FCHL revisited: faster and more accurate quantum machine learning

Anders S. Christensen; Felix A. Faber; Lars A. Bratholm; O. Anatole von Lilienfeld

FCHL revisited: faster and more accurate quantum machine learning

T0 review · 0 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read FCHL19 descriptor yields millisecond-time, chemical-accuracy energy and force predictions.

desk verdict FCHL19 is a genuine, well-tested practical advance for kernel-based ML potentials; the universal parameters are truly out-of-sample on the benchmarks, and the weak spots are disclosed and minor. read the letter →

arxiv 1909.01946 v2 pith:3CCXDWBU submitted 2019-09-04 physics.chem-ph

Anders S. Christensen , Lars A. Bratholm , Felix A. Faber , O. Anatole von Lilienfeld This is my paper · ORCID

classification physics.chem-ph

keywords FCHL19quantummachinelearningmolecularrepresentationkernelridgeregressionforcepredictionatomizationenergytransferablehyperparametersdynamics

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a discretized, parameter-optimized atomic environment representation, FCHL19, makes kernel-based quantum machine learning both more accurate and much faster. The authors show that models built on FCHL19 reach chemical accuracy for molecular energies and forces, and that the representation's universal parameters transfer across datasets without re-optimization. This matters because it lowers the cost of training and using machine-learned potentials, bringing millisecond-per-atom force predictions within reach for routine molecular dynamics.

What carries the argument

The central object is the FCHL19 representation: a compact, atom-centered vector with a two-body part encoding radial distributions via log-normal basis functions and a three-body part encoding angles and mean distances via Fourier cosine and sine terms scaled by an Axilrod-Teller-Muto-type factor. The kernel that carries the argument is the element-screened Gaussian kernel $k(q_I, q_J^*) = \delta_{Z_I Z_J^*} \exp(-\|q_I - q_J^*\|_2^2 / 2\sigma^2)$, where the Kronecker delta restricts comparisons to atoms of the same element and makes kernel evaluations cheaper. The regressor that delivers millisecond force and energy predictions is the operator quantum machine learning approach, whose first-order derivative kernel keeps training fast enough to be practical. The whole construction rests on nine representation parameters fitted once by Monte Carlo optimization and then applied across all benchmark sets without re-optimization.

What would settle it

Re-run the learning curves after re-optimizing the nine representation parameters on each benchmark dataset with the same Monte Carlo procedure; if the re-fitted parameters improve out-of-sample MAE substantially over the fixed defaults on any of QM9, MD17, or Water40, the claimed universality of the parameters is shown not to transfer.

Watch

Extended reading notes

Core claim

FCHL19 is an atom-centered descriptor that represents each atomic environment as a compact vector of two-body radial terms (log-normal basis functions placed on a grid from the cutoff inward) and three-body terms (radial and Fourier angular terms scaled by an Axilrod-Teller-Muto-type factor). The paper's central claim is that this discretized representation, combined with an element-screened Gaussian kernel and a Monte Carlo-fitted set of universal parameters, reproduces reference quantum-chemical energies and forces at chemical accuracy. On QM9 the model reaches an out-of-sample atomization-energy MAE of 0.25 kcal/mol at the largest training split; on the MD17 force benchmarks it matches the best kernel baselines while cutting training data requirements roughly two- to four-fold relative to its predecessor; and on Water40 it reaches a binding-energy MAE below 0.1 kcal/mol per molecule after 3,200 training samples. With the operator quantum machine learning regressor, forces and energies are predicted in a few milliseconds per atom, making the model fast enough for routine molecular dynamics.

Load-bearing premise

The load-bearing premise is that the nine representation parameters, fitted once by Monte Carlo on 576 distorted geometries of small molecules built from carbon, nitrogen, oxygen, and hydrogen, transfer to larger and chemically different benchmark sets without re-optimization.

