betaFeedback

WorksPhysics and Astronomy

Machine Learning the Complete Active-Space Wavefunction Reconstruction

Abstract

Machine-learned interatomic potentials enable multireference-quality reaction dynamics but do not retain the electronic structure underlying the predicted energies and forces. Recovering this information by independently performing complete active space self-consistent field (CASSCF) calculations is both expensive and potentially inconsistent. We introduce the Complete Active-Space Wave Function Reconstruction (CAS-WR) model, which learns the geometry-dependent electronic structure from CASSCF data and reconstructs the CAS wave function without repeated orbital optimization. Combined with uncertainty-based active learning, CAS-WR is demonstrated for the Diels–Alder reaction and \ce{TiC+}-catalyzed C–H activation. The approach enables trajectory-resolved analysis of multiconfigurational electronic structure across reactive ensembles. Preliminary applications to excited-state proton transfer and transition-metal photocatalysis suggest that CAS-WR can provide a general route toward recovering electronic structure along complex ground- and excited-state reactive dynamics.

Introduction

Machine-learned interatomic potentials (MLIPs) have transformed molecular simulation by enabling trajectories that approach the accuracy of an underlying electronic-structure method at a small fraction of its computational cost. [1, 2, 3] Their accuracy has made it possible to study chemical reactions through long trajectories and statistically meaningful ensembles of reactive events. However, conventional MLIPs can only learn the potential energy and its nuclear derivatives. The electronic degrees of freedom are integrated out, and the resulting simulations do not retain direct access to the electronic structure associated with each molecular geometry. Consequently, a MLIP trajectory cannot by itself reveal how the electronic structure evolves during bond breaking and formation.

This limitation restricts the use of MLIPs for electronic structure-informed design. Quantities such as orbital occupations, configuration weights, charge and spin redistribution, and changes in bond order often provide the most direct description of a reaction mechanism. They are also needed to identify orbital-based principles for modifying catalysts, chromophores, and reactive intermediates. [4, 5, 6, 7] In the absence of an explicit electronic representation, such information must be recovered through additional electronic-structure calculations on sub-selected structures, [8, 9] yet it does not provide the ensemble-level description.

A growing class of machine-learning methods addresses this limitation by learning electronic-structure quantities rather than only energies and forces. [10, 11] These approaches have targeted effective Hamiltonians and wave functions, [12, 13, 14] electron densities, [15] and reduced density matrices, [16] thereby retaining access to a broader range of electronic observables than conventional MLIPs. Most such methods, however, have focused on mean-field methods like Hartree–Fock, Kohn–Sham density functional theory, or other predominantly single-reference descriptions, making their extension to reactions with substantial static correlation nontrivial. In regimes such as bond dissociation, open-shell intermediates, avoided crossings, and transition-metal-mediated reactions, multiple electronic configurations may contribute with comparable weight, motivating explicitly multiconfigurational descriptions. [17] Neural quantum Monte Carlo methods provide an alternative by representing correlated wave functions directly with neural-network ansätze and have shown high accuracy for strongly correlated systems and excited states. [19, 20, 21] [18] These approaches, nevertheless, require variational optimization, making independent wave function optimization across the many configurations encountered in a reaction-dynamics ensemble computationally demanding.

Complete active space self-consistent field (CASSCF) theory provides a natural framework for describing the multiconfigurational electronic structure of reactive systems. [22, 23] For CASSCF, both the configuration-interaction coefficients in a selected active space and the molecular orbitals defining that active space are variationally optimized. In our previous work, the weighted active space protocol (WASP) enabled the generation of consistent active-spaces across diverse geometries and the construction of MLIPs with multireference (MC-PDFT[24, 25]) accuracy. [26] This approach made multireference-quality reaction dynamics practical, but the resulting MLIPs still supplied only energies and forces. The CASSCF wave function and the electronic observables derived from it were not retained.

Reconstructing this information by independently repeating a CASSCF calculation at every geometry is problematic for reasons beyond computational cost. CASSCF orbital optimization is a nonlinear, non-convex optimization problem with multiple local minima, and independent calculations at neighboring trajectory geometries may therefore converge to different stationary solutions. [27, 28] This is particularly problematic for reaction trajectories, where substantial structural changes can alter the orbital landscape and cause the optimization to switch between different CASSCF solutions. Repeating CASSCF calculations post-hoc at selected trajectory geometries does not guarantee recovery of the same adiabatic surface, potentially leading to discontinuous or inconsistent electronic-structure assignments between the sampled geometries. The relevant goal is therefore not merely to obtain a converged CASSCF solution at each geometry, but to consistently recover the adiabatic surface represented by the multireference MLIP.

