RLEASE: Reinforcement Learning Efficient Active Space Engine

Abhishek Mitra; Andrew J. Jenkins; Arpan Kundu; Arvin Kakekhani; Dario Rocca; Etinosa Osaro; Kelsey A. Parker; Robert H. Lavroff; Verena A. Neufeld

arxiv: 2606.07879 · v1 · pith:HK4JWSHHnew · submitted 2026-06-05 · ⚛️ physics.chem-ph

RLEASE: Reinforcement Learning Efficient Active Space Engine

Etinosa Osaro , Abhishek Mitra , Andrew J. Jenkins , Kelsey A. Parker , Robert H. Lavroff , Verena A. Neufeld , Arpan Kundu , Arvin Kakekhani

show 1 more author

Dario Rocca

This is my paper

Pith reviewed 2026-06-27 20:02 UTC · model grok-4.3

classification ⚛️ physics.chem-ph

keywords active space selectionreinforcement learningmultireference methodselectronic structure theoryneural networkpotential energy surfacesquantum chemistryorbital descriptors

0 comments

The pith

A neural network trained by reinforcement learning selects compact active spaces that transfer to chemically diverse molecules without retraining.

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

The paper presents RLEASE, which trains a neural network to assign diagnostic scores to orbitals using only inexpensive Hartree-Fock descriptors. Proximal policy optimization tunes a threshold that partitions orbitals into active and inactive sets, with the reward signal coming from the energy discrepancy between sc-NEVPT2 on the chosen space and DMRG references. After training on a limited set of molecules and geometries, the same network produces active spaces that yield competitive potential-energy surfaces on new systems. At deployment the method needs only the cheap descriptors and a forward pass, avoiding any target-system DMRG or retraining. This removes a long-standing expert-intuition bottleneck in multireference electronic-structure work.

Core claim

RLEASE trains a neural network to predict per-orbital scores from Hartree-Fock descriptors and optimizes a learned threshold via proximal policy optimization; the reward is the sc-NEVPT2 versus DMRG energy discrepancy on the selected active space. The resulting policy, trained on a small collection of molecules and geometries, produces compact active spaces that deliver potential-energy surfaces competitive with entropy-based selectors when applied to chemically diverse test systems, all at the cost of a single inexpensive orbital-descriptor evaluation and neural-network inference.

What carries the argument

A neural network that maps inexpensive Hartree-Fock orbital descriptors to per-orbital diagnostic scores, combined with a learned threshold policy optimized by proximal policy optimization using sc-NEVPT2-DMRG discrepancy as the reward signal.

If this is right

High-throughput multireference calculations become possible because deployment requires only cheap orbital descriptors and a single neural-network inference.
The same trained network can be used with multireference perturbation theory or composite coupled-cluster estimators on new molecules.
Active-space selection no longer requires molecule-specific pilot DMRG calculations or retraining.
Compact active spaces are produced that remain competitive with entropy-based methods across diverse chemical systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar reinforcement-learning policies could be explored for other selection tasks in quantum chemistry that currently rely on expensive reference calculations.
The approach might scale to larger molecules where full DMRG references become prohibitive, provided the orbital-descriptor representation remains informative.
Integration with existing multireference workflows could reduce the expert time currently spent on trial-and-error active-space trials.

Load-bearing premise

The reward signal derived from sc-NEVPT2 versus DMRG discrepancy on a small training set will continue to produce accurate and compact active-space choices when the same network is applied to new, chemically diverse molecules.

What would settle it

On a chemically diverse molecule outside the training distribution, the RLEASE-selected active space produces a potential-energy surface whose error relative to DMRG reference values is substantially larger than the error obtained with an established entropy-based active-space selector.

Figures

Figures reproduced from arXiv: 2606.07879 by Abhishek Mitra, Andrew J. Jenkins, Arpan Kundu, Arvin Kakekhani, Dario Rocca, Etinosa Osaro, Kelsey A. Parker, Robert H. Lavroff, Verena A. Neufeld.

