used to construct the Green’s function and evaluate the spectral function.
Results
The dataset construction, first-principles reference calculations, and partitioning protocol are described in § 2. Unless stated otherwise, the water-monomer, water-dimer, and water-cluster results below are obtained with the frequency-conditioned workflow, which is the primary model of this work. The fixed-frequency workflow is introduced in the QM9 subsection as a reference for a direct comparison on the more chemically complex dataset.
Single water molecule HO
For a single water molecule, we evaluated the frequency-conditioned model on the molecular geometries generated by molecular dynamics (MD) using the ABACUS package.[62, 63] The corresponding self-energy matrices were computed with FHI-aims. Figure 2 reports the element-wise absolute error of the complex correlation self-energy matrix in the NAO basis, with the real and imaginary parts combined as . At each frequency the average runs over the test structures and over all matrix elements. The mean error stays between about 0.05 and 0.08 meV across the whole grid, with a grid-averaged value of 0.065 meV, and the upper edge of the one-standard-deviation band remains below 0.22 meV. Neither grows systematically from one end of the frequency range to the other, although the spread is somewhat wider at the upper end of the grid.
Figure 2. Frequency-resolved HO correlation self-energy matrix-error statistics for the frequency-conditioned model. The horizontal axis is the index of the imaginary-frequency point and the vertical axis is the NAO-basis complex matrix-element MAE in meV, with real and imaginary errors combined as . At each frequency the curve is the mean over the test structures and over all matrix elements, and the shaded band is standard deviation of the same distribution.
As a further validation, we analytically continue the self-energy to the real-frequency axis and construct the Green’s function . The spectral function is then evaluated according to Eq. (23), allowing a direct comparison of quasiparticle spectral functions. As shown in Figure 3, the predicted and the reference spectra are visually indistinguishable across the whole range displayed, in the occupied manifold as well as above the gap. Each reference peak is reproduced at the same position and with the same weight, and no spurious structure appears between them: the continuation neither displaces the quasiparticle poles nor generates satellite features that are absent from the target. This is what the imaginary-axis errors of Figure 2 cannot by themselves establish. A sub-meV matrix-element error is only useful if the continuation does not amplify it, and for this system it does not.
Figure 3. HO quasiparticle spectral function reconstructed from the target and predicted correlation self-energies with the causal continuation of Sec. II C. The blue and orange curves denote the target and the neural-network prediction, respectively. A spectral broadening of is used in the Green’s-function reconstruction.
The approach therefore works for the simplest system: is reproduced on the imaginary axis to well below 1 meV per matrix element, and — which is not automatic, given that the continuation of Sec. II C is ill-conditioned with respect to perturbations of its input — that accuracy survives the continuation to the real axis. Two questions follow. Does the same hold once the orbital space grows and the bonding environment becomes more varied, and does a model trained on one set of structures carry over to structures of a kind it has not seen? The next two subsections take these up in turn: first the water-cluster series, in which the system grows while the chemistry stays the same, and then the QM9 set, in which the chemical composition itself varies.
Water-cluster scaling and out-of-distribution test
The water-cluster series tests a different question: whether a model trained on the smaller members of a homologous series transfers to a larger one. Clusters up to the trimer are represented in the training set, using the datasets and the partitioning protocol of § 2 B. The tetramer is held out entirely, so that its self-energy is predicted by a model that has never seen a cluster of that size; it probes size transferability rather than interpolation within a sampled distribution.
Figure 4 compares the target and predicted traces after averaging over randomly selected structures from the corresponding test sets. The upper and lower panels show the real and imaginary trace components, respectively, and both components follow the reference over the complete imaginary-frequency grid. The mean absolute deviations between the averaged target and prediction traces are approximately 0.004, 0.021, 0.022, and 0.099 eV for the real part and 0.004, 0.009, 0.030, and 0.085 eV for the imaginary part, for , , , and , respectively.
Figure 4. Test-set averaged traces of the correlation self-energy for HO, (HO), (HO), and (HO). For each cluster size, traces from randomly selected structures in the corresponding test set are averaged before plotting. The upper and lower panels show and , respectively; filled circles and open squares denote the target and prediction. The close agreement of the averaged traces, including for the larger clusters, illustrates the low trace-level error and the model’s generalization across molecular geometries and cluster sizes.
