Abstract
Last year, at least 30,000 scientific papers used the Kohn–Sham scheme of density functional theory to solve electronic structure problems in a wide variety of scientific fields. Machine learning holds the promise of learning the energy functional via examples, bypassing the need to solve the Kohn–Sham equations. This should yield substantial savings in computer time, allowing larger systems and/or longer timescales to be tackled, but attempts to machinelearn this functional have been limited by the need to find its derivative. The present work overcomes this difficulty by directly learning the densitypotential and energydensity maps for test systems and various molecules. We perform the first molecular dynamics simulation with a machinelearned density functional on malonaldehyde and are able to capture the intramolecular proton transfer process. Learning density models now allows the construction of accurate density functionals for realistic molecular systems.
Introduction
Kohn–Sham density functional theory^{1} (KSDFT) is now enormously popular as an electronic structure method in a wide variety of fields^{2}. Useful accuracy is achieved with standard exchange–correlation (XC) approximations, such as generalized gradient approximations^{3} and hybrids^{4}. Such calculations are playing a key role in the materials genome initiative^{5}, at least for weakly correlated materials^{6}.
There has also been a recent spike of interest in applying machine learning (ML) methods in the physical sciences^{7,8,9,10,11}. The majority of these applications involve predicting properties of molecules or materials from large databases of KSDFT calculations^{12,13,14,15}. A few applications involve finding potential energy surfaces within molecular dynamics (MD) simulations^{16,17,18,19}. Fewer still have focussed on finding the functionals of DFT as a method of performing KS electronic structure calculations without solving the KS equations^{20,21,22,23,24}. If such attempts could be made practical, the possible speedup in repeated DFT calculations of similar species, such as occur in ab initio MD simulations, is enormous.
A key difficulty has been the need to extract the functional derivative of the noninteracting kinetic energy. The noninteracting kinetic energy functional T _{s}[n] of the electron density n is used in two distinct ways in a KS calculation^{1}, as illustrated in Fig. 1: (i) its functional derivative is used in the Euler equation which is solved in the selfconsistent cycle and (ii) when selfconsistency is reached, the groundstate energy of the system is calculated by E[n], an orbitalfree (OF) mapping. The solution of the KS equations performs both tasks exactly. Early results on simple model systems showed that ML could provide highly accurate values for T _{s}[n] with only modest amounts of training^{20}, but that the corresponding functional derivatives are too noisy to yield sufficiently accurate results to (i). Subsequent schemes overcome this difficulty in various ways, but they typically lose a factor of 10 or more in accuracy^{22}, and their computational cost can increase dramatically with system complexity.
Here we present an alternative ML approach, in which we replaced the Euler equation by directly learning the Hohenberg–Kohn (HK) map v(r) → n(r) (red line in Fig. 1a) from the onebody potential of the system of interest to the interacting groundstate density, i.e., we establish an MLHK map. We show that this map can be learned at a much more modest cost than either previous ML approaches to find the functional and its derivative (MLOF) or direct attempts to model the energy as a functional of v(r) (MLKS). Furthermore, we show that it can immediately be applied to molecular calculations by calculating the energies of small molecules over a range of conformers. Moreover, as we have already implemented this approach with a standard quantum chemical code (Quantum Espresso^{25}) using a standard DFT approximation (Perdew–Burke–Ernzerhof (PBE)), this can now be tried on much larger scales.
The MLHK map reflects the underlying computational approach used to generate a particular electron density, but is not restricted to any given electronic structure method. Many molecular properties, including but not limited to the energy, are dependent on the electron density, making the MLHK map more versatile than a direct MLKS mapping. We also establish that densities can be learned with sufficient accuracy to distinguish between different DFT functionals, thus providing a route to future functional development via generation of precise densities for a range of molecules and conformations.
Results
We will first outline theoretical results, most prominently the MLHK map, and then illustrate the approach with simulations of onedimensional (1D) systems and threedimensional (3D) molecules.
MLHK map
Previous results show that for an MLOF approach, the accuracy of MLKS kinetic energy models \(T_{\rm{s}}^{{\rm{ML}}}\left[ n \right]\) improve rapidly with the amount of data. However, minimizing the total energy via gradient descent requires the calculation of the gradient of the kinetic energy model \(T_{\rm{s}}^{{\rm{ML}}}\) (Fig. 1) and calculating this gradient is challenging. Due to the datadriven nature of, e.g., kernel models, the machinelearned kinetic energy functional has no information in directions that point outside the data manifold^{26}. This heavily influences the gradient to an extent that it becomes unusable without further processing^{20}. There have been several suggestions to remedy this problem, but all of them share a significant loss in accuracy compared to T _{s}[n]^{21, 22, 27}.
Here, we propose an interesting alternative to gradients and the MLOF approach. Recently, it has been shown that the HK map for the density as a functional of the potential can be approximated extremely accurately using semiclassical expressions^{28}. Such expressions do not require the solution of any differential equation, and become more accurate as the number of particles increases. Errors can be negligible even for just two distinct occupied orbitals.
Inspired by this success, we suggest how one might circumvent the kinetic energy gradient and directly train a multivariate ML model. We name this the MLHK map:
Here each density grid point is associated with a group of model weights β. Training requires solving an optimization problem for each density grid point. Although this is possible in 1D, it rapidly becomes intractable in 3D, as the number of grid points grows cubically.
The use of a basis representation for the densities, as in
renders the problem tractable even for 3D. A ML model that predicts the basis function coefficients u ^{(l)}[v] instead of the grid points is then formulated.