For the active space sizes that one can afford in reaction dynamics, a substantial fraction of the CASSCF computational cost arises from iterative orbital optimization. If suitable active-space orbitals can instead be supplied directly, the CI problem can be solved by CASCI without further orbital optimization. Approximate orbitals for CASCI can be obtained, for example, from unrestricted natural orbitals derived from broken-symmetry Hartree–Fock or from MP2 natural orbitals. [29, 30] These approaches can reduce the cost relative to CASSCF, but still require an iterative mean-field or post-mean-field calculation at each new geometry and do not explicitly preserve the multireference electronic structure associated with the trajectory. A complementary strategy was introduced by Booth and co-workers, where they used eigenvector continuation to interpolate correlated many-body wave functions by solving a generalized eigenvalue problem within a model space spanned by a set of reference wave functions. [31] These developments provide useful alternatives to repeated CASSCF orbital optimization and illustrate different routes to recovering active-space orbitals or correlated wave functions. In the context of multireference MLIP trajectories, it is also desirable to recover the corresponding electronic structure efficiently, consistently with the underlying adiabatic surface, and in a form that retains access to wave function derived observables.

Here, we introduce the complete active space wave function reconstruction (CAS-WR) model, a machine-learning framework that reconstructs the CAS wave function without repeated orbital optimization. CAS-WR learns a geometry-dependent representation of the converged CASSCF solution from reference data generated across configuration space. From this representation, the molecular orbitals are recovered and used in a subsequent CASCI calculation. The approach therefore does not merely predict selected electronic properties. It reconstructs the multiconfigurational wave function, from which configuration weights, natural-orbital occupations, charge and spin distributions, bond-order measures, and other wavefunction-derived observables can be evaluated.

Combined with uncertainty-based active learning, [26, 32] we demonstrate CAS-WR for the Diels–Alder reaction of 1,4-butadiene and ethylene and the TiC+\mathrm{TiC}^{+}-catalyzed C–H activation of methane. The resulting framework enables ensemble analysis of charge and spin redistribution, bond-order measures, configuration weights, natural-orbital occupations, and other properties derived from the reconstructed wave function. Therefore, this extends multireference-accurate MLIP dynamics from nuclear motion to trajectory-resolved electronic-structure analysis. Finally, we provide a preliminary demonstration of CAS-WR for excited-state transition-metal photocatalysis, highlighting its potential extension to excited states.

The central challenge is to identify a representation that is compact, varies smoothly with nuclear geometry, and retains sufficient information to recover the reference CASSCF solution. We consider two representations: the one-electron reduced density matrix D\mathbf{D} and the effective Fock matrix[22] F\mathbf{F} of the converged CASSCF solution,

Dpq=⟨Ψ∣a^p†a^q∣Ψ⟩,(1)D_{pq} = \langle\Psi| \hat{a}^{\dagger}_{p}\hat{a}_{q} | \Psi\rangle, \tag*{(1)}

and

Fpq=hpq+∑rsDrs[(pq∣rs)−12(pr∣qs)].(2)F_{pq} = h_{pq} + \sum_{rs} D_{rs} \left[ (pq|rs) - \frac{1}{2}(pr|qs) \right] . \tag*{(2)}

We assess their ability to reconstruct ground- and excited-state CAS wave functions for processes involving strong electron correlation, namely the ground-state dissociation of N2\mathrm{N}_{2} and the excited-state isomerization of fulvene. [33, 34] Natural orbitals were obtained by solving (SDS)CD=SCDn(\mathbf{SDS})\mathbf{C}_{D} = \mathbf{S}\mathbf{C}_{D}\mathbf{n}, while Fock-derived orbitals were obtained from FCF=SCFε\mathbf{F}\mathbf{C}_{F} = \mathbf{S}\mathbf{C}_{F}\boldsymbol{\varepsilon}, where S\mathbf{S} is the AO overlap matrix; each orbital set was used in CASCI, and the resulting energies were compared with the reference CASSCF energies. Figure 1 shows that orbitals reconstructed from D\mathbf{D} consistently yield better agreement with the reference CASSCF results, with a particularly pronounced improvement for the excited-state isomerization of fulvene. This advantage can be understood from the information encoded by the two representations: the 1-RDM directly captures the correlated one-electron structure through its natural orbitals and their occupations, whereas F\mathbf{F} is an effective one-electron quantity whose eigenvectors reflect the multiconfigurational wave function only indirectly. This limitation can become especially pronounced in state-averaged calculations, where constructing F\mathbf{F} from weightedaveraged densities further reduces its state-specific orbital character. This interpretation is supported by the orbital overlaps with the reference CASSCF solution, as discussed in Section SI of the Supporting Information. The poorer recovery obtained from F\mathbf{F} also explains why Marquetand and co-workers required subsequent orbital optimization when reconstructing wave functions from machine-learned Fock matrices. [35] Because such optimization would introduce substantial overhead in trajectory-wide electronic-structure analyses, we chose D\mathbf{D} as the target quantity for CAS-WR.

Assessment of electronic representations for CAS-WR

Figure 1. Assessment of electronic representations for CAS-WR. (a) Ground-state N2\mathrm{N}_{2} dissociation profiles from reference CASSCF and CASCI calculations using CAS(6e,6o)/cc-pVDZ. The CASCI orbitals are obtained by diagonalizing either the CASSCF effective Fock matrix, CAS(F\mathbf{F}), or the 1-RDM, CAS(D\mathbf{D}). (b) Corresponding excited-state double-bond torsion profiles for fulvene using SA(2)-CAS(6e,6o)/cc-pVDZ. Orbitals obtained from D\mathbf{D} reproduce the reference profiles more closely, particularly for fulvene. (c) CAS-WR reconstruction of the N2\mathrm{N}_{2} dissociation profile with increasing numbers of training geometries. (d) Corresponding CAS-WR reconstruction of the fulvene torsion profile.