Editorial extensions

If this is right

Energy models on QM7b and QM9 reach chemical accuracy after training for minutes and hours respectively, with a QM9 atomization-energy MAE of 0.25 kcal/mol at the largest training split.
Training a QM9 kernel model takes roughly one day on a single compute node instead of three weeks, a 10- to 20-fold reduction in kernel evaluation time compared with the previous FCHL18 representation.
On the Water40 dataset, training on 3,200 samples gives a binding-energy MAE below 0.1 kcal/mol per molecule, showing transfer to condensed-phase-like systems dominated by non-bonded interactions.
For MD17 force learning, FCHL19 with Gaussian process regression requires about half the training data of FCHL19 with operator quantum machine learning, which in turn requires about half the data of FCHL18 with the same regressor.
FCHL19 with the operator quantum machine learning regressor predicts forces in 5.7 to 25.3 milliseconds per atom depending on molecule size, while remaining two to four times more sample-efficient than its predecessor for the same accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

If the transferability claim holds, the same Monte Carlo parameter-fitting recipe could be applied to other compact two- and three-body descriptors, turning descriptor optimization into a generic preprocessing step rather than a per-dataset chore.
The element-screening delta function in the kernel points toward an alchemical extension: replacing the Kronecker delta with a smooth element-similarity matrix would let the model share information across element types, something the paper mentions only as a future possibility.
The millisecond-per-atom force prediction times make on-the-fly machine-learned potentials practical for interactive and real-time molecular dynamics, a direction the paper itself gestures at in its outlook but does not develop.
A direct stress test of the universal parameters would be to apply the fixed defaults to molecules containing elements outside the CNO/H fitting set; if accuracy degrades sharply, element-specific or dataset-specific parameter fitting would be necessary.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Read this as a strong yes with a couple of caveats. The core claim—that a discretized revision of FCHL18 with optimized parameters gives state-of-the-art accuracy at 10–20x lower kernel cost—holds up on the evidence they present. They ship code, use nested CV with bootstrapped CIs, compare against a wide set of baselines, and report concrete timings. The FCHL19 representation is a genuine modification: log-normal radial functions, an element-screened Gaussian kernel, and a Monte Carlo parameter fit. It is incremental relative to their own FCHL18, but it is a useful step.

The main worry you passed me—that the nine representation parameters are fit on only 576 small CNO geometries and might not transfer—is less damning than it sounds. The parameters are fit once on a separate training set and then used without re-optimization on QM7b, QM9, Water40, and MD17. That is out-of-sample transfer, not circular tuning. The Water40 result is a particularly good test because it involves a different regime (non-bonded, condensed-phase-like), and the parameters do fine. So the paper's central accuracy claims do follow from its own fitting procedure. The stress-test concern is a legitimate be-careful, but not a fatal flaw.

The more honest soft spot is the paper's own disclosure that FCHL19 lags sGDML on toluene and naphthalene. The representation struggles with crowded, symmetric six-membered rings, and the speculative explanation about crowded angular spectra is plausible. That limitation should arguably be in the abstract, not just in the results. The benzene exclusion from MD17 is post hoc, but it is disclosed with a noise justification, and the noise claim is reasonable even if not rigorously proven. The 'state-of-the-art' phrasing is molecule-dependent—sGDML wins on some—but the paper does not hide that.

The timings are a clear plus: milliseconds per atom for force prediction with OQML, and the scaling analysis is transparent. This paper is for anyone building or using kernel-based ML potentials. It is a practical methods paper with reproducible deliverables. It deserves serious peer review: the experiments are carefully done, the code is public, and the claims are scoped.

Referee Report

0 major / 6 minor

Summary. The paper introduces FCHL19, a discretized revision of the FCHL18 atomic-environment representation. It replaces analytic integrals by log-normal radial bins and an angular Fourier expansion, adds element screening in a Gaussian kernel, and fits nine representation parameters once by Monte Carlo optimization on a small set of 576 distorted CNO geometries (Section IV.B, Table III). The resulting universal parameters are then used without per-dataset refitting in KRR/OQML/GPR models. Benchmarks cover atomization energies for QM7b and QM9, extrapolation from QM7b-T to GDB13-T, binding energies of (H2O)40 clusters, and forces/energies for seven MD17 molecules, together with kernel and prediction timings. The main claims are that FCHL19 reaches near-state-of-the-art accuracy at 10-20x lower kernel-evaluation cost than FCHL18 and can predict forces in milliseconds per atom.