Predicting the basis function coefficients not only makes the ML model efficient and allows the extension of the approach to 3D but also permits regularization, e.g., to smooth the predicted densities by removing the highfrequency basis functions, or to further regularize the ML model complexity for specific basis functions.
For orthogonal basis functions, the ML model reduces to several independent regression models and admits an analytical solution analogous to kernel ridge regression (KRR, see Supplementary Eq. (5)):
Here, for each basis function coefficient, λ ^{(l)} are regularization parameters and \({{\bf{K}}_{{{\boldsymbol{\sigma }}^{(l)}}}}\) is a Gaussian kernel with kernel width σ ^{(l)}. The parameters λ ^{(l)} and σ ^{(l)} can be chosen individually for each basis function via independent crossvalidation (see refs. ^{12, 29}). This MLHK model avoids prior gradient descent procedures and, with it, the necessity to “denoise” the gradients. Due to the independence of Eq. (3) for each l, the solution scales favourably.
Functional and densitydriven error
How can the performance of the MLHK map be measured? It has recently been shown how to separate out the effect of the error in the functional F and the error in the density n(r) on the resulting error in the total energy of any approximate, selfconsistent DFT calculation^{30}. Let \(\tilde F\) be an approximation of the many body functional F, and \(\tilde n({\bf{r}})\) the approximate groundstate density when \(\tilde F\) is used in the Euler equation. Defining \(\tilde E[n] = \tilde F[n] + {\int} {{d^3}r\,{\kern 1pt} n({\bf{r}})\,{\kern 1pt} v({\bf{r}})}\) yields
where \(\Delta {E_{\rm{F}}} = \tilde F[n]  F[n]\) is the functionaldriven error, while \(\Delta {E_{\rm{D}}} = \tilde E\left[ {\tilde n} \right]  \tilde E[n]\) is the densitydriven error. In most DFT calculations, ΔE is dominated by ΔE _{F}. The standard DFT approximations can, in some specific cases, produce abnormally large density errors that dominate the total error. In such situations, using a more accurate density can greatly improve the result^{30,31,32}. We will use these definitions to measure the accuracy of the MLHK map.
1D potentials
The following results demonstrate how much more accurate ML is when applied directly to the HK map. The box problem originally introduced in Snyder et al.^{20} is used to illustrate the principle. Random potentials consisting of three Gaussian dips were generated inside a hardwall box of length 1 (atomic units), and the Schrödinger equation for one electron was solved extremely precisely. Up to 200 cases were used to train an ML model \(T_{\rm{s}}^{{\rm{ML}}}[n]\) for the noninteracting kinetic energy functional T _{s}[n] via KRR (for details, see Supplementary Methods).
To measure the accuracy of an approximate HK map, the analysis of the previous section is applied to the KS DFT problem. Here F is just T _{s}, the noninteracting kinetic energy, and
i.e., the error made in an approximate functional on the exact density. Table 1 on the left gives the errors made by MLOF for the total energy, and its different components, when the density is found from the functional derivative. This method works by following a gradient descent of the total energy functional based on the gradient of the ML model \(T_{\rm{s}}^{{\rm{ML}}}\),
where \(\epsilon\) is a small number and P(n ^{(j)}) is a localized PCA projection to denoise the gradient. Here, and for all further 1D results, we use
The densitydriven contribution to the error, ΔE _{D}, which we calculate exactly here using the von Weizsäcker kinetic energy^{33}, is always comparable to, or greater than, the functionaldriven error, ΔE _{F}, due to the poor quality of the ML functional derivative^{20}. The calculation is abnormal and can be greatly improved by using a more accurate density from a finer grid. As the number of training points M grows, the error becomes completely dominated by the error in the density. This shows that the largest source of error lies in the use the ML approximation of T _{s} to find the density by solving the Euler equation.
The next set of columns analyzes the MLHK approach, using a grid basis. The leftmost of these columns shows the energy error we obtain by utilizing the MLHK map:
Note that both ML models, \(T_{\rm{s}}^{{\rm{ML}}}\) and n ^{ML}, have been trained using the same set of M training points.
The MLHK approach is always more accurate than MLOF, and its relative performance improves as M increases. The next column reports the densitydriven error, ΔE _{D}, which is an orderofmagnitude smaller than that of MLOF. Lastly, we list an estimate to the densitydriven error
which uses the ML model \(T_{\rm{s}}^{{\rm{ML}}}\) for the kinetic energy functional in 1D. This proxy is generally a considerable overestimate (a factor of 3 too large), so that the true ΔE _{D} is always significantly smaller. We use it in subsequent calculations (where we cannot calculate \(T_{\rm{s}}^{{\rm{ML}}}\)) to (over)estimate the energy error due to the MLHK map.
The last set of columns are densitydriven errors for other basis sets. Three variants of the MLHK map were tested. First, direct prediction of the grid coefficients: in this case, \({{u}}_i^{(l)} = {n_i}\left( {{x_l}} \right)\), l = 1, …, G. Five hundred grid points were used, as in Snyder, et al.^{20}. This variant is tested in 1D only; in 3D the high dimensionality is prohibitive. Second, a common Fourier basis is tested. The density can be transformed efficiently via the discrete Fourier transform, using 200 Fourier basis functions in total. In 3D, these basis functions correspond to plane waves. The backprojection \({{u}} \mapsto n\) to input space is simple, but although the basis functions are physically motivated, they are very general and not specifically tailored to density functions. The performance is almost identical to the grid, on average, although maximum errors are much less. For M = 20, the error that originates from the basis representation starts to dominate. This is a motivation for exploring, third, a Kernel PCA (KPCA) basis^{34}. KPCA^{35} is a popular generalization of PCA that yields basis functions that maximize variance in a higher dimensional feature space. The KPCA basis functions are datadriven, and computing them requires an eigendecomposition of the Kernel matrix. Good results are achieved with only 25 KPCA basis functions. The KPCA approach gives better results because it can take the nonlinear structure in the density space into account. However, it introduces the preimage problem: it is not trivial to project the densities from KPCA space back to their original (grid) space (Supplementary Note 1). It is thus not immediately applicable to 3D applications.