The 1-RDM provides a compact representation intrinsically connected to configuration mixing and static correlation. [36] Having established the CASSCF 1-RDM as the target representation, we next require a geometry-dependent electronic descriptor from which the 1-RDM can be predicted. We use the one-electron Hamiltonian h(R)\mathbf{h}(\mathbf{R}), which provides an operator-level representation of the nuclear geometry and has recently been explored as an input for learning electronic-structure quantities. [16, 37] Following the work of Pavanello and co-workers, [16, 38] we formulate CAS-WR using kernel ridge regression (KRR), [39] which provides a nonlinear but data-efficient interpolation of the electronic structure over the sampled region of configuration space. We write the predicted CASSCF 1-RDM as

D^CAS−WR[h(R)]=∑i=1NtrainβiK(hi,h(R)),(3)\widehat{\mathbf{D}}_{\mathrm{CAS-WR}}[\mathbf{h}(\mathbf{R})] = \sum_{i=1}^{N_{\mathrm{train}}} \beta_i K(\mathbf{h}_i, \mathbf{h}(\mathbf{R})), \tag*{(3)}

where KK is the kernel function, and βi\beta_i contains the learned regression coefficients for the elements of D\mathbf{D}. We first evaluated CAS-WR on the N2N_2 dissociation and fulvene isomerization benchmarks by reconstructing their state-averaged (SA)-CASSCF potential-energy profiles, as shown in Figure 1. Accurate recovery required training geometries not only at the extrema of the reaction coordinate, but also at intermediate configurations that capture the evolution of the electronic structure between them. This indicates that extending CAS-WR to multidimensional systems requires an automated strategy for identifying and selecting the geometries most informative for training.

We next applied CAS-WR to the Diels–Alder cycloaddition of butadiene and ethylene to form cyclohexene, a multidimensional reaction with substantial multiconfigurational character.[21] The system was modeled using a (6e, 6o) active space comprising its entire π\pi system, with MC-PDFT employing the tPBE on-top functional and the def2-TZVP basis set.[40, 41] Details of the electronic structure setup are given in Section SII. To construct the training set efficiently, we used a two-stage active-learning protocol. Query-by-committee was first used to identify geometries with large disagreement among independently trained KRR models, [42, 43] thereby selecting configurations with high electronic-structure uncertainty. Gaussian-process diversity screening was then applied in the descriptor space to remove redundant structures, [32, 44] Further details are provided in Section SIII of the Supporting Information. The selected configurations were labeled with D and the model was iteratively retrained. Figure 2 shows the complete active learning cycle for CAS-WR. The distribution of configurations selected over successive active-learning rounds is shown in Figure 3a, while Figure 3b reports the corresponding error reduction, quantified by the mean absolute error of the CASCI energies obtained from the CAS-WR predicted 1-RDMs. The error in the predicted 1-RDMs along the active learning rounds are reported in Figure S4.

Active learning workflow to train CAS-WR

Figure 2. Active learning workflow to train CAS-WR

Distribution of configurations, CASCI energy errors, computational costs, and CASCI energy errors over active-learning rounds

Figure 3. (a) Distribution of configurations selected over successive active-learning rounds for the Diels–Alder cycloaddition of butadiene and ethylene. (b) Mean absolute error of CASCI energies evaluated using the CAS-WR-predicted 11-RDMs as a function of active-learning round for the Diels–Alder system. (c) Comparison of the computational cost of a single energy evaluation using conventional CASSCF, CAS-WR reconstruction, and MACE MLIP inference for the TiC+\mathrm{TiC}^{+}-catalyzed C–H activation of methane. MACE MLIP and CASSCF timings are taken from Seal et al.[26] (d) CASCI energy errors obtained from CAS-WR-predicted 11-RDMs over successive active-learning rounds for the TiC++CH4\mathrm{TiC}^{+}+\mathrm{CH}_{4} system.

Next, we demonstrate CAS-WR for an open-shell transition-metal system requiring a multireference description. [45] The reference data were taken from the active-learning data set previously constructed for the WASP-enabled multireference MLIP describing the TiC+\mathrm{TiC}^{+}-catalyzed C–H activation of methane at the SA(2)-CAS(7e,9o)/def2-TZVP level of theory. [26] These calculations already define a consistent active space and CASSCF solution branch across the reactive region, allowing CAS-WR to be trained directly on the same data, as an electronic-structure surrogate model. As shown in Figure 3c, CAS-WR is more computationally expensive than a MLIP(MACE[26, 46]) inference, but remains substantially cheaper than repeating CASSCF calculations. Crucially, CAS-WR restores access to arbitrary electronic observables that can be derived from the reconstructed CAS wave function and are not

available from a conventional MLIP unless explicitly included as additional learning targets. Figure 3d shows the reduction in CAS-WR errors (quantified in terms of CASCI energy errors) over successive active-learning rounds.