Significance. Credit: the manuscript provides reproducible, open-source code in the QML toolkit, uses nested 5-fold cross-validation with bootstrapped confidence intervals on all learning curves, and compares against many baselines (FCHL18, SLATM/aSLATM, SOAP, Coulomb matrix, BoB, SchNet, sGDML/GDML). If the reported results hold, FCHL19 is a practically useful, compact representation: a single-node QM9 energy model in about 27 h versus 548-728 h for earlier representations, and sub-10 ms/atom force predictions. The transferability of the fixed representation parameters is partially demonstrated by out-of-sample results on chemistries (Cl, S, F) not present in the fitting set. The main limitations are the small size and limited chemical diversity of the parameter-fitting set and the molecule-dependent nature of the 'state-of-the-art' label on MD17; these are acknowledged in the text but deserve more explicit qualification.

minor comments (6)

[Section IV.B, Eq. (27), Table III] The Monte Carlo optimization of the universal representation parameters is described only qualitatively (greedy steps, no number of iterations, no restarts, no convergence criterion). Please add these details and, ideally, a sensitivity analysis around the Table III values; the current text's claim that the parameter variances are 'relatively small' is unsupported. This is not a circularity issue because the parameters are fixed before the benchmarks, but it is a reproducibility concern for the central transferability claim.
[Section III.B.1] The exclusion of benzene from the MD17 benchmark is a post hoc dataset choice; please support it with a quantitative estimate of the energy-label noise or explicitly discuss how the exclusion affects the benchmark comparison.
[Section III.B.1 and Conclusion] The label 'state-of-the-art accuracy' for MD17 should be qualified: Fig. 7 shows sGDML outperforming FCHL19 on toluene and naphthalene, so either report per-molecule best methods or use a phrase like 'competitive with state-of-the-art on most MD17 molecules.'
[Section III.C.2] The statement that the O(N^2) scaling of GPR and OQML 'translates to a 4 times increase in prediction speed' appears to confuse training and prediction complexity; prediction scales as O(N) with the number of training samples, so the factor of 4 applies to training time, not prediction speed. Please correct the wording.
[Section III.A] The assumption that differences in cross-validation methodology between this work and literature values 'give rise to negligible differences' is stated without support; please document the original protocols or add a caveat that the comparisons are approximate.
[Conclusion and Section IV.A.5] There are several typos ('discetization' in the Conclusion, 'ab inito' in the MD17 dataset description); please proofread the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: FCHL19 parameters are fit on a separate small-molecule dataset and applied out-of-sample to all benchmarks.

full rationale

The paper's central prediction claims are not constructed from their own targets. The nine FCHL19 representation parameters are optimized by Monte Carlo on a disjoint set of 576 distorted CNO geometries (Sec. IV B, Eq. 27), with a 384/192 train/test split, and then fixed (Table III) for all subsequent benchmarks; QM7b, QM9, Water40, and MD17 labels never enter this fit. Kernel width and regularization are chosen by nested five-fold cross-validation (Sec. IV C), which is standard model selection and does not make the outer-fold MAE a refit. The self-citations to FCHL18 (Ref. 1) and OQML (Ref. 22) supply the prior representation and a regressor, but neither is used as a uniqueness constraint or as the evidence for the accuracy claims; the benchmarks are quantitative and include cases where the proposed model does not win (e.g., sGDML on toluene/naphthalene, MOB on GDB13-T, FCHL18 on QM9). The paper even flags its own empirical choices ('we note that this is an empirical choice', Sec. II A 1; 'must be refitted to each individual dataset', Sec. IV B) and its out-of-sample intent ('the hyperparameters of FCHL19 ... have not been optimized on the QM9 dataset', Sec. III A 1). There is therefore no step in which a prediction reduces by construction to a fitted input or a self-citation chain.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The central claims rest on the representational ansatz (two- and three-body terms), the log-normal basis, the element-screened kernel, and transferability of universal hyperparameters. No new physical entities are introduced; the free parameters are representation hyperparameters fitted on an external training set and per-dataset kernel hyperparameters selected by cross-validation.