Molecules
We next apply the MLHK approach to predict electron densities and energies for a series of small molecules. We test the ML models on KSDFT results obtained using the PBE XC functional^{36} and atomic pseudopotentials with the projector augmentedwave (PAW) method^{37, 38} in the Quantum ESPRESSO code^{25}. As the MLOF approach applied in the previous section becomes prohibitively expensive in 3D due to the poor convergence of the gradient descent procedure (Supplementary Note 2), we compare the MLHK map to the MLKS approach. This approach models the energy directly as a functional of v(r), i.e., it trains a model
via KRR (for details, see Supplementary Methods).
We also apply the MLHK map with Fourier basis functions. Instead of a \(T_{\rm{s}}^{{\rm{ML}}}[n]\) model, we learn an E ^{ML}[n] model
which avoids implementing the potential energy and XC functionals.
Both approaches require the characterization of the Hamiltonian by its external potential. The external (Coulomb) potential diverges for 3D molecules and is, therefore, not a good feature to measure the distance in ML. Instead, we use an artificial Gaussian potential of the form
where R _{ α } are the positions, and Z _{ α } are the nuclear charges of the N ^{a} atoms. The Gaussian potential is used for the ML representation only. The width γ is a hyperparameter of the algorithm. The choice is arbitrary but can be crossvalidated. We find reasonable results with γ = 0.2 Å. The idea of using Gaussians to represent the external potential has been used previously^{39}. The Gaussian potential is discretized on a coarse grid with grid spacing Δ = 0.08. To use the discretized potential in the Gaussian kernel, we flatten it into a vector form and thus use a tensor Frobenius norm.
Our first molecular prototype is H_{2}, with the only degree of freedom, R, denoting the distance between the H atoms. A data set of 150 geometries is created by varying R between 0.5 and 1.5 Å (sampled uniformly). A randomly chosen subset of 50 geometries is designated as the test set and is unseen by the ML algorithms. These geometries are used to measure the outofsample error reported below.
The remaining 100 geometries make up the grand training set. To evaluate the performance of the MLKS map and the MLHK map, subsets of varying sizes M are chosen out of the grand training set to train the E ^{ML}[v] and n ^{ML}[v] models, respectively. Because the required training subsets are so small, careful selection of a subset that covers the complete range of R is necessary. This is accomplished by selecting the M training points out of the grand training set so that the R values are nearly equally spaced (see Supplementary Note 3 for details).
For practical applications, it is not necessary to run DFT calculations for the complete grand training set, only for the M selected training points. Strategies for sampling the conformer space and selecting the grand training set for molecules with more degrees of freedom are explained for H_{2}O and MD simulations later on.
The performance of the MLKS map and MLHK map is compared by evaluating E ^{ML}[v] that maps from the Gaussian potential to the total energy and the combination of n ^{ML}[v] that maps from Gaussian potential to the groundstate density in a 3D Fourier basis representation (l = 25) and E ^{ML}[n] that maps from density to total energy. The prediction errors are listed in Table 2.
The mean average error (MAE) of the energy evaluated using the MLHK map is significantly smaller than that of the MLKS map. This indicates that even for a 3D system, learning the potentialdensity relationship via the HK map is much easier than directly learning the potentialenergy relationship via the KS map.
Figure 1b shows the errors made by the MLKS and the MLHK maps. The error of the MLHK map is smoother than the MLKS error and is much smaller, even for the most problematic region when R is smaller than the equilibrium bond distance of R _{0} = 0.74 Å. The MAE that is introduced by the PBE approximation on the H_{2} data set is 2.3 kcal/mol (compared to exact CI calculations), i.e., well above the errors of the ML model and verifies that the error introduced by the MLHK map is negligible for a DFT calculation.
The next molecular example is H_{2}O, parametrized with three degrees of freedom: two bond lengths and a bond angle. In order to create a conformer data set, the optimized structure (R _{0} = 0.97 Å, θ _{0} = 104.2 using PBE) is taken as a starting point. A total of 350 geometries are then generated by changing each bond length by a uniformly sampled value between ±0.075 Å and varying the angle θ between ±8.59° (±0.15 rad) away from θ _{0} (see Supplementary Fig. 1 for a visualization of the sampled range). For this molecule, the outofsample test set again comprises a random subset of 50 geometries, with the remaining 300 geometries used as the grand training set. Because there are now three parameters, it is more difficult to select equidistant samples for the training subset of M data points. We therefore use a Kmeans approach to find M clusters and select the grand training set geometry closest to each cluster’s center for the training subset (see Supplementary Note 4 for details).
Models are trained as for H_{2}. The results are given in Table 2. As expected, the increase in degrees of freedom for H_{2}O compared to H_{2} requires a larger training set size M. However, even for the more complicated molecule, the MLHK map is consistently more precise than the MLKS map and provides an improved potential energy surface, as shown in Fig. 2. With an MAE of 1.2 kcal/mol for PBE energies relative to CCSD(T) calculations for this data set, we again show that ML does not introduce a new significant source of error.