Whereas the previous study focused on the nuclear dynamics of proton-coupled electron transfer (PCET), here we use CAS-WR to resolve the accompanying electron transfer along the reactive trajectories. Figure 4a compares the CAS-WR energies with the corresponding MC-PDFT reference values, showing that CASCI in the predicted orbital basis recovers the correct minima along the reaction pathway. We then used CAS-WR to analyze the electronic-structure evolution over 100 reactive trajectories. The transferring hydrogen retains a positive Mulliken charge throughout the reaction, supporting proton transfer, consistent with the mechanism proposed by Geng et al.[45] Fragment-resolved Mulliken charges[47] further show that the TiC moiety becomes more positive while the emerging methyl fragment gains electron density (Figure 4b,Figure 4c), indicating electron transfer from TiC to CH3\mathrm{CH}_3.

Trajectory-resolved electronic-structure analysis of TiC⁺-catalyzed methane C–H activation using CAS-WR, with panels showing energy comparison, fragment-resolved Mulliken charges, Wiberg bond indices, and proton- and electron-transfer time distributions

Figure 4. Trajectory-resolved electronic-structure analysis of TiC+^{+}-catalyzed methane C–H activation using CAS-WR. (a) Comparison of MC-PDFT energies evaluated on top of CASCI wave functions obtained in the CAS-WR-predicted orbital basis with those obtained using independently computed HF orbitals and the corresponding reference energies along the reaction coordinate. (b, c) Fragment-resolved Mulliken charges of the TiC and CH3_3 moieties. (d, e) Wiberg bond indices describing cleavage of the Ti–Cc_c bond and formation of the Ti–Cm_m bond over a trajectory. (f) Distribution of the proton-transfer time, tPTt_{\mathrm{PT}} given by t(ηˉPT=0)t(\bar{\eta}_{\mathrm{PT}} = 0) and electron-transfer times, tETt_{\mathrm{ET}} given by t(ηˉET=0)t(\bar{\eta}_{\mathrm{ET}} = 0) over 100 reactive trajectories, where ηˉPT\bar{\eta}_{\mathrm{PT}} and ηˉET\bar{\eta}_{\mathrm{ET}} denote the corresponding normalized coordinates.

The bond rearrangement is captured by the Wiberg bond indices.[48, 49] The Ti−CTiC\mathrm{Ti}-\mathrm{C}_{\mathrm{TiC}} bond order decreases, while the Ti−CMe\mathrm{Ti}-\mathrm{C}_{\mathrm{Me}} bond order increases from approximately zero to one as the metal–methyl bond forms (Figure 4d,Figure 4e). Together with the positive charge on the transferring hydrogen and the increasing electron density on the methyl fragment, these trends support heterolytic cleavage of the methane C−H\mathrm{C}-\mathrm{H} bond, with proton transfer to CTiC\mathrm{C}_{\mathrm{TiC}} and electron transfer toward CH3\mathrm{CH}_3. The same mechanistic picture is obtained using intrinsic atomic orbital partitioning evaluated along a representative reactive trajectory (Figure S5).

To quantify the relative timing of proton and electron transfer, we define two collective variables from nuclear (ηPT\eta_{\mathrm{PT}}) and electronic (ηET\eta_{\mathrm{ET}}) degrees of freedom, respectively,

ηPT=CN(CMe,H)−CN(CTiC,H),ηET=qTiC−qCH3,(4)\eta_{\mathrm{PT}} = \mathrm{CN}(\mathrm{C}_{\mathrm{Me}}, \mathrm{H}) - \mathrm{CN}(\mathrm{C}_{\mathrm{TiC}}, \mathrm{H}), \qquad\eta_{\mathrm{ET}} = q_{\mathrm{TiC}} - q_{\mathrm{CH}_{3}}, \tag*{(4)}

where CN denotes the coordination number and q is the fragment-resolved charge. The corresponding crossing times provide trajectory-resolved proton- and electron-transfer timescales, allowing synchronous and asynchronous PCET pathways to be distinguished. [50] The distribution in Figure 4f indicates that the reaction proceeds through a synchronous PCET mechanism. This illustrates how CAS-WR enables ensemble electronic structure analysis which gives mechanistic insights, inaccessible from the MLIP trajectories alone.

As a preliminary demonstration of CAS-WR for excited-state reactivity, we tested the use of state-averaged 1-RDMs to interpolate excited-state potential-energy cuts. We first considered excited-state intramolecular proton transfer in 10-hydroxybenzo[h]quinoline (HBQ), a system previously studied using linearized-pair density functional theory (L-PDFT), [51, 52] and modeled here with SA(2)-tPBE(4e,4o)/6-31G** along the proton-transfer coordinate (Figure 5a). We then tested CAS-WR for transition-metal photocatalysis, modelling light-activated C–C coupling in dialkyl NiII{}^{\mathrm{II}} complexes. [53] Using (TMEDA)NiII{}^{\mathrm{II}}(CH3{}_{3})2{}_{2} as a model system, we constructed the potential-energy cut along the Ni–C(CH3{}_{3}) stretching coordinate at the SA(10)-tPBE(10e,6o)/6-31G++(d,p) level (Figure 5b). Starting from reference calculations at the endpoints of each reaction coordinate described below, we iteratively added geometries selected to be maximally distinct from those already included in the training set. For both systems, CAS-WR reproduced the full potential-energy profiles using less than 40% of the total sampled configurations. These results provide an initial demonstration that CAS-WR can be extended beyond ground-state wavefunctions to excited-state reactivity.

