Journal of Chemical Theory and Computation最新文献

Spin–orbit coupling (SOC) induces state mixingand distortionsin potential energy curves (PECs), complicating their constructionand interpretation for diatomic molecules. We introduce a fully automatedmachine-learning framework for reconstructing physically consistentSOC-resolved PECs across independent electronic-structure calculationswith different basis-set representations. The approach combines generativedata augmentation using a conditional Wasserstein Generative AdversarialNetwork with gradient penalty (cWGAN-GP) with supervised regressionmodels (Ridge Regression, Gradient Boosting, and Extremely RandomizedTrees) to learn the relationship between raw SOC-affected PECs, spin–orbitcomposition tables, and their physically corrected counterparts. UsingMgCa as a demonstration system, SOC-resolved PECs and spectroscopicconstants were computed at the MRCI level for the ground and 26 low-lyingexcited states. Models trained on a single manually corrected referencebasis accurately predict physically corrected PEC energies for independentbasis sets, achieving sub-percent deviations in dissociation energieswhile preserving asymptotic limits, equilibrium geometries, and fine-structuresplittings without explicit constraints. This framework eliminatesmanual postprocessing of SOC-mixed PECs and provides a scalable strategyfor consistent spin–orbit treatment in high-level electronic-structureworkflows.

Zinc(Zn), the second most abundant transition metalin biologicalsystems, plays essential roles in numerous metalloproteins presentin diverse organisms including humans, plants, and bacteria. Currently,Zn-based proteins have attracted research interest owing to theirinvolvement in a series of vital physiological processes, such asenzymatic reactions, deoxyribonucleic acid and ribonucleic acid binding,and the stability of biological structures. Although experimentalmethods can reveal macroscopic processes, they cannot efficientlycapture the detailed behavior of these compounds. Therefore, theoreticalexploration of Zn-binding conformations is significant. Herein, afluctuating charge force field for Zn-containing metalloproteins isreported based on the atom-bond electronegativity equalization method(ABEEM). We designed our model molecules based on the Protein DataBank (PDB) database, focusing primarily on tetracoordinated Zn proteins.The ABEEM polarizable force field parameters were determined and optimizedusing the quantum mechanical method, which were in good agreementwith computed charge distributions, potential energy surfaces, andgeometric structures. Molecular dynamics simulations were subsequentlyperformed on 10 different protein segments. The developed ABEEM modelreproduces the first Zn–O_w peak position and maintainsthe structural stability of the systems throughout the simulations.Compared with the crystal structures from the PDB database, the ABEEMmodel shows lower root mean square deviation values for heavy atomsin these domains. All protein domains can dynamically capture thechange in charge distributions. This process can be further appliedto simulate the detailed behavior of the molecules and investigatetheir properties. The identified force field for Zn metalloproteinsby the ABEEM provides a promising approach for simulating Zn-relatedbiological systems while simplifying several complex and time-consumingprocesses.

We introduce a GPU-accelerated multigrid Gaussian-Plane-Wavedensityfitting (FFTDF) approach for efficient Fock builds and nuclear gradientevaluations within Kohn–Sham density functional theory, asimplemented in the GPU4PySCF module of PySCF. OurCUDA kernels employ a grid-based parallelization strategy for contractingGaussian basis function pairs and achieve 50% to 80% of the FP64 peakperformance on NVIDIA GPUs, with no loss of efficiency for high angularmomentum (up to f-shell) functions. Benchmark calculationson molecules and solids with up to 1536 atoms and 20480 basis functionsshow up to 25 × speedup on an H100 GPU relative to the CPU implementationon a 28-core shared-memory node. For a 256-water cluster, the ground-stateenergy and nuclear gradients can be computed in ∼30 secondson a single H100 GPU. This implementation serves as an open-sourcefoundation for many applications, such as ab initio molecular dynamics and high-throughput calculations.

High-resolution spectroscopy of weaklybound molecularcomplexesrequires theoretical predictions with uncertainties below 0.1 cm^–1. Achieving this level of accuracy places stringentdemands on the underlying full-dimensional potential energy surfaces(PESs). However, constructing such PESs remains challenging becauseflexible configuration space requires dense sampling and the ab initio energies used for fitting must meet spectroscopicaccuracy requirements. Here, we develop a physics-constrained synergeticactive- and transfer-learning framework (SynATL) that combines uncertainty-drivenactive learning, high-level refinement via transfer learning, andan explicit long-range representation to construct spectroscopicallyaccurate full-dimensional PESs. SynATL was used to construct a full-dimensionalPES of the H₂O–N₂ complex at the post-CCSD(T)level, achieving a weighted root-mean-square error (RMSE) of 0.815cm^–1 on the test set and sub-cm^–1 RMSEs in the spectroscopically relevant low-energy and attractiveregions. Applied to rovibrational calculations of H₂O–N₂ and D₂O–N₂, the resulting PESreproduces microwave transitions with a root-mean-square deviation(RMSD) of 1.6 × 10^–4 cm^–1, about 80% lower than previous theoretical values, and gives 10^–3 cm^–1 level band-specific RMSDsfor 235 assigned infrared lines across the H₂O–N₂ bending/overtone and D₂O–N₂ stretching/bendingbands. SynATL therefore provides an efficient and transferable frameworkfor full-dimensional PESs for quantitative spectroscopy of weaklybound complexes.

