PBE0+MBD MB-PIP: Dispersion & Many-Body Modeling

• PBE0+MBD MB-PIP potentials are transferable atom-based many-body models that use symmetry-adapted Morse-type polynomial functions to capture dispersion interactions.
• They combine a low-level PBE0+MBD electronic structure reference with a compact Δ-machine learning correction to approach high-level CCSD(T) accuracy.
• The framework informs related approaches in ethanol PES and molecular crystal embedding by emphasizing essential many-body corrections and symmetry adaptations.

Searching arXiv for the cited papers and related records. arxiv_search.query({"search_query":"id:(Qu et al., 22 Sep 2025) OR id:(Nandi et al., 2024) OR id:(List et al., 18 Dec 2025) OR id:(Li et al., 15 Feb 2026)","start":0,"max_results":10,"sort_by":"submittedDate","sort_order":"descending"}) Running a broader title/ID search on arXiv. arxiv_search.search({"query":"(Qu et al., 22 Sep 2025, Nandi et al., 2024, List et al., 18 Dec 2025, Li et al., 15 Feb 2026) PBE0+MBD MB-PIP","max_results":10}) Searching arXiv by title keywords and identifiers. to=arxiv_search.search 平台开号 北京赛车的 arguments as JSON {"query":"PBE0+MBD MB-PIP (Qu et al., 22 Sep 2025, Nandi et al., 2024, List et al., 18 Dec 2025, Li et al., 15 Feb 2026)","max_results":10} Searching arXiv for relevant papers. to=arxiv_search.search 天天彩票网 彩神争霸苹果={"query":"PBE0+MBD MB-PIP (Qu et al., 22 Sep 2025, Nandi et al., 2024, List et al., 18 Dec 2025, Li et al., 15 Feb 2026)","max_results":10} {"query":"PBE0+MBD MB-PIP (Qu et al., 22 Sep 2025, Nandi et al., 2024, List et al., 18 Dec 2025, Li et al., 15 Feb 2026)","max_results":10} to=arxiv_search.query {"search_query":"all:PBE0+MBD AND all:MB-PIP","start":0,"max_results":5} PBE0+MBD MB-PIP potentials are permutationally invariant polynomial potential-energy models in which the underlying electronic-structure reference is PBE0 augmented by the many-body dispersion model of Tkatchenko, DiStasio, Car, and Scheffler. In the current literature, the term refers most directly to a transferable atom-based many-body permutationally invariant polynomial potential for linear alkanes, and more broadly to closely related PBE0+MBD-based polynomial correction and embedding strategies that use the same symmetry-adapted polynomial machinery without always constituting a standalone many-body intermolecular potential in the strict MB-PIP sense (Qu et al., 22 Sep 2025, Nandi et al., 2024, List et al., 18 Dec 2025, Li et al., 15 Feb 2026). The common technical thread is the use of Morse-type distance variables, permutational symmetry, compact polynomial representations, and, in several cases, a $\Delta$-machine-learning correction from a lower-cost baseline toward a correlated reference.

1. Scope and nomenclature

The most direct realization of a PBE0+MBD MB-PIP is reported in the linear-alkane study "Gold-Standard" $\Delta$-Machine Learned and Transferable Potential for Linear Alkanes" (Qu et al., 22 Sep 2025). There, PBE0+MBD serves both as a standalone low-level transferable many-body permutationally invariant polynomial potential and as the low-level component in a $\Delta$-machine-learning workflow that elevates the surface to a PNO-LCCSD(T)-F12 benchmark. The model is explicitly many-body, transferable across chain length, and truncated at 4-body terms.

Two adjacent literatures are often grouped with this topic but are formally distinct. The ethanol study "Δ-Machine Learning to Elevate DFT-based Potentials and a Force Field to the CCSD(T) Level Illustrated for Ethanol" constructs a PBE0+MBD PIP potential and a compact PIP correction, but it is a single-molecule global intramolecular PES rather than a many-body intermolecular decomposition (Nandi et al., 2024). The molecular-crystal study "Multimer Embedding for Molecular Crystals Utilizing up to Tetramer Interactions" uses PBE0+MBD in a subtractive multimer embedding framework on top of periodic PBE+MBD, but it does not fit a permutationally invariant polynomial potential at all (List et al., 18 Dec 2025). A further framework paper, "Monomeric machine learning potential for general covalent molecules: linear alkanes as an example," extends MB-PIPNet to covalent systems with monomer-based PIP descriptors and neural networks, yet its reference level is B3LYP/cc-pVDZ rather than PBE0+MBD (Li et al., 15 Feb 2026).