Figure 5. Excited-state reactivity with CAS-WR. (a) L-PDFT potential-energy profiles along the proton-transfer coordinate of 10-hydroxybenzo[h]quinoline. CAS-WR energies are obtained by performing multistate CASCI in the CAS-WR-predicted orbital basis followed by L-PDFT, and are compared with the corresponding reference L-PDFT energies. (b) L-PDFT potential-energy profiles along the Ni−C(CH3)\mathrm{Ni}-\mathrm{C}(\mathrm{CH}_{3}) bond-dissociation coordinate of the (TMEDA)NiII(CH3)2(\mathrm{TMEDA})\mathrm{Ni}^{\mathrm{II}}(\mathrm{CH}_{3})_{2} complex, obtained analogously from multistate CASCI wave functions in the CAS-WR-predicted orbital basis and compared with reference L-PDFT calculations. The inset shows the Ni−C\mathrm{Ni}-\mathrm{C} bond-breaking process. The training geometries for both the systems are marked by black dashes. (c) Natural orbitals of the (10e,6o)(10e, 6o) active space at the transition state of the Ni−C\mathrm{Ni}-\mathrm{C} photo-dissociation with the ground-state occupation numbers.

In summary, we introduced CAS-WR a machine-learning framework for reconstructing geometry-dependent CAS wave functions without repeated CASSCF orbital optimization.

CAS-WR learns the one-electron reduced density matrix of a reference CASSCF solution, uses its natural orbitals to define the active-space orbital basis, and recovers the corresponding multiconfigurational wave function through a subsequent CASCI calculation. This procedure reduces the computational cost by approximately a factor of 30 while preserving access to wavefunction-derived quantities. Among the representations examined, the 1-RDM reproduced the reference CASSCF orbital solution more reliably than the effective Fock matrix for both ground- and excited-state test cases.

For multidimensional reactive systems, we combined CAS-WR with a two-stage active-learning procedure based on 1-RDM uncertainty and structural diversity. Applications to the Diels–Alder reaction and TiC+^{+}-catalyzed methane activation show that the reconstructed wave functions provide natural-orbital occupations, configuration weights, charge and spin populations, and bond-order measures across complete reactive ensembles. In the methane-activation example, this information identifies the redistribution of charge during C–H cleavage and establishes the synchronous character of the proton-coupled electron-transfer mechanism—information that cannot be obtained from the MLIP trajectory alone. The results therefore demonstrate that CAS-WR goes beyond predicting individual electronic properties: it reconstructs the CAS wave function that can be interrogated through a broad range of electronic observables. Preliminary calculations for excited-state proton transfer and Ni–C photodissociation indicate that CAS-WR can be applied to excited state properties.

Further developments will focus on a differentiable implementation for evaluating gradients,[54] the calculation of spin–orbit couplings from the reconstructed multistate wave functions,[55] and direct learning of the 1- and 2-RDMs.[56] The latter would remove the CASCI step and permit direct evaluation of one- and two-electron observables and multireference energies. CAS-WR thus provides a route to retaining correlated electronic information across configuration space at a cost compatible with large-scale sampling.

Supporting Information

Computational details of the electronic-structure calculations and kernel ridge regression models, orbital-overlap quantification, active-learning workflow and associated error statis 马会 tics, and supplementary figures.

Data Availability

The CAS-WR code was developed using PySCF[57, 58, 59] (version 2.12.1) and scikit-learn[60] (version 1.7.2), and is available at https://github.com/aniruddha-seal/cas-wr.

Acknowledgments

A.S. acknowledges support from the Benjamin Ball Freud Merit Scholarship Fellowship from the Department of Chemistry at the University of Chicago. A.S., A.L.F., and L.G. acknowledge support from the Catalyst Design for Decarbonization Center, an Energy Frontier Research Center funded by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award No. DE-SC0023383. B.J. and L.G. acknowledge support from the Computational Chemical Sciences Program under Award No. DE-SC0023382, funded by the U.S. Department of Energy, Office of Basic Energy Sciences, the Chemical Sciences, Geosciences, and Biosciences Division. The authors also acknowledge the University of Chicago Research Computing Center for providing computational resources that supported this work.

Conflict of Interest Disclosure

A.L.F. is a co-founder and consultant of Evozyne, Inc. and a co-author of US Patent Applications 16/887,710 and 17/642,582, US Provisional Patent Applications 62/853,919, 62/900,420, 63/314,898, 63/479,378, 63/521,617, 63/510,130, 63/669,836, and 63/987,554 and International Patent Applications PCT/US2020/035206, PCT/US2020/050466, PCT/US24/10805, PCT/US24/34369, and PCT/US25/35833.