**Figure 2.** Figure 2: FIG. 2. Per-molecule active-space sizes selected by [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Per-molecule Jaccard similarity between [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Distribution of [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

**Figure 5.** Figure 5: FIG. 5. Relative PES binding curves for CASCI + sc-NEVPT2 using active spaces selected by [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6. Relative PES binding curves for ASF-CCSD using active spaces selected by [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗

**Figure 7.** Figure 7: FIG. 7. Relative PES binding curves for ASF-CCSD(T) using active spaces selected by [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗

**Figure 8.** Figure 8: FIG. 8. Relative PES binding curves for [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗

**Figure 9.** Figure 9: FIG. 9. The six active orbitals selected by [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗

read the original abstract

Selecting the active space for multireference electronic-structure calculations is a long-standing bottleneck that often requires expert chemical intuition and costly trial-and-error. We introduce RLEASE (Reinforcement Learning Efficient Active Space Engine), a low-cost method for automatic, geometry-dependent active-space selection. A neural network predicts per-orbital diagnostic scores ($\hat{s}_{1}$) from inexpensive Hartree-Fock orbital descriptors, and a learned threshold partitions orbitals into active and inactive sets. The threshold policy is optimized with proximal policy optimization, using the discrepancy between sc-NEVPT2 energies computed with the selected active space and DMRG reference energies as the reward. After training, the same RLEASE-selected active spaces can be used with multireference perturbation theory or composite coupled-cluster energy estimators. Despite being trained on a small set of molecules and geometries, RLEASE transfers to chemically diverse test systems, producing compact active spaces and competitive potential-energy surfaces relative to established entropy-based selectors. Because deployment requires only inexpensive orbital descriptors and neural-network inference, RLEASE enables high-throughput multireference workflows without molecule-specific retraining or target-system pilot DMRG calculations.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

RLEASE trains a PPO policy on DMRG discrepancies to threshold NN scores from HF descriptors and claims the resulting active-space rule transfers to new molecules without retraining.

read the letter

RLEASE trains a neural network on Hartree-Fock orbital features to predict scores, then uses PPO to optimize a threshold policy so that the selected active space makes sc-NEVPT2 close to DMRG energies. The headline result is that this learned policy transfers to new molecules and geometries without retraining or additional DMRG runs.

What stands out is the shift from static rules or per-system entropy calculations to a learned, geometry-aware selector. The paper reports that the resulting active spaces stay compact and produce potential energy surfaces that compete with established methods on the test cases.

The method is practical at inference time because it only needs cheap HF calculations plus a forward pass. That part is a clear win for throughput.

The soft spot is the dependence on DMRG during training. The reward function ties the policy directly to reproducing DMRG-quality spaces on the training set, so success on new systems rests on how well the HF descriptors and the learned threshold capture the relevant electronic features across different bonding and elements. The abstract claims good transfer, but the strength of that claim depends on the actual spread of the training molecules, the quantitative gaps versus entropy selectors, and whether error bars or controls rule out post-hoc tuning.

The circularity is partial rather than total, since deployment avoids DMRG, but it still means the method is supervised by an expensive reference.

This work is for computational chemists running multireference calculations who want to cut down on trial-and-error active space selection. A reader looking for a drop-in tool would get value if the transfer results hold with solid controls.

I would send this to peer review. The idea is sound enough and the problem important enough that referees should check the experiments and the generalization evidence.

Referee Report

2 major / 1 minor

Summary. The paper introduces RLEASE, a reinforcement learning method (using PPO) for automatic, geometry-dependent active-space selection in multireference calculations. A neural network predicts per-orbital diagnostic scores from inexpensive Hartree-Fock orbital descriptors; a learned threshold partitions orbitals into active/inactive sets. The policy is optimized with a reward equal to the discrepancy between sc-NEVPT2 energies (using the selected space) and DMRG reference energies. After training on a small set of molecules and geometries, the fixed network is claimed to transfer to chemically diverse test systems, yielding compact active spaces whose subsequent multireference energies are competitive with those from established entropy-based selectors, all at the cost of only HF descriptors plus NN inference.