The trace is extensive, so these values are not comparable across cluster size as they stand. Per water molecule they become 0.004, 0.011, 0.007 and 0.025 eV for the real part and 0.004, 0.005, 0.010 and 0.021 eV for the imaginary part, so that the per-molecule deviation grows with cluster size by roughly a factor of five from the monomer to the tetramer. It nevertheless remains small on the scale of the traces themselves, which reach several eV in the real part and of order a hundred eV in the imaginary part. The trace alone, however, is blind to the off-diagonal elements and admits cancellation between diagonal errors of opposite sign, so we also quote the element-wise error of , defined exactly as in Figure 2. Averaged over the test structures and over the whole frequency grid it is 0.065, 0.173, 0.131 and 0.607 meV for through . All four lie well below 1 meV, and the dimer and trimer are within a factor of three of the monomer. The tetramer, which the model has never seen, is about an order of magnitude above the monomer; this is the price of extrapolating beyond the training set, and it is paid on a quantity that is still an order of magnitude smaller than the accuracy floor of the continuation itself. The accuracy reached on the monomer, dimer and trimer therefore carries over to a cluster size the model has never seen, with a degradation that is gradual rather than abrupt — the behaviour a local, symmetry-preserving representation of is expected to provide, and what makes the approach useful for systems larger than those that can be afforded at the level. One caveat attaches to the trace values: they are deviations of ensemble-averaged traces, so errors may cancel between structures; the element-wise figures quoted above are free of that ambiguity.
The corresponding real-frequency spectra are shown in Figure 5. Panel (a) is the water-dimer spectrum, and panels (b) and (c) those of the trimer and the tetramer test structures. For the dimer and the trimer the predicted and the target curves overlap throughout, as they did for the monomer. For the tetramer the peak positions remain aligned, but the heights of some of the stronger peaks differ visibly, the prediction overshooting the reference. The deviation is therefore one of spectral weight rather than of quasiparticle energy, which is the benign way for it to appear, since the frontier levels are read from peak positions; and it is consistent with the element-wise error quoted above being about an order of magnitude larger for the tetramer than for the monomer.
Figure 5. Spectral-function comparisons for the water-cluster systems. Panel (a) shows the water-dimer spectral function corresponding to the dimer result presented in this work, while panels (b) and (c) show the spectra of and , respectively. In each panel, the target and predicted spectra are reconstructed from the reference and the frequency-conditioned neural-network correlation self-energy. The horizontal coordinate is . The spectra use the self-energy broadening of Eq. (20) and the Green’s-function imaginary shift .
QM9 Dataset
We next consider the more complex molecular set drawn from QM9, using the dataset and partitioning protocol described in § 2[48].
For these larger molecules, which typically contain more than ten atoms, the learning task is substantially more demanding than in the water-cluster tests. The fixed-frequency workflow is compared directly with the frequency-conditioned model here. The aggregate statistics and the structure-by-structure comparison below use all fifteen molecules of the QM9 test set; the full reconstruction pipeline completed on every one of them for both workflows. Four of the fifteen, drawn at random, are shown as representative spectra in Figure 7, and one as the orbital-resolved map of Figure 9. Rather than comparing all quasiparticle levels, we focus on the frontier quantities HOMO, LUMO, and the resulting gap, together with representative spectral functions, because these observables are both physically transparent and directly relevant to downstream spectroscopy.
Figure 7. Spectral-function comparisons for four representative QM9 test structures, obtained with the causal continuation of Sec. II C. Panels (a)-(d), in reading order, correspond to 2-propanol, ; propanal, ; 1-propanol, ; and methoxyethane, . Each panel compares the GW target with the predictions of the frequency-conditioned and fixed-frequency workflows.
Figure 9. Orbital-resolved absolute correlation self-energy errors for a QM9 test structure at the representative high-frequency point , with eV. Panel (a) shows the frequency-conditioned workflow, with a matrix-element MAE of 0.27 meV, whereas panel (b) shows the fixed-frequency workflow, with a matrix-element MAE of 0.80 meV. Both panels use the same logarithmic color scale.
Figure 6 shows that the frequency-conditioned workflow has the lower mean absolute error for all three frontier observables on the common QM9 evaluation set of fifteen structures. The mean absolute errors for the frequency-conditioned and fixed-frequency workflows are 0.019 and 0.065 eV for the HOMO, 0.023 and 0.068 eV for the LUMO, and 0.027 and 0.123 eV for the gap. The structure-by-structure panel resolves how uniform this advantage is: the frequency-conditioned workflow has the smaller error for 9, 10 and 12 of the fifteen test structures for the HOMO, the LUMO and the gap, respectively, and for 12 of fifteen when the three frontier errors are averaged. The margin is therefore clearest for the gap and least pronounced for the HOMO, where the two workflows are closer on an appreciable fraction of the structures.