free parameters (9)

w (log-normal radial width) = 0.41 Å² (E); 0.32 Å² (E+F)
Fitted by Monte Carlo optimization on 576 distorted geometries; controls two-body radial basis width.
η3 (three-body radial width) = 0.97 Å⁻² (E); 2.7 Å⁻² (E+F)
Fitted by Monte Carlo optimization; controls three-body radial basis width.
N2 (two-body scaling exponent) = 2.4 (E); 1.8 (E+F)
Fitted by Monte Carlo optimization; sets decay of two-body weight with distance.
N3 (three-body scaling exponent) = 2.4 (E); 0.57 (E+F)
Fitted by Monte Carlo optimization; sets decay of three-body Axilrod-Teller-Muto term.
c3 (three-body weight) = 45.8 Å^N3 (E); 13.4 Å^N3 (E+F)
Fitted by Monte Carlo optimization; balances two-body vs three-body contributions.
nRs2 (number of two-body radial bins) = 22 (E); 24 (E+F)
Fitted by Monte Carlo optimization; determines discretization resolution of two-body term.
nRs3 (number of three-body radial bins) = 17 (E); 20 (E+F)
Fitted by Monte Carlo optimization; determines discretization resolution of three-body term.
Kernel width σ = Selected per dataset from {1,2,4,8,16,32} via nested CV
Standard hyperparameter of the Gaussian kernel; chosen by inner CV for each benchmark, affects all accuracy numbers.
Regularization λ (KRR/GPR) or εmin (OQML) = λ from {1e-10,...,1e-6}; εmin from {0,1e-12,...,1e-6} per dataset
Selected via nested CV; controls numerical stability and regression fit, affects reported errors.

assumptions (5)

domain assumption Two- and three-body terms are sufficient to represent atomic environments for accurate energy and force learning; one-body and higher-order terms are omitted.
Section II A states 'we do not consider terms of higher order than three-body' and the model achieves chemical accuracy on benchmarks, but this truncation is an ansatz.
ad hoc to paper Log-normal distributions are a suitable radial basis for the representation.
Section II A 1: 'We found it advantageous to use log-normal distribution functions... We note that this is an empirical choice.'
domain assumption The element-screened Gaussian kernel (Eq. 26) is a suitable similarity measure for atomic environments.
Section II B 5; the kernel is introduced without derivation and is standard for local atomic representations.
domain assumption Reference energies and forces in the benchmark datasets (DFT, MP2) are accurate enough to serve as ground truth.
Section IV A describes the datasets; the paper relies on their reported values without independent verification.
domain assumption Monte Carlo optimization of representation parameters on a separate 576-geometry set yields parameters that transfer across datasets.
Section IV B: two universal parameter sets are fitted a priori on the small set and used for all benchmarks; transferability is an empirical claim tested by the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of FCHL revisited: faster and more accurate quantum machine learning." pith.science (2026). https://pith.science/paper/3CCXDWBU

@misc{pith2026190901946,
  author       = {Pith},
  title        = {Pith review of: FCHL revisited: faster and more accurate quantum machine learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3CCXDWBU}},
  note         = {Machine review of arXiv:1909.01946}
}

read the original abstract

We introduce the FCHL19 representation for atomic environments in molecules or condensed-phase systems. Machine learning models based on FCHL19 are able to yield predictions of atomic forces and energies of query compounds with chemical accuracy on the scale of milliseconds. FCHL19 is a revision of our previous work [Faber et al. 2018] where the representation is discretized and the individual features are rigorously optimized using Monte Carlo optimization. Combined with a Gaussian kernel function that incorporates elemental screening, chemical accuracy is reached for energy learning on the QM7b and QM9 datasets after training for minutes and hours, respectively. The model also shows good performance for non-bonded interactions in the condensed phase for a set of water clusters with an MAE binding energy error of less than 0.1 kcal/mol/molecule after training on 3,200 samples. For force learning on the MD17 dataset, our optimized model similarly displays state-of-the-art accuracy with a regressor based on Gaussian process regression. When the revised FCHL19 representation is combined with the operator quantum machine learning regressor, forces and energies can be predicted in only a few milliseconds per atom. The model presented herein is fast and lightweight enough for use in general chemistry problems as well as molecular dynamics simulations.