The ML maps can also be used to find the minimum energy configuration. The total energy is minimized as the geometry varies with respect to both bond lengths and angles. For optimization, we use Powell’s method^{40}, which requires a starting point and an evaluation function to be minimized. For the H_{2}O case, the search is restricted to symmetric configurations with a random symmetric geometry used as the starting point. Results are reported in Table 2. The optimizations consistently converge to the correct minimum regardless of starting point, consistent with the maps being convex, i.e., the potential energy curves are sufficiently smooth to avoid introducing artificial local minima.
For larger molecules, generating random conformers that sample the full configurational space becomes difficult. Therefore, we next demonstrate that MD using a classical force field can also be used to create the grand training set. As an example, we use benzene (C_{6}H_{6}) with only small fluctuations in atomic positions out of the molecular plane (Supplementary Fig. 2). Appropriate conformers are generated via isothermal MD simulations at 300, 350, and 400 K using the General Amber Force Field (GAFF)^{41} in the PINY_MD package^{42}. Saving snapshots from the MD trajectories generates a large set of geometries that are sampled using the Kmeans approach to obtain 2000 representative points for the grand training set. Training of n ^{ML}[v] and E ^{ML}[n] is performed as above by running DFT calculations on M = 2000 points. We find that the ML error is reduced by creating the training set from trajectories at both the target temperature and a higher temperature to increase the representation of more distorted geometries. The final ML model is tested on 200 conformational snapshots taken from an independent MD trajectory at 300 K (Fig. 3a). The MAE of the MLHK map for this data set using training geometries from 300 and 350 K trajectories is only 0.37 kcal/mol for an energy range that spans more than 10 kcal/mol (Table 3).
For benzene, we further quantify the precision of the MLHK map in reproducing PBE densities. In Fig. 4, it is clear that the errors in the Fourier basis representation are larger than the errors introduced by the MLHK map by two orders of magnitude. Furthermore, the MLHK errors in density (as evaluated on a grid in the molecular plane of benzene) are also considerably smaller than the difference in density between density functionals (PBE vs. LDA^{43}). This result verifies that the MLHK map is specific to the density used to train the model and should be able to differentiate between densities generated with other electronic structure approaches.
Ethane (C_{2}H_{6}), with a small energy barrier for the relative rotation of the methyl groups, is also evaluated in the same way. Using geometries sampled with Kmeans from 300 and 350 K classical trajectories, the MLHK model reproduces the energy of conformers with a MAE of 0.23 kcal/mol for an independent MD trajectory at 300 K (Fig. 3b). This test set includes conformers from the sparselysampled eclipsed configuration (Supplementary Fig. 3). Using points from a 400 K trajectory improves the MLHK map due to the increased probability of higher energy rotamers in the training set (Table 3). The training set could also be constructed by including explicit rotational conformers, as is commonly done when fitting classical force field parameters^{41}. In either case, generating appropriate conformers for training via computationally cheap classical MD significantly decreases the cost of the MLHK approach.
As additional proof of the versatility of the MLHK map, we show that this approach is also able to interpolate energies for proton transfer in the enol form of malonaldehyde (C_{3}H_{4}O_{2}). This molecule is a wellknown example of intramolecular proton transfer, and our previous AIMD and ab initio path integral studies^{44} found classical and quantumfree energy barrier values of 3.5 and 1.6 kcal/mol, respectively, from gradientcorrected DFT. In this work, classical MD trajectories are run for each tautomer separately, with a fixed bonding scheme, then combined for Kmeans sampling to create the grand training set. The training set also includes an artificially constructed geometry that is the average of tautomer atomic positions. For the test set, we use snapshots from a computationally expensive Born–Oppenheimer ab initio MD trajectory at 300 K. Figure 5a shows that the MLHK map is able to predict DFT energies during a proton transfer event (MAE of 0.27 kcal/mol) despite being trained on classical geometries that did not include these intermediate points.
We show, finally, that the MLHK map can also be used to generate a stable MD trajectory for malonaldehyde at 300 K (Fig. 5b). In principle, analytic gradients could be obtained for each timestep, but for this first proofofconcept trajectory, a finitedifference approach was used to determine atomic forces. The MLHKgenerated trajectory samples the same molecular configurations as the ab inito MD simulation (see Fig. 6 and Supplementary Table 1) with a mean absolute energy error of 0.77 kcal/mol, but it typically underestimates the energy for outofplane molecular fluctuations at the extremes of the classical training set (maximum error of 5.7 kcal/mol, see Supplementary Fig. 4). Even with the underestimated energy values, however, the atomic forces are sufficiently large to return the molecule to the equilibrium configuration, thus resulting in a stable trajectory. The new set of coordinates could be further sampled to expand the training set in a selfconsistent manner. Using iterative MLHKgenerated MD trajectories would eliminate the need to run computationally expensive MD simulations with KS DFT and would provide an iterative approach to reduce the energy errors for conformations not included in the classical training set.
Discussion
For several decades, DFT has been a crossdisciplinary area between theoretical physics, chemistry, and materials sciences. The methods of each field crossfertilize advances in other fields. This has led to its enormous popularity and widespread success, despite its wellknown limitations in both accuracy and the systems and properties to which it can be applied.