Figure 6. Frontier quasiparticle-level errors for the common QM9 evaluation set, obtained with the causal continuation of Sec. II C. The left panel reports the mean absolute errors of HOMO, LUMO, and the gap, while the right panel resolves the same errors structure by structure for the two workflows.
The same conclusion is visible at the spectral-function level. As shown in Figure 7, the frequency-conditioned workflow follows the GW target more closely across the four structures and reproduces the principal peak positions and spectral weights with smaller deviations. Across the four panels the curve-wise mean absolute error of the frequency-conditioned workflow ranges from 0.009 to 0.020 eV, against 0.086 to 0.092 eV for the fixed-frequency workflow; averaged over the four structures, 0.016 versus 0.089 eV, a factor of 5.6. The fixed-frequency error is almost independent of the molecule, whereas that of the frequency-conditioned model varies with it, which places the residual discrepancy of the fixed-frequency workflow in its representation of the frequency dependence rather than in the chemistry of the individual structures. To determine whether the distinction between the two workflows is already present at the imaginary-frequency self-energy level, we further compare their matrix-element error on the same QM9 evaluation set.
As shown in Figure 8, the frequency-conditioned workflow has the lower error throughout the displayed frequency range in both panels. The mean element-wise MAE is 0.55 versus 4.55 meV, and the mean relative Frobenius error is 0.0059 versus 0.049, for the frequency-conditioned and fixed-frequency workflows, respectively. Thus, on the reported imaginary-axis diagnostics, the frequency-conditioned model is better by approximately factors of 8.3 and 8.4, while also retaining a continuously queryable frequency representation. The margin is not uniform across the grid. It is widest at low and intermediate frequencies and narrows at the upper end: in panel (b) the relative Frobenius error of the frequency-conditioned model rises over the last part of the range, while that of the fixed-frequency reference falls, so that the two differ by about a factor of two at the highest sampled frequency rather than by an order of magnitude. This is the region in which is smallest, and therefore the region that an absolute-error training objective constrains least; it is also the region that the analytic high-frequency behaviour discussed in § 2.4 would address.
Figure 8. Frequency-resolved comparison of QM9 correlation self-energy errors for the frequency-conditioned and fixed-frequency workflows on the common evaluation set. Panel (a) shows the element-wise mean absolute error of in eV, whereas panel (b) shows its relative Frobenius-norm error.
The orbital-resolved comparison in Figure 9 provides a direct view of how these errors are distributed within the self-energy matrix of the CHO test structure. At eV, the matrix-element MAE is 0.27 meV for the frequency-conditioned workflow and 0.80 meV for the fixed-frequency workflow. Because the two panels share the same logarithmic color scale, the approximately threefold reduction reflects a genuine decrease in the orbital-resolved errors rather than a change in color normalization. This panel is a local diagnostic for one structure and one frequency; the aggregate imaginary-axis comparison is given in Figure 8.
The comparisons in Figure 6–Figure 9 consistently favor the frequency-conditioned workflow on the present QM9 evaluation: it has lower mean errors for the HOMO, LUMO, and gap, smaller deviations in the representative spectral functions, and lower values in both reported imaginary-axis matrix-error diagnostics. The orbital-resolved heatmap provides an additional local example, with a matrix-element MAE of 0.27 meV versus 0.80 meV for the fixed-frequency reference. Thus, explicit frequency conditioning improves the predictive accuracy observed here while also providing a smooth, continuously queryable representation; the shared-backbone fixed-frequency workflow remains a useful reference at comparable parameter scale.
Summary and Prospects
The question posed in § 1 was whether a machine-learned self-energy can reproduce the route by which ab initio is carried out in a numerical atom-centered orbital basis: evaluate on an imaginary-frequency grid, then continue it to the real axis. The answer obtained here is that it can. A frequency-conditioned -equivariant network predicts the imaginary-axis correlation self-energy of molecules from the atomic structure alone to well below 1 meV per matrix element, and — which is the part that is not automatic — those matrices survive the continuation: inserted into the same Dyson equation and the same spectral analysis that the reference calculations use, they reproduce the spectral functions peak by peak and the frontier quasiparticle levels to a few tens of meV, for water clusters including a tetramer absent from the training set and for chemically diverse molecules from QM9. The learned self-energy is therefore a drop-in replacement for the most expensive step of an established workflow rather than a parallel construction that has to be validated on its own terms. Once the model is trained, obtaining for a new geometry requires no explicit calculation at all.