TOC Graphic

Figure 6 (PDF p. 23).

References

  1. [1]Deringer, V. L.; Caro, M. A.; Csányi, G. Machine Learning Interatomic Potentials as Emerging Tools for Materials Science. Advanced Materials 2019, 31, 1902765.DOI
  2. [2]Schran, C.; Thiemann, F. L.; Rowe, P.; Müller, E. A.; Marsalek, O.; Michaelides, A. Machine learning potentials for complex aqueous systems made simple. Proceedings of the National Academy of Sciences 2021, 118, e2110077118.DOI
  3. [3]Unke, O. T.; Chmiela, S.; Sauceda, H. E.; Gastegger, M.; Poltavsky, I.; Schütt, K. T.; Tkatchenko, A.; Müller, K.-R. Machine Learning Force Fields. Chemical Reviews 2021, 121, 10142–10186.DOI
  4. [4]Cui, L.; Feng, J.; Wu, S.; Xu, S.; Wu, X.; Zhang, S.; Zhang, Q.; Fu, G.; Wang, Y.; Xie, S. Photocatalyst Design with an Electrical-Polarization-Ability Descriptor for Exclusive C–H Bond Activation in Methanol. Journal of the American Chemical Society 2025, 147, 22509–22516.DOI
  5. [5]Wang, X. et al. eg occupancy as an effective descriptor for the catalytic activity of perovskite oxide-based peroxidase mimics. Nature Communications 2019, 10, 704.DOI
  6. [6]Zhang, Y.; Wu, Q.; Seow, J. Z. Y.; Jia, Y.; Ren, X.; Xu, Z. J. Spin states of metal centers in electrocatalysis. Chemical Society Reviews 2024, 53, 8123–8136.DOI
  7. [7]Suntivich, J.; Gasteiger, H. A.; Yabuuchi, N.; Nakanishi, H.; Goodenough, J. B.; Shao-Horn, Y. Design principles for oxygen-reduction activity on perovskite oxide catalysts for fuel cells and metal–air batteries. Nature Chemistry 2011, 3, 546–550.DOI
  8. [8]Yan, Y.; Yu, Y.; Hu, J.; Sun, X.; Luo, Q.; Chu, Z.; Dai, J.; Niu, H. Atomistic simulation study of diamond doping based on machine learning potential. Physical Chemistry Chemical Physics 2026, 28, 397–407.DOI
  9. [9]Yang, M.; Raucci, U.; Parrinello, M. Reactant-induced dynamics of lithium imide surfaces during the ammonia decomposition process. Nature Catalysis 2023, 6, 829–836.DOI
  10. [10]Tang, Z.; Chen, H.; Li, Y.; Qian, Y.; Wang, Y.; Fu, W.; Li, J.; Si, C.; Duan, W.; Chen, J.; Xu, Y. Deep-learning electronic structure calculations. Nature Computational Science 2025, 5, 1133–1146.DOI
  11. [11]Hindges, T. S.; He, W.; Li, J. Recent advances in electronic structure learning. Modelling and Simulation in Materials Science and Engineering 2026, 34, 053001.DOI
  12. [12]Schütt, K. T.; Gastegger, M.; Tkatchenko, A.; Müller, K.-R.; Maurer, R. J. Unifying machine learning and quantum chemistry with a deep neural network for molecular wavefunctions. Nature Communications 2019, 10, 5024.
  13. [13]Cignoni, E.; Suman, D.; Nigam, J.; Cupellini, L.; Mennucci, B.; Ceriotti, M. Electronic Excited States from Physically Constrained Machine Learning. ACS Central Science 2024, 10, 637–648.DOI
  14. [14]Gong, X.; Li, H.; Zou, N.; Xu, R.; Duan, W.; Xu, Y. General framework for E(3)-equivariant neural network representation of density functional theory Hamiltonian. Nature Communications 2023, 14, 2848.DOI
  15. [15]Grisafi, A.; Fabrizio, A.; Meyer, B.; Wilkins, D. M.; Corminboeuf, C.; Ceriotti, M. Transferable Machine-Learning Model of the Electron Density. ACS Central Science 2018, 5, 57–64.DOI
  16. [16]Shao, X.; Paetow, L.; Tuckerman, M. E.; Pavanello, M. Machine learning electronic structure methods based on the one-electron reduced density matrix. Nature Communications 2023, 14, 6281.arxiv.org/abs/2302.10741
  17. [17]Wardzala, J. J.; Hennefarth, M. R.; Agarawal, V.; Jangid, B.; Seal, A.; Hermes, M. R.; King, D. S.; Gagliardi, L. Multireference Methods for Chemistry and Materials Sci- PFAU, D. Ab initio solution of the many-electron Schrödinger equation with deep neural networks. Physical Review Research 2020, 2.
  18. [19]Hermann, J.; Schätzle, Z.; Noé, F. Deep-neural-network solution of the electronic Schrödinger equation. Nature Chemistry 2020, 12, 891–897.
  19. [20]Pfau, D.; Axelrod, S.; Sutterud, H.; von Glehn, I.; Spencer, J. S. Accurate computation of quantum excited states with neural networks. Science 2024, 385, eadn0137.DOI