A common misconception is therefore to treat every PBE0+MBD-plus-PIP construction as an MB-PIP in the same sense. The direct alkane model is an explicit atom-based many-body polynomial PES; the ethanol work is a global PIP plus $\Delta$-correction over intramolecular configuration space; the crystal work is a subtractive many-body correction to a periodic baseline; and MB-PIPNet is a monomeric effective-energy model rather than an explicit $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$ expansion.

2. Atom-based many-body polynomial form

In the alkane realization, the total potential is written as

$$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$

with the expansion truncated at 4-body terms (Qu et al., 22 Sep 2025). For a covalent system such as a single alkane molecule, the “bodies” are atoms rather than molecular fragments. The 1-body term is fixed as the sum of isolated atomic energies and is not fit.

Because the systems contain C and H atoms, the fitted classes are organized by chemical composition. For 2-body terms the classes are CC, CH, and HH; for 3-body terms, CCC, CCH, CHH, and HHH; and for 4-body terms, CCCC, CCCH, CCHH, CHHH, and HHHH. Each class has its own coefficient set, although classes with equivalent permutation symmetry may share the same symmetry-adapted PIP basis. The 3-body term is written explicitly as a sum over CCC, CCH, CHH, and HHH contributions in Morse-type variables $y_{ij}$ with permutation symmetries $A_3$ and $A_2B$.

The geometric variables are Morse-like functions of interatomic distance. In the related ethanol PIP formalism the transformed coordinate is written as

$y_{a\beta}=\exp(-r_{a\beta}/\lambda),$

with $\Delta$0 typically $\Delta$1–$\Delta$2 bohr (Nandi et al., 2024). The alkane MB-PIP uses the same Morse-type construction and applies distinct range parameters by body order. The basis functions are also purified: any polynomial that does not vanish when one atom is infinitely far from the others is removed. This makes the many-body terms physically proper and compatible with range switching.

The low-level PBE0+MBD MB-PIP uses a single Morse range parameter of $\Delta$3 bohr for the 2-body part, maximum polynomial power $\Delta$4, and cutoff distance $\Delta$5 bohr, giving $\Delta$6 total 2-body coefficients. For the 3-body part the Morse range parameter is $\Delta$7 bohr, maximum polynomial order $\Delta$8, and a switching function is applied when the maximum distance in a trimer is between $\Delta$9 and $\Delta$0 bohr, with the contribution set to zero beyond $\Delta$1 bohr. For the 4-body part the Morse range parameter is $\Delta$2 bohr, maximum polynomial order $\Delta$3, and switching occurs between $\Delta$4 and $\Delta$5 bohr. The resulting numbers of PIPs are $\Delta$6, $\Delta$7, $\Delta$8, $\Delta$9, $\Delta$0, $\Delta$1, $\Delta$2, $\Delta$3, and $\Delta$4 for CCC, CCH, CHH, HHH, CCCC, CCCH, CCHH, CHHH, and HHHH, respectively, for a total of $\Delta$5 undetermined linear coefficients, including the $\Delta$6 2-body coefficients.

3. Electronic-structure definition and $\Delta$7-machine learning

PBE0+MBD enters the direct alkane MB-PIP as a dispersion-aware low-level reference. The motivation for moving beyond the earlier B3LYP MB-PIP is explicit: the previous model was built from straight B3LYP energies that do not account for dispersion, whereas alkane folding is controlled by intramolecular dispersion and the associated self-solvation effect (Qu et al., 22 Sep 2025). PBE0+MBD energies were computed with FHI-aims using its "intermediate" basis settings. The correlated target for the $\Delta$8-correction is PNO-LCCSD(T)-F12b/AVTZ', with AVTZ' defined as cc-pVTZ for H and aug-cc-pVTZ for C, computed in Molpro 2023.

The correction protocol is the standard $\Delta$9-ML relation

$1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$0

where $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$1 is the low-level PES and $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$2 is fitted to the difference between correlated and low-level electronic energies. In the present application,

$1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$3

The same logic also underlies the ethanol work, in which a full PIP PES fitted to PBE0+MBD energies and gradients is corrected by a compact PIP fit to CCSD(T)-minus-PBE0+MBD energies (Nandi et al., 2024).