The fixed-frequency formulation is used as a targeted reference rather than as a second main method. Its comparison with the frequency-conditioned model on QM9 shows that explicit frequency conditioning provides both a smooth, continuously queryable representation and lower errors for the frontier levels, representative spectra, and the reported imaginary-axis matrix diagnostics. The shared-backbone fixed-frequency workflow remains a useful reference with a comparable parameter scale, while the present QM9 results support frequency conditioning as the preferred workflow for the downstream observables examined here.
Several directions may further improve the method. The high-frequency behaviour of is known analytically, and imposing the tail as a constraint rather than leaving it to be learned from a grid that is sparsest there — would target the region in which the present models are least accurate. A related point concerns the loss: because falls by orders of magnitude across the grid, an absolute-error criterion is dominated by the low-frequency points, and a relative error, or a per-frequency normalization of the targets, would give the high-frequency region a weight commensurate with its role in the analytic continuation. We leave a systematic study of both to future work. Beyond this, the static and the dynamical parts of the quasiparticle Hamiltonian could be supplied by two networks trained together. The static ingredients that the present model takes from the DFT reference enter only through the single combination , a Hermitian one-body operator in the same NAO basis and therefore a single equivariant target of exactly the kind that Hamiltonian models already predict; [23, 33] the present model supplies the dynamical part . Together with the basis overlap matrix, which is purely geometric, the two would assemble the full frequency-dependent effective Hamiltonian from predicted quantities alone, with no self-consistent-field calculation. A predicted self-energy that is accurate enough on the imaginary axis further opens the way to self-consistent schemes in which it is re-inserted into the Dyson equation. A further direction concerns the cost of the architecture itself, which is dominated by the Clebsch–Gordan tensor products and grows steeply with . Cheaper equivalents have recently been developed, based either on local rotations or on Gaunt tensor products evaluated with fast Fourier transforms,[65, 66] and adapting them to the parity structure of the orbital blocks would reduce the cost of both workflows.
Extending the present framework to solids is another natural next step. Periodic systems introduce additional complications, such as -point sampling and Brillouin-zone integration, which will require corresponding architectural refinements. Reference data for such systems can be generated with low-scaling NAO implementations, which agree with the present all-electron results to within numerical noise while scaling far more favourably for large bulk cells.[45]
If successful, adaptation to periodic materials would open the way to high-throughput screening and accurate electronic-structure calculations at a fraction of the cost of conventional ab initio methods. More broadly, it would further integrate many-body physics and symmetry constraints into data-driven electronic-structure modeling.
Acknowledgments
We thank Lei Wang and Zechen Tang for helpful discussions. We acknowledge the funding support from the Advanced Materials–National Science and Technology Major Project (Grant No. 2025ZD0618401) and the National Natural Science Foundation of China (Grant Nos. 12374067, 12134012 and 12188101). This work was also supported by the Strategic Priority Research Program of the Chinese Academy of Sciences under Grant No. XDB0500201, by the National Key Research and Development Program of China (Grant Nos. 2022YFA1403800 and 2023YFA1507004), and by the robotic AI-Scientist platform of the Chinese Academy of Sciences. M.-Y.Z. acknowledges support from the IOP–Humboldt Postdoctoral Fellowship in Physics at the Institute of Physics, Chinese Academy of Sciences (Grant No. 202402).
References
References
- [1]G. Onida, L. Reining, and A. Rubio, Electronic excitations: density-functional versus many-body Green’s function approaches, Rev. Mod. Phys. 74, 601 (2002).DOI
- [2]P. Hohenberg and W. Kohn, Inhomogeneous electron gas, Phys. Rev. 136, B864 (1964).
- [3]W. Kohn and L. J. Sham, Self-consistent equations including exchange and correlation effects, Phys. Rev. 140, A1133 (1965).DOI
- [4]J. P. Perdew, W. Yang, K. Burke, Z. Yang, E. K. U. Gross, M. Scheffler, G. E. Scuseria, T. M. Henderson, I. Y. Zhang, A. Ruzsinszky, H. Peng, J. Sun, E. Trushin, and A. Görling, Understanding band gaps of solids in generalized Kohn–Sham theory, Proc. Natl. Acad. Sci. U.S.A. 114, 2801 (2017).DOI
- [5]A. Seidl, A. Görling, P. Vogl, J. A. Majewski, and M. Levy, Generalized Kohn-Sham schemes and the band-gap problem, Phys. Rev. B 53, 3764 (1996).pubmed.ncbi.nlm.nih.gov/9983927
- [6]L. Hedin, New method for calculating the one-particle Green’s function with application to the electron-gas problem, Physical Review 139, A796 (1965).DOI
- [7]R. M. Martin, L. Reining, and D. M. Ceperley, Interacting electrons (Cambridge University Press, 2016).