The present work represents a key step forward toward adding an additional ingredient to this mix, namely the construction of functionals via ML. While previous work showed proofs of principle in 1D, this work is a demonstration in 3D, using real molecules and productionlevel codes. We also demonstrate that molecular conformers used in the training set can be generated by a range of methods, including informed scans and classical MD simulations. This opens the possibility that ML methods, which complement all existing approaches to functional approximation, could become a new and very different approach to this problem with the potential to reduce the computational cost of routine DFT calculations significantly.
Our method, directly learning the HK densitypotential map, overcomes a major bottleneck in previous methodologies that arises in 3D. Our approach avoids solving an intermediate more general problem (the gradient descent) to find the solution of the more specific problem (finding the groundstate density). This is called transductive inference by the ML community and is thought to be key to the success of statistical inference methods^{45}. Following a direct prediction approach with the MLHK map increases the accuracy consistently on both 1D examples and 3D molecules. We are also able to learn density models that outperform energy models trained on much more data. This quantitative observation allows us to conclude that learning density models is much easier than learning energy models. Such a finding should be no surprise to practitioners of the art of functional construction (see, e.g., ref. ^{28}), but the present work quantifies this observation using standard statistical methods. As the MLHK map accurately reflects the training densities, more exact methods could also be used to generate the training set densities for functional development.
We have also derived a scheme for using basis functions to render the approach computationally feasible, which allows for facile integration of the method into existing DFT codes. Another advantage is the possibility to take the innate structure of the densities into account, i.e., spatial correlations are preserved by using lowfrequency basis functions. Again, this fits with the intuition of experienced practitioners in this field, but here we have quantified this in terms of machinelearned functionals.
Direct prediction of energies (e.g., the MLKS map) always has the potential to lead to conceptually easier methods. But such methods must also abandon the insights and effects that have made DFT a practical and usefully accurate tool over the past half century. Many usefully accurate DFT approximations already exist, and the corrections to such approximations can be machinelearned in precisely the same way as the entire functional has been approximated here^{23}. If ML corrections require less data, the method becomes more powerful by taking advantage of existing successes. Furthermore, existing theorems, such as the viral theorem^{46}, might also be used to construct the kinetic energy functional directly from an MLHK map. In the case of orbitaldependent functionals, such as metaGGA’s or global hybrids, the method presented here must be extended to learn, e.g., the full density matrix instead of just the density.
We also note that, for all the 3D calculations shown here, we machinelearned E[n], the entire energy (not just the kinetic energy), which includes some densityfunctional approximation for XC. However, with a quantum chemical code, we could have trained on much more accurate quantum chemical densities and energies. Thus, the MLHK maps, in principle, allow the construction of (nearly) exact density functionals for molecular systems, with the potential to significantly reduce the computational cost of quantum chemistrybased MD simulations. All this points to useful directions in which to expand on the results shown here.
Methods
Kohn–Sham DFT
The KSDFT computational electronic structure method determines the properties of manybody systems using functionals of the electron density. The foundation is the HK theorem^{47}, which establishes a onetoone relationship between potential and density, i.e., at most one potential can give rise to a given groundstate density.
Kohn–Sham DFT avoids direct approximation of manybody effects by imagining a fictitious system of noninteracting electrons with the same density as the real one^{1} (see Supplementary Note 5 for details). Its accuracy is limited by the accuracy of existing approximations to the unknown XC energy, while its computational bottleneck is the solution of the Kohn–Sham equations that describe the noninteracting particles.
Here, 3D DFT calculations for ML models are performed with the Quantum ESPRESSO code^{25} using the PBE XC functional^{36} and PAWs^{37, 38} with Troullier–Martins pseudization for describing the ionic cores^{48}. All molecules are simulated in a cubic box (L = 20 bohr) with a wave function cutoff of 90 Ry (see Supplementary Note 3 for details). The 1D data set is taken from ref. ^{20}.
Kernel ridge regression
KRR^{49, 50} is a ML method for regression. It is a kernelized version of ridge regression that minimizes the least squares error and applies an \({\ell _2}\) (Tikhonov) regularization. Let \({{\bf{x}}_1}, \ldots ,{{\bf{x}}_m} \in {{\Bbb R}^d}\) be the training data points and let Y = \({\left( {{{\bf{y}}_1}, \ldots ,{{\bf{y}}_m}} \right)^{\bf{T}}}\) be their respective labels. KRR then performs the following optimization:
where k is the kernel function and λ is a regularization parameter. K is the kernel matrix with K _{ ij } = k(x _{ i }, x _{ j }). Eq. (13) has the following analytical solution:
Most popular is the Gaussian (radial basis function) kernel that leads to a smooth, nonlinear model function in input space corresponding to a linear function in an infinite dimensional feature space^{29}.
For the MLHK map (Supplementary Methods), the canonical error is given by the \({{\cal L}_2}\) distance between predicted and true densities
The ML model coefficients β ^{(l)} can be optimized independently for each basis coefficient l via
Crossvalidation
Note that all model parameters and hyperparameters are estimated on the training set; the hyperparameter choice makes use of standard crossvalidation procedures (see Hansen et al.^{12}). Once the model is fixed after training, it is applied unchanged out of sample.
Exact calculations
Relative energy errors of the ML models trained on KSDFT calculations are determined by comparing to accurate energies from the Molpro Quantum Chemistry Software^{51} using the full configuration interaction method for H_{2} and CCSD(T)^{52} for H_{2}O.
Molecular dynamics
For benzene, ethane, and malonaldehyde, GAFF parameters^{41} were assigned using the AmberTools package^{53}. Geometry optimizations were performed using MP2/631g(d) in Gaussian09^{54}. Atomic charge assignments are from RESP fits^{55} to HF/631g(d) calculations at optimized geometries and two additional rotational conformers for ethane.