The low-level alkane MB-PIP is trained on the same approximately $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$4 configurations of $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$5 used previously for the B3LYP MB-PIP. Its overall fitting RMSE is $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$6, i.e. $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$7, corresponding to $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$8 or $1\mathrm{B}/2\mathrm{B}/3\mathrm{B}/\dots$9 per atom. The $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$0-fit uses $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$1 energy differences for $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$2 only. Those configurations comprise $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$3 structures from NVE molecular dynamics started from $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$4 stationary points on the low-level PES, $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$5 random displacements around the same $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$6 stationary points, and $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$7 configurations randomly selected from the large low-level data set. The $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$8-PIP retains the same MB-PIP form but uses lower maximum orders of $$V(1,\cdots,N)=\sum_{i=1}^N V_{1-b}(i)+\sum_{i>j}^N V_{2-b}(i,j)+\sum_{i>j>k}^N V_{3-b}(i,j,k)+\sum_{i>j>k>l}^N V_{4-b}(i,j,k,l)+\cdots,$$9, $y_{ij}$0, and $y_{ij}$1 for 2-body, 3-body, and 4-body terms, respectively, producing $y_{ij}$2 unknown coefficients. The reported fit quality is $y_{ij}$3 for the $y_{ij}$4-PBE0+MBD data, compared with $y_{ij}$5 for the $y_{ij}$6-B3LYP data. The corresponding energy-difference range is $y_{ij}$7 for PNO-LCCSD(T)-F12 minus PBE0+MBD, versus $y_{ij}$8 for PNO-LCCSD(T)-F12 minus B3LYP, indicating that PBE0+MBD is already a substantially closer low-level surface.

4. Transferability and alkane conformational energetics

The direct validation target for the transferable PBE0+MBD MB-PIP is the hairpin-minus-linear energy difference across linear alkanes from approximately $y_{ij}$9 to $A_3$0, evaluated at PBE0+MBD optimized geometries (Qu et al., 22 Sep 2025). This observable probes the onset of folding driven by intramolecular dispersion. The data show that bare B3LYP MB-PIP is qualitatively wrong, because without dispersion it over-stabilizes the extended conformer over the whole range. By contrast, bare PBE0+MBD MB-PIP already predicts the linear-to-hairpin transition in the correct region, although it somewhat overstabilizes the hairpin. The $A_3$1-corrected PBE0+MBD MB-PIP improves the magnitudes further and reduces the average error scale to about $A_3$2.

Representative values make the trend explicit. For $A_3$3, the benchmark CCSD(T) hairpin-minus-linear energy is $A_3$4, PBE0+MBD gives $A_3$5, and $A_3$6-PBE0+MBD gives $A_3$7. For $A_3$8, the same quantity is $A_3$9 for CCSD(T), $A_2B$0 for PBE0+MBD, and $A_2B$1 after correction. For $A_2B$2, the values are $A_2B$3, $A_2B$4, and $A_2B$5, and for $A_2B$6 they are $A_2B$7, $A_2B$8, and $A_2B$9.

$y_{a\beta}=\exp(-r_{a\beta}/\lambda),$0	CCSD(T)	PBE0+MBD / $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$1-PBE0+MBD
12	1.20	0.92 / 1.10
14	0.37	-0.10 / 0.22
16	-0.42	-1.24 / -0.98
28	-5.39	-6.88 / -5.86

The sign change in the benchmark occurs between $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$2 and $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$3, whereas the bare PBE0+MBD MB-PIP changes sign between $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$4 and $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$5. Across the listed chain lengths, the mean absolute error of bare PBE0+MBD against the benchmark is about $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$6, while the $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$7-corrected PBE0+MBD is about $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$8. The physical conclusion stated in the study is that PBE0+MBD is a much better low-level reference than B3LYP for alkane conformational energetics because dispersion is essential to folding.

5. Related PBE0+MBD polynomial and many-body correction schemes

The ethanol study provides a directly relevant but narrower analogue of PBE0+MBD MB-PIP methodology (Nandi et al., 2024). Its target system is ethanol, not a transferable alkane family, and the model is a single-molecule global PES over intramolecular configuration space rather than a many-body decomposition over monomers or fragments. The low-level PBE0+MBD PES is fitted with maximum polynomial order $y_{a\beta}=\exp(-r_{a\beta}/\lambda),$9, permutation symmetry $\Delta$00, a basis size of $\Delta$01 PIPs, and a training/test split of $\Delta$02 DFT geometries from the MDQM21 ethanol dataset. The $\Delta$03-correction to CCSD(T)-F12a/aug-cc-pVDZ uses only $\Delta$04 training and $\Delta$05 test geometries, maximum polynomial order $\Delta$06, the same permutation symmetry, and a basis of only $\Delta$07 PIPs. This size reduction is a central methodological result: the correction surface is much simpler than the full PES.