- [8]F. Aryasetiawan and O. Gunnarsson, The GW method, Reports on progress in Physics 61, 237 (1998).DOI
- [9]L. Reining, The GW approximation: content, successes and limitations, Wiley Interdisciplinary Reviews: Computational Molecular Science 8, e1344 (2018).DOI
- [10]D. Golze, M. Dvorak, and P. Rinke, The GW compendium: A practical guide to theoretical photoemission spectroscopy, Frontiers in chemistry 7, 377 (2019).DOI
- [11]M. S. Hybertsen and S. G. Louie, First-principles theory of quasiparticles: calculation of band gaps in semiconductors and insulators, Physical review letters 55, 1418 (1985).pubmed.ncbi.nlm.nih.gov/10031814
- [12]M. Shishkin and G. Kresse, Implementation and performance of the frequency-dependent GW method within the PAW framework, Physical Review B: Condensed Matter and Materials Physics 74, 035101 (2006).DOI
- [13]H. Jiang and P. Blaha, GW with linearized augmented plane waves extended by high-energy local orbitals, Physical Review B 93, 115203 (2016).DOI
- [14]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).
- [15]J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, Neural message passing for quantum chemistry, in International conference on machine learning (PMLR, 2017) pp. 1263–1272.arxiv.org/abs/1704.01212
- [16]T. Cohen and M. Welling, Group equivariant convolutional networks, in International conference on machine learning (PMLR, 2016) pp. 2990–2999.arxiv.org/abs/1602.07576
- [17]M. Weiler, M. Geiger, M. Welling, W. Boomsma, and T. S. Cohen, 3d steerable cnns: Learning rotationally equivariant features in volumetric data, Advances in Neural information processing systems 31 (2018).arxiv.org/abs/1807.02547
- [18]N. Thomas, T. Smidt, S. Kearnes, L. Yang, L. Li, K. Kohlhoff, and P. Riley, Tensor field networks: Rotation- and translation-equivariant neural networks for 3d point clouds, arXiv preprint arXiv:1802.08219 (2018).arxiv.org/abs/1802.08219
- [19]J. Brandstetter, R. Hesselink, E. van der Pol, E. J. Bekkers, and M. Welling, Geometric and physical quantities improve E(3)-equivariant message passing, arXiv preprint arXiv:2110.02905 (2021).arxiv.org/abs/2110.02905
- [20]M. Geiger and T. Smidt, e3nn: Euclidean neural networks, arXiv preprint arXiv:2207.09453 (2022).arxiv.org/abs/2207.09453
- [21]S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials, Nature communications 13, 2453 (2022).DOI
- [22]P. W. Battaglia, J. B. Hamrick, V. Bapst, A. Sanchez Gonzalez, V. Zambaldi, M. Malinowski, A. Tacchetti, D. Raposo, A. Santoro, R. Faulkner, et al., Relational inductive biases, deep learning, and graph networks, arXiv preprint arXiv:1806.01261 (2018).arxiv.org/abs/1806.01261
- [23]O. Unke, M. Bogojeski, M. Gastegger, M. Geiger, T. Smidt, and K.-R. Müller, SE (3)-equivariant prediction of molecular wavefunctions and electronic densities, Advances in Neural Information Processing Systems 34, 14434 (2021).arxiv.org/abs/2106.02347
- [24]E. Prodan and W. Kohn, Nearsightedness of electronic matter, Proceedings of the National Academy of Sciences 102, 11635 (2005).pubmed.ncbi.nlm.nih.gov/16087868
- [25]Z. Qiao, A. S. Christensen, W. Welborn, F. R. Manby, A. Anandkumar, and T. F. Miller III, Informing geometric deep learning with electronic interactions to accelerate quantum chemistry, Proceedings of the National Academy of Sciences 119, e2205221119 (2022).DOI
- [26]M. Su, J.-H. Yang, H.-J. Xiang, and X.-G. Gong, Efficient determination of the Hamiltonian and electronic properties using graph neural network with complete local coordinates, Machine Learning: Science and Technology 4, 035010 (2023).arxiv.org/abs/2205.05475
- [27]X. Gong, H. Li, N. Zou, R. Xu, W. Duan, and Y. Xu, General framework for E(3)-equivariant neural network representation of density functional theory Hamiltonian, Nature Communications 14, 2848 (2023).DOI
- [28]H. Yu, Z. Xu, X. Qian, X. Qian, and S. Ji, Efficient and equivariant graph networks for predicting quantum Hamiltonian, in International Conference on Machine Learning (PMLR, 2023) pp. 40412–40424.arxiv.org/abs/2306.04922
- [29]S. Yin, X. Pan, X. Zhu, T. Gao, H. Zhang, F. Wu, and L. He, Towards harmonization of SO(3)-equivariance and expressiveness: A hybrid deep learning framework for electronic-structure Hamiltonian prediction, Machine Learning: Science and Technology 5, 045038 (2024).arxiv.org/abs/2401.00744
- [30]S. Yin, X. Pan, F. Wang, and L. He, TraceGrad: A framework learning expressive $SO(3)$-equivariant non-linear representations for electronic-structure Hamiltonian prediction, in *Proceedings of the 42nd International Conference on Machine Learning (ICML)*, Proceedings of Machine Learning Research, Vol. 267 (PMLR, 2025) pp. 72364–72392.