Significance. If the transferability claim holds, RLEASE would remove a major practical bottleneck for high-throughput multireference workflows by eliminating molecule-specific DMRG pilot calculations and expert trial-and-error while still producing usable active spaces. The low inference cost is a genuine practical strength.

major comments (2)

[Abstract] Abstract: the central transferability claim ("Despite being trained on a small set of molecules and geometries, RLEASE transfers to chemically diverse test systems") is load-bearing for the deployment narrative, yet the reward is defined directly as sc-NEVPT2–DMRG discrepancy; this creates an explicit dependence on the external DMRG benchmark during training that must be shown not to produce overfitting to the training distribution's electronic motifs.
[Abstract] Abstract: the claim that the learned threshold policy produces "competitive potential-energy surfaces relative to established entropy-based selectors" on test systems requires explicit quantitative support (e.g., error statistics or PES plots) demonstrating that the fixed network does not systematically select too many or too few orbitals when bonding types, spin states, or elements differ from the training set.

minor comments (1)

Notation for the predicted per-orbital score (â§s_{1}) should be defined once and used consistently; the abstract introduces it without prior definition.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We address each major comment below and outline the revisions we will make to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the central transferability claim ("Despite being trained on a small set of molecules and geometries, RLEASE transfers to chemically diverse test systems") is load-bearing for the deployment narrative, yet the reward is defined directly as sc-NEVPT2–DMRG discrepancy; this creates an explicit dependence on the external DMRG benchmark during training that must be shown not to produce overfitting to the training distribution's electronic motifs.

Authors: We acknowledge that the reward function relies on DMRG reference energies during training, which could in principle lead to overfitting to motifs present in the training distribution. The manuscript already demonstrates transfer to chemically diverse test systems without retraining, but to more explicitly rule out overfitting we will add a dedicated subsection describing the chemical and electronic diversity of the training molecules and geometries, together with additional out-of-distribution performance metrics on the test set. revision: partial
Referee: [Abstract] Abstract: the claim that the learned threshold policy produces "competitive potential-energy surfaces relative to established entropy-based selectors" on test systems requires explicit quantitative support (e.g., error statistics or PES plots) demonstrating that the fixed network does not systematically select too many or too few orbitals when bonding types, spin states, or elements differ from the training set.

Authors: We agree that the competitiveness claim requires explicit quantitative backing. In the revised manuscript we will include tables reporting mean absolute errors and maximum deviations in sc-NEVPT2 energies relative to DMRG for both RLEASE and entropy-based selections across the full test set, as well as representative potential-energy-surface plots that illustrate active-space sizes and energy errors for molecules with bonding types, spin states, and elements outside the training distribution. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation or training pipeline.

full rationale

The PPO reward is the sc-NEVPT2 vs. DMRG discrepancy on a fixed training collection; this is an external oracle used to supervise the policy. Deployment applies the fixed network to new molecules using only HF descriptors, with no further DMRG or retraining. The transferability claim is an empirical assertion that can be (and is) tested on held-out systems rather than being true by construction. No self-citation is invoked to justify uniqueness or to close the loop, and no equation or definition reduces the output active-space selection to the training reward itself. The chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the premise that a policy trained via DMRG-supervised reinforcement learning on a small molecular set will generalize; no free parameters are explicitly named beyond the learned NN weights and threshold, no new physical entities are introduced, and the axioms are standard RL convergence assumptions plus the domain assumption that sc-NEVPT2/DMRG discrepancy is a faithful proxy for active-space quality.

free parameters (1)

neural-network weights and PPO policy parameters
Learned during training on the small molecular set; the abstract does not report their number or regularization details.

axioms (2)

domain assumption sc-NEVPT2 energy discrepancy with DMRG is a suitable scalar reward for active-space quality
Invoked in the reward definition; no justification or sensitivity analysis is provided in the abstract.
domain assumption Hartree-Fock orbital descriptors contain sufficient information to predict useful active-space membership
Implicit in the choice of input features to the neural network.

pith-pipeline@v0.9.1-grok · 5762 in / 1629 out tokens · 24102 ms · 2026-06-27T20:02:54.002721+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 6 canonical work pages

[1]

B. O. Roos, R. Lindh, P. ˚A. Malmqvist, V. Veryazov, and P.-O. Widmark,Multiconfigurational Quantum Chemistry, 1st ed. (John Wiley & Sons, 2016)

2016
[2]