Figures

Figures reproduced from arXiv: 1909.01946 by the authors.

**Figure 2.** FIG. 2. The three-body basis functions are plotted for the two unique [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** FIG. 3. Learning curves for the QM9 dataset: The mean absolute [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 5.** Figure 5: FIG. 5. The two figures display the out-of-training error for models trained on subsets of the QM7b-T dataset. In (A) the models predict [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: FIG. 6. Learning curves for the Water40 dataset: The mean absolute [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

**Figure 7.** Figure 7: FIG. 7. Here we present learning curves for force and energy learning for seven molecules from the MD17 dataset. Learning curves are [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

**Figure 8.** Figure 8: FIG. 8. The time to calculate the training kernel for 1,000 training [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Regression-clustering for Improved Accuracy and Training Cost with Molecular-Orbital-Based Machine Learning
physics.chem-ph 2019-09 conditional novelty 6.0 of 10

Clustering molecular-orbital training data into locally linear groups before regression cuts MOB-ML training time by up to 35,000-fold while retaining chemical accuracy.
Physical machine learning outperforms "human learning" in Quantum Chemistry
physics.chem-ph 2019-08 conditional novelty 6.0 of 10

A neural network using only a Hartree-Fock electron density as input predicts CCSD(T)-level energies (MAE 0.87 kcal/mol) and coupled-cluster densities for QM9 molecules, outperforming DFT and prior ML methods.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · cited by 2 Pith papers

[1]

Each of the nRs2 basis functions in this set is placed on an equidistant grid from rcut nRs2 to rcut, with rcut being the cutoff radius

Two-body function For a given central atom, a set of radial basis functions is constructed for each unique type of element in the data set. Each of the nRs2 basis functions in this set is placed on an equidistant grid from rcut nRs2 to rcut, with rcut being the cutoff radius. We found it advantageous to use log-normal distribu- tion functions for the radi...

work page
[2]

operator quantum machine learning

Three-body function The three-body function encodes the distances from an atom to neighboring pairs of atoms in the environment of the atom, as well as the angle between the triplet, and the element types of the neighbors. The resulting function is a product of the following terms: G3-body = ξ3G3-body Radial G3-body Angular fcut (rIJ ) fcut (rJK ) fcut (r...

work page
[3]

1,24,25,33–36,43 In KRR the energy, U∗, of a query compound, c, can be decomposed into the sum of atomic energies

Kernel Ridge Regression (KRR) It is well-established that KRR—despite its simplicity— is one of the most powerful methods to learn energies of chemical compounds. 1,24,25,33–36,43 In KRR the energy, U∗, of a query compound, c, can be decomposed into the sum of atomic energies. These are calculated in a basis of ker- nel functions placed on the atoms of th...

work page
[4]

In the operator quantum machine learning (OQML) approach introduced in ref

Operator Quantum Machine Learning (OQML) It is advantageous to also include forces in the training step if available, as this both improves energy and force prediction. In the operator quantum machine learning (OQML) approach introduced in ref. 22, the model is trained on the energy and forces simultaneously. The kernel is expanded in a basis of kernel fu...

work page
[5]

Gaussian Process Regression Including Derivatives It is also possible to deﬁne models that incorporate deriva- tives in the training set within the framework of Gaussian pro- cess regression.41 The relevant equations for training a model on energies and forces for chemical compounds are presented below. For their derivation we refer the reader to the work...

work page
[6]