For the three larger molecules, classical isothermal MD simulations were run using the PINY_MD package^{42} with massive Nosé–Hoover chain (NHC) thermostats^{56} for atomic degrees of freedom (length = 4, τ = 20 fs, Suzuki–Yoshida order = 7, multiple time step = 4) and a time step of 1 fs. The rRESPA multiple time step approach^{57} was employed to compute rapidly varying forces more frequently (torsions every 0.5 fs, bonds/bends every 0.1 fs). Systems were equilibrated for 100 ps before collecting snapshots every 100 fs from 1 ns trajectories. Snapshots were aligned to a reference molecule prior to DFT calculations for the ML model. For malonaldehyde, the ML training set geometries were selected from trajectories for both enol tautomers as the GAFF force field does not permit changes in chemical bond environments.
For malonaldehyde, an additional Born–Oppenheimer MD simulation using DFT was run using the QUICKSTEP package^{58} in CP2K v. 2.6.2^{59}. The PBE XC functional^{36} was used in the Gaussian and planewave scheme^{60} with DZVPMOLOPTGTH (mDZVP) basis sets^{61} paired with the appropriate dualspace GTH pseudopotentials^{62} optimized for the PBE functional^{63}. Wave functions were converged to 1E7 Hartree in the energy using the orbital transformation method^{64} on a multiple grid (n = 5) with a cutoff of 900 Ry for the system in a cubic box (L = 20 bohr). A temperature of 300 K was maintained using massive NHC thermostats^{56} (length = 4, τ = 10 fs, Suzuki–Yoshida order = 7, multiple time step = 4) and a time step of 0.5 fs.
In order to generate the MD trajectory with the MLHK model, we used the Atomistic Simulation Environment^{65} with a 0.5 fs timestep and a temperature of 300 K maintained via a Langevin thermostat. A thermostat friction value of 0.01 atomic units (0.413 fs^{−1}) was chosen to reproduce the fluctuations in C atoms observed for the DFTbased trajectory (Supplementary Note 3). In this proofofconcept work, atomic forces were calculated using central finite differences, with \(\epsilon\) = 0.001 Å chosen to conserve the total energy during the simulation. The last 1 ps of a 4 ps trajectory was used to evaluate the performance of the MLHK model.
Data availability
All data sets used in this work are available at http://quantummachine.org/datasets/.
References
 1.
Kohn, W. & Sham, L. J. Selfconsistent equations including exchange and correlation effects. Phys. Rev. 140, A1133–A1138 (1965).
 2.
PribramJones, A., Gross, D. A. & Burke, K. DFT: a theory full of holes? Annu. Rev. Phys. Chem. 66, 283–304 (2015).
 3.
Perdew, J. P. Densityfunctional approximation for the correlation energy of the inhomogeneous electron gas. Phys. Rev. B 33, 8822–8824 (1986).
 4.
Becke, A. D. Density functional thermochemistry. III. The role of exact exchange. J. Chem. Phys. 98, 5648–5652 (1993).
 5.
National Science And Technology. Materials Genome Initiative for Global Competitiveness (General Books, 2011).
 6.
Jain, A. et al. A highthroughput infrastructure for density functional theory calculations. Comput. Mater. Sci. 50, 2295–2310 (2011).
 7.
Pozun, Z. D. et al. Optimizing transition states via kernelbased machine learning. J. Chem. Phys. 136, 174101 (2012).
 8.
McGibbon, R. T. & Pande, V. S. Learning kinetic distance metrics for Markov state models of protein conformational dynamics. J. Chem. Theory. Comput. 9, 2900–2906 (2013).
 9.
Fletcher, T. L., Davie, S. J. & Popelier, P. L. Prediction of intramolecular polarization of aromatic amino acids using kriging machine learning. J. Chem. Theory. Comput. 10, 3708–3719 (2014).
 10.
Rupp, M., Tkatchenko, A., Müller, K.R. & von Lilienfeld, O. A. Fast and accurate modeling of molecular atomization energies with machine learning. Phys. Rev. Lett. 108, 058301 (2012).
 11.
Hautier, G., Fischer, C. C., Jain, A., Mueller, T. & Ceder, G. Finding nature’s missing ternary oxide compounds using machine learning and density functional theory. Chem. Mater. 22, 3762–3767 (2010).
 12.
Hansen, K. et al. Assessment and validation of machine learning methods for predicting molecular atomization energies. J. Chem. Theory Comput. 9, 3404–3419 (2013).
 13.
Schütt, K. T. et al. How to represent crystal structures for machine learning: towards fast prediction of electronic properties. Phys. Rev. B 89, 205118 (2014).
 14.
Hansen, K. et al. Machine learning predictions of molecular properties: accurate manybody potentials and nonlocality in chemical space. J. Phys. Chem. Lett. 6, 2326–2331. PMID: 26113956 (2015).
 15.
Schütt, K. T., Arbabzadah, F., Chmiela, S., Müller, K. R. & Tkatchenko, A. Quantumchemical insights from deep tensor neural networks. Nat. Commun. 8, 13890 (2017).
 16.
Behler, J. & Parrinello, M. Generalized neuralnetwork representation of highdimensional potentialenergy surfaces. Phys. Rev. Lett. 98, 146401 (2007).
 17.
Seko, A., Takahashi, A. & Tanaka, I. Sparse representation for a potential energy surface. Phys. Rev. B 90, 024101 (2014).