P. G. Szalay, T. M¨ uller, G. Gidofalvi, H. Lischka, and R. Shepard, Chemical Reviews112, 108 (2012), https://doi.org/10.1021/cr200137a

work page doi:10.1021/cr200137a 2012
[3]

D. I. Lyakh, M. Musia l, V. F. Lotrich, and R. J. Bartlett, Chemical Reviews112, 182 (2012), https://doi.org/10.1021/cr2001417

work page doi:10.1021/cr2001417 2012
[4]

B. O. Roos, P. R. Taylor, and P. E. Sigbahn, Chem. Phys.48, 157 (1980)

1980
[5]

P. E. M. Siegbahn, J. Alml¨ of, A. Heiberg, and B. O. Roos, J. Chem. Phys.74, 2384 (1981), https://doi.org/10.1063/1.441359

work page doi:10.1063/1.441359 1981
[6]

Siegbahn, A

P. Siegbahn, A. Heiberg, B. Roos, and B. Levy, Phys. Scr.21, 323 (1980)

1980
[7]

S. R. White, Physical Review Letters69, 2863 (1992)

1992
[8]

G. K.-L. Chan and S. Sharma, Annual Review of Physical Chemistry62, 465 (2011)

2011
[9]

R. J. Buenker and S. D. Peyerimhoff, Theoretica chimica acta35, 33 (1974)

1974
[10]

R. J. Buenker, S. D. Peyerimhoff, and W. Butscher, Mol. Phys.35, 771 (1978)

1978
[11]

Angeli, R

C. Angeli, R. Cimiraglia, S. Evangelisti, T. Leininger, and J.-P. Malrieu, Journal of Chemical Physics114, 10252 (2001)

2001
[12]

Angeli, S

C. Angeli, S. Borini, M. Cestari, and R. Cimiraglia, J. Chem. Phys.121, 4043 (2004), https://doi.org/10.1063/1.1778711

work page doi:10.1063/1.1778711 2004
[13]

Angeli, R

C. Angeli, R. Cimiraglia, and J.-P. Malrieu, Chem. Phys. Lett.350, 297 (2001)

2001
[14]

J. J. Wardzala, M. R. Hennefarth, V. Agarawal, B. Jangid, A. Seal, M. R. Hermes, D. S. King, and L. Gagliardi, Chemical Reviews126, 4592 (2026), pMID: 41930730, https://doi.org/10.1021/acs.chemrev.5c00866

work page doi:10.1021/acs.chemrev.5c00866 2026
[15]

Lloyd, Science273, 1073–1078 (1996)

S. Lloyd, Science273, 1073–1078 (1996)

1996
[16]

Verma, A

S. Verma, A. Mitra, Q. Wang, R. D’Cunha, B. Jangid, M. R. Hennefarth, V. Agarawal, L. Otis, S. Haldar, M. R. Hermes, and L. Gagliardi, Chemical Reviews126, 184 (2026), pMID: 41481354, https://doi.org/10.1021/acs.chemrev.5c00486

work page doi:10.1021/acs.chemrev.5c00486 2026
[17]

Veryazov, P

V. Veryazov, P. ˚A. Malmqvist, and B. O. Roos, International Journal of Quantum Chemistry111, 3329 (2011)

2011
[18]

Legeza and J

¨O. Legeza and J. S´ olyom, Physical Review B68, 195116 (2003)

2003
[19]

Rissler, R

J. Rissler, R. M. Noack, and S. R. White, Chemical Physics323, 519 (2006)

2006
[20]

C. J. Stein and M. Reiher, Journal of Chemical Theory and Computation12, 1760 (2016)

2016
[21]

E. R. Sayfutyarova, Q. Sun, G. K.-L. Chan, and G. Knizia, Journal of Chemical Theory and Computation13, 4063 (2017)

2017
[22]

Jeong, S

W. Jeong, S. J. Stoneburner, D. King, R. Li, A. Walker, R. Lindh, and L. Gagliardi, Journal of Chemical Theory and Computation16, 2389 (2020)

2020
[23]

Golub, A

P. Golub, A. Antalik, L. Veis, and J. Brabec, Journal of Chemical Theory and Computation17, 6053 (2021)

2021
[24]