Gradient-Domain Machine Learning (GDML) Since we will be comparing numerical results from the GDML28 and the closely related sGDML29 methods, we also brieﬂy review these approaches for the sake of completeness. GDML can be seen as equivalent to the GPR approach de- tailed above, with the difference that the energy is left out of the training data, such th...

work page
[7]

universal

Kernel Function We introduce a variant of the Gaussian kernel function, aug- mented with an elemental screening function that only com- pares representations for atomic environments of atoms of the same element type: k (qI, q∗ J) = δZIZ∗ J exp ( −‖qI− q∗ J‖2 2 2σ 2 ) (26) where δ is the Kronecker delta function and the subscripts ZI and Z∗ J are the nucle...

work page
[8]

3 we compare the predictive accuracy of a num- ber of kernel-based models for the atomization energy of molecules in the QM9 dataset

Results for QM9 In Fig. 3 we compare the predictive accuracy of a num- ber of kernel-based models for the atomization energy of molecules in the QM9 dataset. 46 We compare FCHL19 to ﬁve other well-performing representations: the SOAP multi- kernel model,34,35 SchNet17 and PhysNet21 which are two the best performing neural networks for this dataset, SLATM ...

work page

Show all 19 references

[9]

Results for QM7b Similarly, Fig. 4 compares the predictive accuracy of a number of kernel-based models for the atomization energy of the QM7b dataset.48 We compare our model to the following representations: FCHL18,1 SLATM,36 the Coulomb matrix,24 Bags-of-Bonds (BoB), 25 and ﬁ...
[10]

Results for QM7b-T and GDB13-T While the QM7b and QM9 datasets contain energies for the equilibrium geometry of small molecules, the QM7b-T49 and GDB13-T49 datasets contain non-equilibrium geometries of molecules from QM7b48 and GDB-13.50 In addition to gaug- ing the accuracy ...
[11]

22 As such, this dataset probes the performance of ML models on chemical systems that approach the condensed phase behavior

Results for Water40 The Water40 dataset consists of 10,000 MD snapshots of a water cluster with 40 water molecules for which a DFT single- point energy has been calculated. 22 As such, this dataset probes the performance of ML models on chemical systems that approach the conde...
[12]

Results for MD17 Fig. 7 reports the MAE force and energy prediction as a function of the number of training samples taken from 7 molecules from the MD17 dataset.28 We note that the original MD17 dataset also includes a dataset for benzene. However, due to low accuracy in the r...
[13]

For FCHL18 and aSLATM, all parameters are set to the default values in QML, and for FCHL19, the values in Appendix A are used

Timings for Energy Learning Using the implementations in the QML software package,54 we compare timings for calculating kernels for three represen- tations that all use a decomposition of the kernel into atomic contributions, namely FCHL19, FCHL18, and aSLATM. For FCHL18 and a...
[14]

These timings are given in Table II

Timings for Force Learning Next, we report timings for kernel evaluations for calculat- ing the training kernel for force and energies for a set of 1K molecules taken from the MD17 dataset. These timings are given in Table II. Again, in all cases, the training times scale as O...
[15]

For each molecule, the Perdew–Burke-Ernzerhof (PBE) equilib- rium geometry is available along with 13 different properties also calculated at the DFT level

QM7b The QM7b dataset 48 is based on a subset of the GDB-13 database,50 and consists of 7,211 molecules with up to 7 atoms of the elements CNOSCl, saturated with hydrogen atoms. For each molecule, the Perdew–Burke-Ernzerhof (PBE) equilib- rium geometry is available along with ...
[16]

uncharacterized

QM9 The QM9 46 dataset is similar to QM7b, only it is based on a subset of GDB-17 database. 55 In contrast to QM7b, the QM9 dataset is much larger, and contains 133 885 molecules with up to 9 atoms of the type CNOF saturated with hydrogen atoms. For each , the B3LYP equlibrium...
[17]

QM7b-T and GDB13-T The QM7b-T and GDB13-T datasets 49 consist of non- equilibrium geometries sampled from ab initio molecu- lar dynamics simulations at 350K. QM7b-T contains non- equlibrium structures of molecules from QM7b, while GDB13-T contains non-equilibrium structures of...
[18]