 18.
Li, Z., Kermode, J. R. & De Vita, A. Molecular dynamics with onthefly machine learning of quantummechanical forces. Phys. Rev. Lett. 114, 096405 (2015).
 19.
Chmiela, S. et al. Machine learning of accurate energyconserving molecular force fields. Sci. Adv. 3, e1603015 (2017).
 20.
Snyder, J. C., Rupp, M., Hansen, K., Müller, K.R. & Burke, K. Finding density functionals with machine learning. Phys. Rev. Lett. 108, 253002 (2012).
 21.
Snyder, J. C. et al. Orbitalfree bond breaking via machine learning. J. Chem. Phys. 139, 224104 (2013).
 22.
Li, L. et al. Understanding machinelearned density functionals. Int. J. Quantum Chem. 116, 819–833 (2016).
 23.
Li, L., Baker, T. E., White, S. R. & Burke, K. Pure density functional for strong correlation and the thermodynamic limit from machine learning. Phys. Rev. B 94, 245129 (2016).
 24.
Yao, K. & Parkhill, J Kinetic Energy of Hydrocarbons as a Function of Electron Density and Convolutional Neural Networks. J. Chem. Theory Comput. 12, 1139–1147 (2016).
 25.
Giannozzi, P. et al. QUANTUM ESPRESSO: a modular and opensource software project for quantum simulations of materials. J. Phys. Condens. Matter. 21, 395502 (2009).
 26.
Snyder, J. C., Mika, S., Burke, K. & Müller, K.R. in Empirical Inference (eds Schölkopf, B., Luo, Z. & Vovk, V.) 245–259 (Springer, 2013).
 27.
Snyder, J. C., Rupp, M., Müller, K.R. & Burke, K. Nonlinear gradient denoising: finding accurate extrema from inaccurate functional derivatives. Int. J. Quantum. Chem. 115, 1102–1114 (2015).
 28.
Ribeiro, R. F., Lee, D., Cangi, A., Elliott, P. & Burke, K. Corrections to ThomasFermi densities at turning points and beyond. Phys. Rev. Lett. 114, 050401 (2015).
 29.
Müller, K.R., Mika, S., Rätsch, G., Tsuda, K. & Schölkopf, B. An introduction to kernelbased learning algorithms. IEEE Trans. Neural. Netw. 12, 181–201 (2001).
 30.
Kim, M.C., Sim, E. & Burke, K. Understanding and reducing errors in density functional calculations. Phys. Rev. Lett. 111, 073003 (2013).
 31.
Kim, M.C., Sim, E. & Burke, K. Ions in solution: density corrected density functional theory (DCDFT). J. Chem. Phys. 140, 18A528 (2014).
 32.
Kim, M.C., Park, H., Son, S., Sim, E. & Burke, K. Improved DFT potential energy surfaces via improved densities. J. Phys. Chem. Lett. 6, 3802–3807 (2015).
 33.
Dreizler, R. M. & Gross, E. K. U. Density Functional Theory: an Approach to the Quantum ManyBody Problem (SpringerVerlag, 1990).
 34.
Schölkopf, B. et al. Input space versus feature space in kernelbased methods. IEEE Trans. Neural Netw. 10, 1000–1017 (1999).
 35.
Schölkopf, B., Smola, A. & Müller, K.R. Nonlinear component analysis as a kernel eigenvalue problem. Neural Comput. 10, 1299–1319 (1998).
 36.
Perdew, J. P., Burke, K. & Ernzerhof, M. Generalized gradient approximation made simple. Phys. Rev. Lett. 77, 3865–3868 (1996).
 37.
Kresse, G. & Joubert, D. From ultrasoft pseudopotentials to the projector augmentedwave method. Phys. Rev. B 59, 1758–1775 (1999).
 38.
Blöchl, P. E. Projector augmentedwave method. Phys. Rev. B 50, 17953–17979 (1994).
 39.
Bartók, A. P., Payne, M. C., Kondor, R. & Csányi, G. Gaussian approximation potentials: the accuracy of quantum mechanics, without the electrons. Phys. Rev. Lett. 104, 136403 (2010).
 40.
Powell, M. J. D. An efficient method for finding the minimum of a function of several variables without calculating derivatives. Comput. J. 7, 155–162 (1964).
 41.
Wang, J., Wolf, R. M., Caldwell, J. W., Kollman, P. A. & Case, D. A. Development and testing of a general Amber force field. J. Comput. Chem. 25, 1157–1174 (2004).
 42.
Tuckerman, M. E., Yarne, D., Samuelson, S. O., Hughes, A. L. & Martyna, G. J. Exploiting multiple levels of parallelism in molecular dynamics based calculations via modern techniques and software paradigms on distributed memory computers. Comput. Phys. Commun. 128, 333–376 (2000).
 43.
Perdew, J. P. & Zunger, A. Selfinteraction correction to densityfunctional approximations for manyelectron systems. Phys. Rev. B 23, 5048–5079 (1981).
 44.
Tuckerman, M. E. & Marx, D. Heavyatom skeleton quantization and proton tunneling in “intermediatebarrier” hydrogen bonds. Phys. Rev. Lett. 86, 4946–4949 (2001).
 45.
Vapnik, V. The Nature of Statistical Learning Theory (Information Science and Statistics) (Springer, 2000).
 46.
Cangi, A., Lee, D., Elliott, P., Burke, K. & Gross, E. K. U. Electronic structure via potential functional approximations. Phys. Rev. Lett. 106, 236404 (2011).