- [31]S. Yin, Z. Dai, X. Pan, and L. He, Advancing universal deep learning for electronic-structure Hamiltonian prediction of materials, in *The Fourteenth International Conference on Learning Representations (ICLR)* (2026).arxiv.org/abs/2509.19877
- [32]H. Li, Z. Wang, N. Zou, M. Ye, R. Xu, X. Gong, W. Duan, and Y. Xu, Deep-learning density functional theory Hamiltonian for efficient ab initio electronic-structure calculation, Nature Computational Science 2, 367 (2022).DOI
- [33]Y. Zhong, H. Yu, M. Su, X. Gong, and H. Xiang, Transferable equivariant graph neural networks for the Hamiltonians of molecules and solids, npj Computational Materials 9, 182 (2023).DOI
- [34]Z. Tang, H. Li, P. Lin, X. Gong, G. Jin, L. He, H. Jiang, X. Ren, W. Duan, and Y. Xu, A deep equivariant neural network approach for efficient hybrid density functional calculations, Nature Communications 15, 8815 (2024).DOI
- [35]Y. Zhong, H. Yu, J. Yang, X. Gong, X. Xiang, and X. Gong, Universal machine learning Kohn–Sham Hamiltonian for materials, Chin. Phys. Lett. 41, 077103 (2024).DOI
- [36]Y. Zhong, R. Wang, X. Gong, and H. Xiang, A universal spin–orbit-coupled Hamiltonian model for accelerated quantum material discovery, Nat. Mach. Intell. 8, 403 (2026).arxiv.org/abs/2504.19586
- [37]L.-F. m. c. Arsenault, A. Lopez-Bezanilla, O. A. von Lilienfeld, and A. J. Millis, Machine learning for many-body physics: The case of the Anderson impurity model, Phys. Rev. B 90, 155136 (2014).DOI
- [38]E. J. Sturm, M. R. Carbone, D. Lu, A. Weichselbaum, and R. M. Konik, Predicting impurity spectral functions using machine learning, Phys. Rev. B 103, 245118 (2021).DOI
- [39]E. Sheridan, C. Rhodes, F. Jamet, I. Rungger, and C. Weber, Data-driven dynamical mean-field theory: An error-correction approach to solve the quantum many-body problem using machine learning, Phys. Rev. B 104, 205120 (2021).DOI
- [40]C. Venturella, C. Hillenbrand, J. Li, and T. Zhu, Machine learning many-body Green’s functions for molecular excitation spectra, Journal of Chemical Theory and Computation 20, 143 (2023).arxiv.org/abs/2310.09911
- [41]X. Dong, E. Gull, and L. Wang, Equivariant neural network for Green’s functions of molecules and materials, Physical Review B 109, 075112 (2024).DOI
- [42]C. Venturella, J. Li, C. Hillenbrand, X. Leyva Peralta, J. Liu, and T. Zhu, Unified deep learning framework for many-body quantum chemistry via Green’s functions, Nat. Comput. Sci. 5, 502 (2025).DOI
- [43]X. Ren, P. Rinke, V. Blum, J. Wieferink, A. Tkatchenko, A. Sanfilippo, K. Reuter, and M. Scheffler, Resolution-of-identity approach to Hartree–Fock, hybrid density functionals, RPA, MP2 and GW with numeric atom-centered orbital basis functions, New Journal of Physics 14, 053020 (2012).DOI
- [44]X. Ren, F. Merz, H. Jiang, Y. Yao, M. Rampp, H. Lederer, V. Blum, and M. Scheffler, All-electron periodic $G_0W_0$ implementation with numerical atomic orbital basis functions: Algorithm and benchmarks, Physical Review Materials 5, 013807 (2021).DOI
- [45]M.-Y. Zhang, P. Lin, R. Shi, and X. Ren, Low-scaling GW calculations of quasi-particle energies for extended systems within the numerical atomic orbital framework, J. Chem. Theory Comput. 22, 5770 (2026).DOI
- [46]H. Gong, M.-Y. Zhang, P. Lin, B. Jia, Z. Guan, L. He, and X. Ren, $G^0W^0$ implementation based on the pseudopotential and numerical-atomic-orbital basis-set framework: Algorithms and benchmarks, Phys. Rev. B 114, 175139 (2026).arxiv.org/abs/2605.11512
- [47]B. Jia, M.-Y. Zhang, Z. Guan, H. Gong, and X. Ren, All-electron quasiparticle self-consistent GW for molecules and periodic systems within the numerical atomic orbital framework, J. Chem. Phys. 164, 194115 (2026).DOI
- [48]R. Ramakrishnan, P. O. Dral, M. Rupp, and O. A. von Lilienfeld, Quantum chemistry structures and properties of 134 kilo molecules, Scientific Data 1, 140022 (2014).