For each sample, a dispersion-corrected DFT singlepoint energy is calculated at the PBEh-3c level of theory

Water40 The Water40 dataset 1 consists of 10 000 MD snapshots from a molecular dynamics simulation of a water cluster with 40 water molecules sampled at 300K. For each sample, a dispersion-corrected DFT singlepoint energy is calculated at the PBEh-3c level of theory. 57 In con...
[19]

universal

MD17 The MD17 dataset 28 contains snapshots from ab inito molecular dynamics on a number of small organic molecules for which reference force and energies are calculated at the DFT level. Out of the dataset we benchmark our models on force and energy data from the molecules et...

2018 arXiv

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Each of the nRs2 basis functions in this set is placed on an equidistant grid from rcut nRs2 to rcut, with rcut being the cutoff radius

Two-body function For a given central atom, a set of radial basis functions is constructed for each unique type of element in the data set. Each of the nRs2 basis functions in this set is placed on an equidistant grid from rcut nRs2 to rcut, with rcut being the cutoff radius. We found it advantageous to use log-normal distribu- tion functions for the radi...

work page

[2] [2]

operator quantum machine learning

Three-body function The three-body function encodes the distances from an atom to neighboring pairs of atoms in the environment of the atom, as well as the angle between the triplet, and the element types of the neighbors. The resulting function is a product of the following terms: G3-body = ξ3G3-body Radial G3-body Angular fcut (rIJ ) fcut (rJK ) fcut (r...

work page

[3] [3]

1,24,25,33–36,43 In KRR the energy, U∗, of a query compound, c, can be decomposed into the sum of atomic energies

Kernel Ridge Regression (KRR) It is well-established that KRR—despite its simplicity— is one of the most powerful methods to learn energies of chemical compounds. 1,24,25,33–36,43 In KRR the energy, U∗, of a query compound, c, can be decomposed into the sum of atomic energies. These are calculated in a basis of ker- nel functions placed on the atoms of th...

work page

[4] [4]

In the operator quantum machine learning (OQML) approach introduced in ref

Operator Quantum Machine Learning (OQML) It is advantageous to also include forces in the training step if available, as this both improves energy and force prediction. In the operator quantum machine learning (OQML) approach introduced in ref. 22, the model is trained on the energy and forces simultaneously. The kernel is expanded in a basis of kernel fu...

work page

[5] [5]

Gaussian Process Regression Including Derivatives It is also possible to deﬁne models that incorporate deriva- tives in the training set within the framework of Gaussian pro- cess regression.41 The relevant equations for training a model on energies and forces for chemical compounds are presented below. For their derivation we refer the reader to the work...

work page

[6] [6]

Gradient-Domain Machine Learning (GDML) Since we will be comparing numerical results from the GDML28 and the closely related sGDML29 methods, we also brieﬂy review these approaches for the sake of completeness. GDML can be seen as equivalent to the GPR approach de- tailed above, with the difference that the energy is left out of the training data, such th...

work page

[7] [7]

universal

Kernel Function We introduce a variant of the Gaussian kernel function, aug- mented with an elemental screening function that only com- pares representations for atomic environments of atoms of the same element type: k (qI, q∗ J) = δZIZ∗ J exp ( −‖qI− q∗ J‖2 2 2σ 2 ) (26) where δ is the Kronecker delta function and the subscripts ZI and Z∗ J are the nucle...

work page

[8] [8]

3 we compare the predictive accuracy of a num- ber of kernel-based models for the atomization energy of molecules in the QM9 dataset

Results for QM9 In Fig. 3 we compare the predictive accuracy of a num- ber of kernel-based models for the atomization energy of molecules in the QM9 dataset. 46 We compare FCHL19 to ﬁve other well-performing representations: the SOAP multi- kernel model,34,35 SchNet17 and PhysNet21 which are two the best performing neural networks for this dataset, SLATM ...

work page