Angeli, R

C. Angeli, R. Cimiraglia, and J.-P. Malrieu, Journal of Chemical Physics117, 9138 (2002)

2002
[25]

Sunet al., WIREs Computational Molecular Science8, e1340 (2018)

Q. Sunet al., WIREs Computational Molecular Science8, e1340 (2018)

2018
[26]

R. J. Bartlett and M. Musia l, Reviews of Modern Physics79, 291 (2007)

2007
[27]

Raghavachari, G

K. Raghavachari, G. W. Trucks, J. A. Pople, and M. Head-Gordon, Chemical Physics Letters157, 479 (1989)

1989
[28]

D. S. King and L. Gagliardi, Journal of Chemical Theory and Computation17, 2817 (2021)

2021
[29]

D. S. King, M. R. Hermes, D. G. Truhlar, and L. Gagliardi, Journal of Chemical Theory and Computation18, 6065 (2022)

2022
[30]

Schulman, F

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, arXiv preprint arXiv:1707.06347 (2017)

Pith/arXiv arXiv 2017
[31]

Zhai and G

H. Zhai and G. K.-L. Chan, Journal of Chemical Physics154, 224116 (2021)

2021
[32]

H. Zhai, H. R. Larsson, S. Lee, Z.-H. Cui, T. Zhu, C. Sun,et al., Journal of Chemical Physics159, 234801 (2023)

2023
[33]

A. E. Clark and E. R. Davidson, J. Org. Chem.68, 3387 (2003)

2003
[34]

Batatia, D

I. Batatia, D. P. Kov´ acs, G. N. C. Simm, C. Ortner, and G. Cs´ anyi, inAdvances in Neural Information Processing Systems, Vol. 35 (2022)

2022
[35]

Batzner, A

S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nature Communications13, 2453 (2022)

2022

[1] [1]

B. O. Roos, R. Lindh, P. ˚A. Malmqvist, V. Veryazov, and P.-O. Widmark,Multiconfigurational Quantum Chemistry, 1st ed. (John Wiley & Sons, 2016)

2016

[2] [2]

P. G. Szalay, T. M¨ uller, G. Gidofalvi, H. Lischka, and R. Shepard, Chemical Reviews112, 108 (2012), https://doi.org/10.1021/cr200137a

work page doi:10.1021/cr200137a 2012

[3] [3]

D. I. Lyakh, M. Musia l, V. F. Lotrich, and R. J. Bartlett, Chemical Reviews112, 182 (2012), https://doi.org/10.1021/cr2001417

work page doi:10.1021/cr2001417 2012

[4] [4]

B. O. Roos, P. R. Taylor, and P. E. Sigbahn, Chem. Phys.48, 157 (1980)

1980

[5] [5]

P. E. M. Siegbahn, J. Alml¨ of, A. Heiberg, and B. O. Roos, J. Chem. Phys.74, 2384 (1981), https://doi.org/10.1063/1.441359

work page doi:10.1063/1.441359 1981

[6] [6]

Siegbahn, A

P. Siegbahn, A. Heiberg, B. Roos, and B. Levy, Phys. Scr.21, 323 (1980)

1980

[7] [7]

S. R. White, Physical Review Letters69, 2863 (1992)

1992

[8] [8]

G. K.-L. Chan and S. Sharma, Annual Review of Physical Chemistry62, 465 (2011)

2011

[9] [9]

R. J. Buenker and S. D. Peyerimhoff, Theoretica chimica acta35, 33 (1974)

1974

[10] [10]

R. J. Buenker, S. D. Peyerimhoff, and W. Butscher, Mol. Phys.35, 771 (1978)

1978

[11] [11]

Angeli, R

C. Angeli, R. Cimiraglia, S. Evangelisti, T. Leininger, and J.-P. Malrieu, Journal of Chemical Physics114, 10252 (2001)

2001

[12] [12]

Angeli, S

C. Angeli, S. Borini, M. Cestari, and R. Cimiraglia, J. Chem. Phys.121, 4043 (2004), https://doi.org/10.1063/1.1778711

work page doi:10.1063/1.1778711 2004

[13] [13]