  20. [21]Foster, A.; Schätzle, Z.; Szabó, P. B.; Cheng, L.; Köhler, J.; Cassella, G.; Gao, N.; Li, J.; Noé, F.; Hermann, J. An ab initio foundation model of wavefunctions that accurately describes chemical bond breaking. Nature Communications 2026,DOI
  21. [22]Roos, B. O.; Taylor, P. R.; Sigbahn, P. E. M. A complete active space SCF method (CASSCF) using a density matrix formulated super-CI approach. Chemical Physics 1980, 48, 157–173.DOI
  22. [23]Roos, B. O. Advances in Chemical Physics; John Wiley & Sons, Ltd, 1987; pp 399–445.
  23. [24]Li Manni, G.; Carlson, R. K.; Luo, S.; Ma, D.; Olsen, J.; Truhlar, D. G.; Gagliardi, L. Multiconfiguration Pair-Density Functional Theory. Journal of Chemical Theory and Computation 2014, 10, 3669–3680.
  24. [25]Sharma, P.; Bao, J. J.; Truhlar, D. G.; Gagliardi, L. Multiconfiguration Pair-Density Functional Theory. Annual Review of Physical Chemistry 2021, 72, 541–564.DOI
  25. [26]Seal, A.; Perego, S.; Hennefarth, M. R.; Raucci, U.; Bonati, L.; Ferguson, A. L.; Parrinello, M.; Gagliardi, L. Weighted active space protocol for multireference machine-learned potentials. Proceedings of the National Academy of Sciences 2025, 122, e2513693122.DOI
  26. [27]Saade, S.; Burton, H. G. A. Excited State-Specific CASSCF Theory for the Torsion of Ethylene. Journal of Chemical Theory and Computation 2024, 20, 5105–5114.arxiv.org/abs/2402.13046
  27. [28]Marie, A.; Burton, H. G. A. Excited States, Symmetry Breaking, and Unphysical Solutions in State-Specific CASSCF Theory. The Journal of Physical Chemistry A 2023, 127, 4538–4552.DOI
  28. [29]Bofill, J. M.; Pulay, P. The unrestricted natural orbital–complete active space (UNO–CAS) method: An inexpensive alternative to the complete active space–self-consistent-field (CAS–SCF) method. The Journal of Chemical Physics 1989, 90, 3637–3646.DOI
  29. [30]Abrams, M. L.; Sherrill, C. D. Natural orbitals as substitutes for optimized orbitals in complete active space wavefunctions. Chemical Physics Letters 2004, 395, 227–232.DOI
  30. [31]Rath, Y.; Booth, G. H. Interpolating numerically exact many-body wave functions for accelerated molecular dynamics. Nature Communications 2025, 16, 2005.DOI
  31. [32]Perego, S.; Bonati, L. Data efficient machine learning potentials for modeling catalytic reactivity via active learning and enhanced sampling. npj Computational Materials 2024, 10, 1–13.
  32. [33]Cramer, C. J. Essentials of Computational Chemistry: Theories and Models; John Wiley & Sons, 2013; Google-Books-ID: k4R6cf7I7q0C.
  33. [34]Bearpark, M. J.; Bernardi, F.; Olivucci, M.; Robb, M. A.; Smith, B. R. Can Fulvene S1 Decay Be Controlled? A CASSCF Study with MMVB Dynamics. Journal of the American Chemical Society 1996, 118, 5254–5260.DOI
  34. [35]Van Workum, R.; Malhado, J.; Marquetand, P. CASNet: Learning Complete Active Space Orbitals using Message Passing Neural Networks. 2023; https://chemrxiv.org/doi/full/10.26434/chemrxiv-2023-lwj87.DOI
  35. [36]Zhang, D.; Truhlar, D. G. Multiconfigurational Effects on the Density Coherence. Journal of Chemical Theory and Computation 2020, 16, 6915–6925.DOI
  36. [37]Nigam, J.; Smidt, T.; Dusson, G. Machine learning of electronic structure and atomistic properties from the external potential. The Journal of Chemical Physics 2026, 165, 034119.
  37. [38]Rana, B.; Viot, N.; Martinez B, J. A.; Shao, X.; Ramos, P.; Pavanello, M. Learning the One-Electron Reduced Density Matrix at SCF Convergence Thresholds. Journal of Chemical Theory and Computation 2025, 21, 12652–12663.DOI
  38. [39]Hastie, T.; Tibshirani, R.; Friedman, J. The Elements of Statistical Learning; Springer Series in Statistics; Springer: New York, NY, 2009.
  39. [40]Seal, A.; Gagliardi, L.; Ferguson, A. L. Computing Reaction Kinetics with MC-PDFT–OPESf: Combining Multireference Electronic Structure Theory and Enhanced Sampling. The Journal of Physical Chemistry Letters 2025, 16, 11458–11463.DOI
  40. [41]Mitchell, E. C.; Scott, T. R.; Bao, J. J.; Truhlar, D. G. Application of Multiconfiguration Pair-Density Functional Theory to the Diels–Alder Reaction. The Journal of Physical Chemistry A 2022, 126, 8834–8843.DOI