Hybrid density functionalapproximations are among themost widelyused electronic-structure methods, yet the physical origin of theirrepresentation of correlation effects remains incompletely understood.In this work, we analyze correlation in hybrid functionals from theperspectives of Kohn–Sham exchange–correlation potentials,correlated densities, and energy-related observables. We show thatthe dominant qualitative features of correlation in hybrid DFAs donot arise primarily from the semilocal correlation functional itself,but instead from the difference between the semilocal exchange andexact-exchange contributions. This exchange-difference term is essentialfor restoring the physically correct structure of the correlationcomponent of the exchange–correlation potential and the correspondingcorrelated density, whereas the semilocal correlation part acts mainlyas a compensating contribution in the core and valence regions. Basedon this analysis, we propose an interpretable construction principlefor hybrid functionals in which the semilocal/exact-exchange differenceis scaled in a controlled way. As a proof of concept, this methodologyyields the simple global hybrid UMK26 functional. Despite its uncomplicatedglobal-hybrid form, the functional produces correlation potentialsand correlated densities in good agreement with CCSD(T) references,delivers accurate ionization potentials from orbital energies, andyields very good charge-transfer excitation energies, performing ona comparable footing with several recently proposed range-separatedhybrids. Although UMK26 is not uniformly optimal across all thermochemicalbenchmarks, it performs particularly well for self-interaction-sensitivesystems and provides a physically transparent route for the rationaldesign of future hybrid density functionals.

Themany-body expansions of Hartree–Fock (HF)and correlationcontributions to water cluster MP2 binding energies converge differently:the HF contribution (one-body (1B) repulsive, two-body (2B) and three-body(3B) attractive) converges at the 3B term, whereas the MP2 correlationcontribution (1B and 2B attractive) converges at the 2B term, with99.7% arising from localized double excitations spanning one and twomonomers. Two orbital excitation truncation approximations, MP2-1B2B(geometry based) and localized LMP2-1M2M (excitation space based),recover ≥99.2% of the canonical MP2 binding energies.

Wepresent a near-full configuration interaction frameworkbasedon iterative configuration interaction with selection and second-orderperturbation theory (iCIPT2) for simulating vibrationally resolvedX-ray absorption spectra. By combining high-fidelity potential energycurves with quantum wavepacket dynamics, our approach captures strongcorrelation and nuclear dynamics. Benchmarking on six diatomic systemsyields sub-0.1 eV accuracy and quantitative agreement with high-resolutionexperiments. This work demonstrates a promising approach for interpretingvibrationally resolved core-level spectroscopy, established here fordiatomic systems as an initial step, and lays a theoretical foundationfor applications in more-complex ultrafast experiments at large-scaleX-ray facilities.

Efficient and reliable identification of transition states(TS)is critical for reaction modeling. Among the available approaches,the combination of double-ended TS search with eigenvector-following,referred to as hierarchical TS search, is an effective tool for locatingTSs starting from reactant and product geometries. However, systematicbenchmarking of these methods on transition metal-catalyzed reactionsremains limited. Here, we benchmark eight pipelined hierarchical methodsusing out-of-the-box settings on a diverse set of 30 single-step,closed-shell metal-catalyzed reactions, evaluating their success rateand computational efficiency. Most workflows achieve high successrates across the data set, indicating that modern double-ended methodsare generally robust for locating TSs in elementary organometallicreactions, and the main differences between methods lie in their computationalefficiency. The benchmark also reveals several key trends: (i) thefinal eigenvector-following step requires an accurate Hessian matrixor explicitly calculated min-mode for a high successrate, (ii) TS searches from interpolation-only paths (e.g., image-dependentpair potential, IDPP) and bracket methods (e.g., Dewar-Healy-Stewart,DHS) are often more efficient than popular methods, such as climbingimage nudged elastic band (CI-NEB), (iii) different hierarchical workflowscan converge to different TSs even when starting from the same twominima, (iv) larger structural changes between reactants and productsare associated with lower success rates and higher computational costs.We also show that further parameter tuning can reduce computationalexpense without reducing the success rate on our data set. These conclusionsare further supported by two case studies involving industrially relevantorganometallic reactions.

Excitationenergy transfer (EET) in molecular aggregates is a fundamentalprocess in a wide range of condensed-phase systems, including photosyntheticcomplexes and organic semiconductors. Among the key factors influencingEET efficiency, thermal fluctuations play a critical role; however,accurately capturing their impact at the atomistic level remains computationallychallenging. To address this, we developed an efficient simulationframework that combines the Frenkel exciton model, the nonadiabaticmolecular dynamics (NAMD) simulation, and the molecular mechanicswith Shepard interpolation correction (MMSIC) method. The MMSIC significantlyaccelerates the evaluation of excitation energies during the NAMDsimulation, enabling atomistic treatment of thermal fluctuations.We applied this framework to two representative systems: exciton diffusionin crystalline anthracene and EET in the Fenna–Matthews–Olsonlight-harvesting complex. In both cases, the calculated results arein good agreement with previous studies, demonstrating the capabilityof the proposed approach to explore the interplay between thermalfluctuations and EET dynamics in complex molecular systems.

Polarizable particle systems, including charged colloids,polarizableions, biomolecular assemblies, and soft nanomaterials, can exhibitcontact electrostatic interactions that depart strongly from Coulombbehavior when the dielectric mismatch and geometric singularitiesamplify polarization effects. Here, we use charged dielectric spheresas a model system and show that these polarization contributions canbe canceled by jointly tuning size, charge, and dielectric asymmetries.By extending a recently developed image-charge formula to contactingdielectric spheres, at the two-body level, we derive analytical conditionsunder which the two-body contact interaction reduces to the bare Coulombform. At the selected design points, accurate two-sphere calculationsgive root-mean-square (RMS) contact force deviations below 3% fromthe Coulomb target. Separately, many-body molecular dynamics simulationsthat retain nonadditive polarization show that under the tested assemblyconditions, the structural correlations of six systems satisfyingthese two-body rules closely match those of their pure Coulomb references.These results establish asymmetry as a route for turning electrostaticcomplexity into Coulombic simplicity at contact, with implicationsfor predictable and controlled self-assembly and material design.