- [49]A. C. Ihrig, J. Wieferink, I. Y. Zhang, M. Ropo, X. Ren, P. Rinke, M. Scheffler, and V. Blum, Accurate localized resolution of identity approach for linear-scaling hybrid density functionals and for many-body perturbation theory, New J. Phys. 17, 093020 (2015).DOI
- [50]P. Lin, X. Ren, and L. He, Accuracy of Localized Resolution of the Identity in Periodic Hybrid Functional Calculations with Numerical Atomic Orbitals, J. Phys. Chem. Lett. 11, 3082 (2020).DOI
- [51]K. Schütt, O. Unke, and M. Gastegger, Equivariant message passing for the prediction of tensorial properties and molecular spectra, in *International conference on machine learning* (PMLR, 2021) pp. 9377–9388.arxiv.org/abs/2102.03150
- [52]H. J. Vidberg and J. W. Serene, Solving the Eliashberg equations by means of $N$-point Padé approximants, J. Low Temp. Phys. 29, 179 (1977).DOI
- [53]M. J. van Setten, F. Caruso, S. Sharifzadeh, X. Ren, M. Scheffler, F. Liu, J. Lischner, L. Lin, J. R. Deslippe, S. G. Louie, C. Yang, F. Weigend, J. B. Neaton, F. Evers, and P. Rinke, GW100: Benchmarking $G_0W_0$ for molecular systems, J. Chem. Theory Comput. 11, 5665 (2015).
- [54]M. Jarrell and J. E. Gubernatis, Bayesian inference and the analytic continuation of imaginary-time quantum Monte Carlo data, Phys. Rep. 269, 133 (1996).
- [55]K. S. D. Beach, R. J. Gooding, and F. Marsiglio, Reliable Padé analytical continuation method based on a high-accuracy symbolic computation algorithm, Phys. Rev. B 61, 5147 (2000).