 47.
Hohenberg, P. & Kohn, W. Inhomogeneous electron gas. Phys. Rev 136, B864–B871 (1964).
 48.
Troullier, N. & Martins, J. L. Efficient pseudopotentials for planewave calculations. Phys. Rev. B 43, 1993–2006 (1991).
 49.
Hastie, T., Tibshirani, R. & Friedman, J. in The Elements of Statistical Learning—Data Mining, Inference, and Prediction 2nd edn (Springer, 2009).
 50.
Vu, K. et al. Understanding Kernel ridge regression: common behaviors from simple functions to density functionals. Int. J. Quantum. Chem. 115, 1115–1128 (2015).
 51.
Werner, H.J. et al. Molpro, Version 2015.1, a Package of Ab Initio Programs http://molpro.net (2015).
 52.
Adler, T. B., Knizia, G. & Werner, H.J. A simple and efficient CCSD(T)F12 approximation. J. Chem. Phys. 127, 221106–224100 (2007).
 53.
Wang, J., Wang, W., Kollman, P. A. & Case, D. A. Antechamber: an accessory software package for molecular mechanical calculations. J. Am. Chem. Soc. 222, U403 (2001).
 54.
Frisch, M. et al. Gaussian 09 (Revision D01) (Gaussian, Inc., 2013).
 55.
Bayly, C. I., Cieplak, P., Cornell, W. & Kollman, P. A. A wellbehaved electrostatic potential based method using charge restraints for deriving atomic charges: the RESP model. J. Phys. Chem. 97, 10269–10280 (1993).
 56.
Martyna, G. J., Klein, M. L. & Tuckerman, M. Nosé–hoover chains: the canonical ensemble via continuous dynamics. J. Chem. Phys. 97, 2635–2643 (1992).
 57.
Tuckerman, M., Berne, B. J. & Martyna, G. J. Reversible multiple time scale molecular dynamics. J. Chem. Phys. 97, 1990–2001 (1992).
 58.
VandeVondele, J. et al. Quickstep: fast and accurate density functional calculations using a mixed Gaussian and plane waves approach. Comput. Phys. Commun. 167, 103–128 (2005).
 59.
Hutter, J., Iannuzzi, M., Schiffmann, F. & VandeVondele, J. CP2K: atomistic simulations of condensed matter systems. Comput. Mol. Sci. 4, 15–25 (2014).
 60.
Lippert, G., Hutter, J. & Parrinello, M. A hybrid Gaussian and plane wave density functional scheme. Mol. Phys. 92, 477–488 (2010).
 61.
VandeVondele, J. & Hutter, J. Gaussian basis sets for accurate calculations on molecular systems in gas and condensed phases. J. Chem. Phys. 127, 114105 (2007).
 62.
Goedecker, S., Teter, M. & Hutter, J. Separable dualspace Gaussian pseudopotentials. Phys. Rev. B 54, 1703–1710 (1996).
 63.
Krack, M. Pseudopotentials for H to Kr optimized for gradientcorrected exchangecorrelation functionals. Theor. Chim. Acta 114, 145–152 (2005).
 64.
VandeVondele, J. & Hutter, J. An efficient orbital transformation method for electronic structure calculations. J. Chem. Phys. 118, 4365 (2003).
 65.
Bahn, S. R. & Jacobsen, K. W. An objectoriented scripting interface to a legacy electronic structure code. Comput. Sci. Eng. 4, 56–66 (2002).
Acknowledgements
We thank J.C. Snyder for early discussions and H. Glawe for helpful guidance regarding the 3D reference computations. We thank IPAM at UCLA for repeated hospitality. Work at UC Irvine supported by NSF CHE1464795. K.R.M. and F.B. thank the Einstein Foundation for generously funding the ETERNAL project. This work was supported by Institute for Information & Communications Technology Promotion (IITP) grant funded by the Korea government (no. 2017000451). Work at NYU supported by the US Army Research Office under contract/grant number W911NF1310387 (M.E.T. and L.V.). Ab initio trajectories were run using High Performance Computing resources at NYU. Other DFT simulations were run using High Performance Computing resources at MPI Halle.
Author information
Affiliations
Contributions
F.B. performed DFT calculations and ML experiments. L.V. performed classical and ab initio MD simulations. L.L. performed FCI and CCSD(T) calculations. K.B. and K.R.M. initiated the work, and K.B., K.R.M., and M.E.T. contributed to the theory and experiments. All authors contributed to the manuscript.
Corresponding authors
Ethics declarations
Competing interests
The authors declare no competing financial interests.
Additional information
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Electronic supplementary material
Rights and permissions
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
About this article
Cite this article
Brockherde, F., Vogt, L., Li, L. et al. Bypassing the KohnSham equations with machine learning. Nat Commun 8, 872 (2017). https://doi.org/10.1038/s41467017008393
Received:
Accepted:
Published:
Further reading

Learning to Approximate Density Functionals
Accounts of Chemical Research (2021)

Dataset Construction to Explore Chemical Space with 3D Geometry and Deep Learning
Journal of Chemical Information and Modeling (2021)

DeePKS: A Comprehensive DataDriven Approach toward Chemically Accurate Density Functional Theory
Journal of Chemical Theory and Computation (2021)

Learning DFT
The European Physical Journal Special Topics (2021)

Bypassing the computational bottleneck of quantumembedding theories for strong electron correlations with machine learning
Physical Review Research (2021)
Comments
By submitting a comment you agree to abide by our Terms and Community Guidelines. If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.