Angeli, R

C. Angeli, R. Cimiraglia, and J.-P. Malrieu, Chem. Phys. Lett.350, 297 (2001)

2001

[14] [14]

J. J. Wardzala, M. R. Hennefarth, V. Agarawal, B. Jangid, A. Seal, M. R. Hermes, D. S. King, and L. Gagliardi, Chemical Reviews126, 4592 (2026), pMID: 41930730, https://doi.org/10.1021/acs.chemrev.5c00866

work page doi:10.1021/acs.chemrev.5c00866 2026

[15] [15]

Lloyd, Science273, 1073–1078 (1996)

S. Lloyd, Science273, 1073–1078 (1996)

1996

[16] [16]

Verma, A

S. Verma, A. Mitra, Q. Wang, R. D’Cunha, B. Jangid, M. R. Hennefarth, V. Agarawal, L. Otis, S. Haldar, M. R. Hermes, and L. Gagliardi, Chemical Reviews126, 184 (2026), pMID: 41481354, https://doi.org/10.1021/acs.chemrev.5c00486

work page doi:10.1021/acs.chemrev.5c00486 2026

[17] [17]

Veryazov, P

V. Veryazov, P. ˚A. Malmqvist, and B. O. Roos, International Journal of Quantum Chemistry111, 3329 (2011)

2011

[18] [18]

Legeza and J

¨O. Legeza and J. S´ olyom, Physical Review B68, 195116 (2003)

2003

[19] [19]

Rissler, R

J. Rissler, R. M. Noack, and S. R. White, Chemical Physics323, 519 (2006)

2006

[20] [20]

C. J. Stein and M. Reiher, Journal of Chemical Theory and Computation12, 1760 (2016)

2016

[21] [21]

E. R. Sayfutyarova, Q. Sun, G. K.-L. Chan, and G. Knizia, Journal of Chemical Theory and Computation13, 4063 (2017)

2017

[22] [22]

Jeong, S

W. Jeong, S. J. Stoneburner, D. King, R. Li, A. Walker, R. Lindh, and L. Gagliardi, Journal of Chemical Theory and Computation16, 2389 (2020)

2020

[23] [23]

Golub, A

P. Golub, A. Antalik, L. Veis, and J. Brabec, Journal of Chemical Theory and Computation17, 6053 (2021)

2021

[24] [24]

Angeli, R

C. Angeli, R. Cimiraglia, and J.-P. Malrieu, Journal of Chemical Physics117, 9138 (2002)

2002

[25] [25]

Sunet al., WIREs Computational Molecular Science8, e1340 (2018)

Q. Sunet al., WIREs Computational Molecular Science8, e1340 (2018)

2018

[26] [26]

R. J. Bartlett and M. Musia l, Reviews of Modern Physics79, 291 (2007)

2007

[27] [27]

Raghavachari, G

K. Raghavachari, G. W. Trucks, J. A. Pople, and M. Head-Gordon, Chemical Physics Letters157, 479 (1989)

1989

[28] [28]

D. S. King and L. Gagliardi, Journal of Chemical Theory and Computation17, 2817 (2021)

2021

[29] [29]

D. S. King, M. R. Hermes, D. G. Truhlar, and L. Gagliardi, Journal of Chemical Theory and Computation18, 6065 (2022)

2022

[30] [30]

Schulman, F

J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, arXiv preprint arXiv:1707.06347 (2017)

Pith/arXiv arXiv 2017

[31] [31]

Zhai and G

H. Zhai and G. K.-L. Chan, Journal of Chemical Physics154, 224116 (2021)

2021

[32] [32]

H. Zhai, H. R. Larsson, S. Lee, Z.-H. Cui, T. Zhu, C. Sun,et al., Journal of Chemical Physics159, 234801 (2023)

2023

[33] [33]

A. E. Clark and E. R. Davidson, J. Org. Chem.68, 3387 (2003)

2003

[34] [34]

Batatia, D

I. Batatia, D. P. Kov´ acs, G. N. C. Simm, C. Ortner, and G. Cs´ anyi, inAdvances in Neural Information Processing Systems, Vol. 35 (2022)

2022

[35] [35]

Batzner, A

S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, Nature Communications13, 2453 (2022)

2022