- [56]J. Schött, I. L. M. Locht, E. Lundin, O. Granäs, O. Eriksson, and I. Di Marco, Analytic continuation by averaging Padé approximants, Phys. Rev. B 93, 075104 (2016).DOI
- [57]J. Fei, C.-N. Yeh, and E. Gull, Nevanlinna analytical continuation, Phys. Rev. Lett. 126, 056402 (2021).DOI
- [58]Z. Huang, E. Gull, and L. Lin, Robust analytic continuation of Green’s functions via projection, pole estimation, and semidefinite relaxation, Phys. Rev. B 107, 075151 (2023).DOI
- [59]Y. Nakatsukasa, O. Sète, and L. N. Trefethen, The AAA algorithm for rational approximation, SIAM J. Sci. Comput. 40, A1494 (2018). E. Fabiano, M. R. Farrow, F. Fiebig, J. Filser, L. Foppa, L. Gallandi, A. Garcia, R. Gehrke, S. Ghan, L. Ghir-inghelli, M. Glass, S. Goedecker, D. Golze, M. Gram-zow, J. A. Green, A. Grisafi, A. Grünew, J. J. Günzl, S. Gutzeit, S. J. Hall, F. Hanke, V. Havu, X. He, J. Hekele, O. Hellman, U. Herath, J. Hermann, D. Her-nangómez-Pérez, O. T. Hofmann, J. Hoja, S. Hollweger, L. Hörmann, B. Hourahine, W. B. How, W. P. Huhn, M. Hülsberg, T. Jacob, S. Panahian Jand, H. Jiang, E. Johnson, W. Jürgens, J. M. Kahk, Y. Kanai, K. Kang, P. Karpov, E. Keller, R. Kempt, D. Khan, M. Kick, B. P. Klein, J. Kloppenburg, A. Knoll, F. Knoop, F. Knuth, S. S. Köcher, J. Kockläuner, S. Kokott, T. Körzdörfer, H.-H. Kowalski, P. Kratzer, P. Kus, R. Laasner, B. Lang, B. Lange, M. F. Langer, A. H. Larsen, H. Lederer, S. Lehtola, M.-O. Lenz-Himmer, M. Leucke, S. Levchenko, A. Lewis, O. A. von Lilienfeld, K. Lion, W. Lipsunpen, J. Lischner, Y. Litman, C. Liu, Q.-L. Liu, S. Liu, A. J. Logsdail, M. Lorke, Z. Lou, I. Mandzhieva, A. Marek, J. T. Margraf, R. J. Maurer, T. Melson, F. Merz, J. Meyer, G. S. Miche-litsch, T. Mizoguchi, E. Moerman, D. Morgan, J. Mor-genstein, J. Moussa, A. S. Nair, L. Nemec, H. Oberhofer, A. Otero-de-la-Roza, R. L. Panadés-Barrueta, T. Pat-lolla, M. Pogodaeva, A. Pöppl, A. J. A. Price, T. A. R. Purcell, J. Quan, N. Raimbault, M. Rampp, K. Rasim, R. Redmer, X. Ren, K. Reuter, N. A. Richter, S. Ringe, P. Rinke, S. P. Rittmeyer, H. I. Rivera-Arrieta, M. Ropo, M. Rossi, V. Ruiz, N. Rybin, A. Sanfilippo, M. Schef-fler, C. Scheurer, C. Schober, F. Schubert, T. Shen, C. Shepard, H. Shang, K. Shibata, A. Sobolev, R. Song, A. Soon, D. T. Speckhard, P. V. Stishenko, E. Stocco, M. N. Tahir, I. Takahara, J. Tang, Z. Tang, T. Theis, F. Theiss, A. Tkatсhenko, M. Todorović, G. Trenins, O. T. Unke, Á. Vázquez-Mayagoitia, O. van Vuren, D. Waldschmidt, H. Wang, Y. Wang, J. Wieferink, J. Wilhelm, S. Woodley, J. Xu, Y. Xu, Y. Yao, Y. Yao, M. Yoon, V. W.-z. Yu, Z. Yuan, M. Zacharias, I. Y. Zhang, M.-Y. Zhang, W. Zhang, X. Zhang, R. Zhao, S. Zhao, R. Zhou, Y. Zhou, and T. Zhu, Roadmap on advancements of the FHI-aims software package, Electron. Struct. 10.1088/2516-1075/ae8067 (2026).
- [61]J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized gradient approximation made simple, Phys. Rev. Lett. 77, 3865 (1996).pubmed.ncbi.nlm.nih.gov/10062328
- [62]P. Li, X. Liu, C. Liu, M. Chen, P. Lin, X. Ren, L. Lin, C. Yang, and L. He, Large-scale ab initio simulations based on systematically improvable atomic basis, Computational Materials Science 112, 503 (2016).arxiv.org/abs/1503.00097
- [63]W. Zhou, D. Zheng, Q. Liu, et al., ABACUS: An electronic structure analysis package for the AI era, The Journal of Chemical Physics 163, 192501 (2025).arxiv.org/abs/2501.08697
- [64]I. Loshchilov and F. Hutter, Decoupled weight decay regularization, in The Seventh International Conference on Learning Representations (ICLR) (2019).
- [65]S. Passaro and C. L. Zitnick, Reducing SO(3) convolutions to SO(2) for efficient equivariant GNNs, in Proceedings of the 40th International Conference on Machine Learning (ICML), Proceedings of Machine Learning Research, Vol. 202 (PMLR, 2023) pp. 27420–27438.arxiv.org/abs/2302.03655
- [66]Z. Liang, Y. Wang, C. Ding, J. Wang, H.-T. Wang, D. Xing, and J. Sun, Hot-Man: an accurate and efficient E(3)-equivariant machine-learning electronic structures calculation framework, Chin. Phys. Lett. 43, 020704 (2026).