For PBE0+MBD specifically, the stationary-point energetics are already good at the baseline level and improve further after correction. Relative to trans, direct PBE0+MBD gives gauche at $\Delta$08, TS1 at $\Delta$09, and TS2 at $\Delta$10; direct CCSD(T) gives $\Delta$11, $\Delta$12, and $\Delta$13; and the corrected PBE0+MBD surface gives $\Delta$14, $\Delta$15, and $\Delta$16. The paper also reports excellent agreement of corrected harmonic frequencies with direct CCSD(T), substantial reduction in gradient errors despite the absence of CCSD(T) gradients in training, and torsional barriers comparable to direct CCSD(T). The result is not a true MB-PIP potential, but it demonstrates that a PBE0+MBD analytic PIP baseline can be elevated toward CCSD(T) quality by a compact polynomial correction.

The molecular-crystal embedding work is many-body in a different sense (List et al., 18 Dec 2025). It starts from periodic PBE+MBD and adds high-minus-low corrections for monomers, dimers, trimers, and tetramers computed at PBE0+MBD:

$\Delta$17

This is not a fitted PIP potential, but it is highly informative for the many-body orders that matter when targeting PBE0+MBD-quality crystal observables. On X23 lattice energies, periodic PBE+MBD alone has an MAE of $\Delta$18 relative to periodic PBE0+MBD; ME3(4 Å) reaches MAE $\Delta$19; and closed-tetramer ME4(5 Å) or ME4(6 Å) gives MAE $\Delta$20. For stresses and optimized cell volumes, trimers are crucial: the stress MAE drops from $\Delta$21 for ME2(4 Å) to $\Delta$22 for ME3(4 Å), and the cell-volume mean absolute relative error drops to $\Delta$23 for ME3(4 Å). For $\Delta$24-point frequencies, the MAE decreases from $\Delta$25 for periodic PBE+MBD to $\Delta$26 for ME3(4 Å). These results identify 3-body corrections as essential for stress, volume, and low-frequency vibrational fidelity, while 4-body corrections slightly improve lattice energies and can be screened aggressively.

6. Conceptual limits and methodological implications

Current work on PBE0+MBD MB-PIP potentials spans a spectrum of related constructions rather than a single fixed architecture. The direct transferable alkane model is an explicit atom-based many-body PIP at the PBE0+MBD level, with a compact $\Delta$27-correction to PNO-LCCSD(T)-F12 (Qu et al., 22 Sep 2025). The ethanol work uses the same analytic PIP machinery but does not perform a many-body decomposition across monomers or clusters (Nandi et al., 2024). The crystal work decomposes high-minus-low corrections into monomer, dimer, trimer, and tetramer pieces, yet the total embedded energy remains a periodic low-level baseline plus finite-cluster corrections rather than a standalone high-level many-body PES (List et al., 18 Dec 2025).

A further conceptual boundary is shown by MB-PIPNet (Li et al., 15 Feb 2026). That framework represents the total energy as a sum of effective monomeric contributions,

$\Delta$28

with species-dependent neural networks acting on monomer self-descriptors and summed pairwise environment descriptors built from PIPs. For linear alkanes, the monomers are $\Delta$29 and $\Delta$30 groups, the reference data are B3LYP/cc-pVDZ single-point energies, and no cutoff radius was applied in constructing the environmental descriptors. The resulting model is therefore an MB-PIP-based framework for covalent molecules, but not a PBE0+MBD MB-PIP and not an explicit many-body expansion in the classic MB-pol sense.

These distinctions matter for interpretation. The alkane results show that PBE0+MBD is a strong physically informed low-level reference for dispersion-dominated conformational energetics, and that the residual to a correlated benchmark can be represented by a much smaller correction model. The ethanol study shows the same economy for a global intramolecular PIP, where a $\Delta$31-PIP low-level surface is corrected by a $\Delta$32-PIP $\Delta$33 surface. The crystal work shows that when the target is periodic PBE0+MBD behavior in molecular crystals, 3-body local corrections are critical for stresses, volumes, and phonons, while selective 4-body refinement is most useful for lattice energies. This suggests that PBE0+MBD MB-PIP development is presently best viewed as a family of dispersion-aware, symmetry-adapted polynomial strategies: some are standalone many-body PESs, some are compact $\Delta$34-corrections to analytic baselines, and some are many-body correction hierarchies that define which interaction orders must be reproduced for specific observables.