  41. [42]Seung, H. S.; Opper, M.; Sompolinsky, H. Query by committee. Proceedings of the fifth annual workshop on Computational learning theory. New York, NY, USA, 1992; pp 287–294.
  42. [43]Schran, C.; Brezina, K.; Marsalek, O. Committee neural network potentials control generalization errors and enable active learning. The Journal of Chemical Physics 2020, 153, 104105.DOI
  43. [44]Vandermause, J.; Xie, Y.; Lim, J. S.; Owen, C. J.; Kozinsky, B. Active learning of reactive Bayesian force fields applied to heterogeneous catalysis dynamics of H/Pt. Nature Communications 2022, 13, 5183.
  44. [45]Geng, C.; Weiske, T.; Li, J.; Shaik, S.; Schwarz, H. Intrinsic Reactivity of Diatomic 3d Transition-Metal Carbides in the Thermal Activation of Methane: Striking Electronic Structure Effects. Journal of the American Chemical Society 2019, 141, 599–610.DOI
  45. [46]Batat ia, I.; Kovacs, D. P.; Simm, G.; Ortner, C.; Csanyi, G. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. Advances in Neural Information Processing Systems 2022, 35, 11423–11436.
  46. [47]Mulliken, R. S. Electronic Population Analysis on LCAO–MO Molecular Wave Functions. I. The Journal of Chemical Physics 1955, 23, 1833–1840.DOI
  47. [48]Wiberg, K. B. Application of the pople-santr y-segal CNDO method to the cyclopropyl-carb inyl and cyclobutyl cation and to bicyclobutane. Tetrahedron 1968, 24, 1083–1096.
  48. [49]Mayer, I. Bond order and valence indices: A personal account. Journal of Computational Chemistry 2007, 28, 204–221.pubmed.ncbi.nlm.nih.gov/17066501
  49. [50]Cukier, R. I.; Nocera, D. G. Proton-Coupled Electron Transfer. Annual Review of Physical Chemistry 1998, 49, 337–369.pubmed.ncbi.nlm.nih.gov/9933908
  50. [51]Hennefarth, M. R.; Hermes, M. R.; Truhlar, D. G.; Gagliardi, L. Linearized Pair-Density Functional Theory. Journal of Chemical Theory and Computation 2023, 19, 3172–3183.DOI
  51. [52]Clifford, H. S.; Seal, A.; Gagliardi, L. Ultrafast excited-state proton transfer dynamics using linearized pair-density functional theory. Chemical Science 2026, 17, 15607–15615.DOI
  52. [53]Cusumano, A. Q.; Chaffin, B. C.; Cagan, D. A.; DiLuzio, S.; Sutcliffe, E.; Hadt, R. G.; Doyle, A. G. Light-Promoted C(sp3)–C(sp3) Reductive Elimination from Dialkyl NiII Complexes. Journal of the American Chemical Society 2025, 147, 32941–32950.DOI
  53. [54]Zhang, X.; Chan, G. K.-L. Differentiable quantum chemistry with PySCF for molecules and materials at the mean-field level and beyond. The Journal of Chemical Physics 2022, 157, 204801.DOI
  54. [55]Jangid, B.; Hennefarth, M. R.; Hermes, M. R.; Truhlar, D. G.; Gagliardi, L. Linearized Pair-Density Functional Theory with Spin–Orbit Coupling. Journal of Chemical Theory and Computation 2026, 22, 318–333.DOI
  55. [56]B, J. A. M.; Rana, B.; Shao, X.; Pernal, K.; Pavanello, M. Machine learning the two-electron reduced density matrix in molecules and condensed phases. 2026; http: //arxiv.org/abs/2603.06882, arXiv:2603.06882 [physics.chem-ph].arxiv.org/abs/2603.06882
  56. [57]Sun, Q.; Berkelbach, T. C.; Blunt, N. S.; Booth, G. H.; Guo, S.; Li, Z.; Liu, J.; Mc-Clain, J. D.; Sayfutyarova, E. R.; Sharma, S.; Wouters, S.; Chan, G. K.-L. PySCF: The Python-based Simulations of Chemistry Framework. Wiley Interdiscip. Rev. Comput. Mol. Sci. 2018, 8, e1340.arxiv.org/abs/1701.08223
  57. [58]Sun, Q. et al. Recent Developments in the PySCF Program Package. J. Chem. Phys. 2020, 153, 024109.DOI
  58. [59]Sun, Q. et al. The Python simulations of chemistry framework: 10 years of an open-source quantum chemistry project. The Journal of Chemical Physics 2026, 165, 102502.arxiv.org/abs/2603.14155
  59. [60]Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; others Scikit-learn: Machine learning in Python. the Journal of machine Learning research 2011, 12, 2825–2830.arxiv.org/abs/1201.0490

Paper details

vlnc:W_rq9kjsd8
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents