Monomeric machine learning potential for general covalent molecules: linear alkanes as an example

Abstract: Machine-learning potentials (MLPs) have become important tools for modern molecular simulations. However, developing models that simultaneously achieve high accuracy and high computational efficiency remains a significant challenge. In this work, we extend the recently proposed MB-PIPNet framework to general covalently bonded molecular systems by combining monomer-based energy decomposition, permutationally invariant polynomial (PIP) descriptors, and neural networks within a fragmentation-based strategy. Within this framework, the total potential energy is represented as a sum of effective monomeric contributions, where PIPs provide compact and chemically motivated descriptions of both monomer internal structures and their local chemical environments. As a proof-of-concept application, we apply the MB-PIPNet framework to linear alkanes, using n-Tetradecane as a representative system, and benchmark its performance against established atomistic machine-learning models. The resulting MB-PIPNet potential accurately reproduces reference ab initio electronic energies and reliably captures key molecular properties, including torsional potential energy profiles, harmonic vibrational frequencies, and vibrational power spectra obtained from molecular dynamics simulations. Importantly, MB-PIPNet demonstrates a substantial advantage in computational efficiency over other MLP models for combined energy and force evaluations. These results establish MB-PIPNet as a scalable and efficient framework for constructing MLPs, providing an additional route for large-scale quantum and classical simulations of complex molecular systems.

• The paper extends MB-PIPNet to covalent molecules by fragmenting n-tetradecane into methyl and methylene monomers represented with permutationally invariant polynomial descriptors and neural networks.
• The monomeric model achieves 16.0 meV test RMSE, substantially outperforming DeePMD at 65.7 meV and approaching MB-PES at 12.5 meV while accurately reproducing torsional scans and most vibrational modes.
• MB-PIPNet evaluates energies and gradients about 7.5× faster than DeePMD and over 5× faster than MB-PES, but broader transferability requires fragmentation schemes for branched, cyclic, and heteroatom-containing molecules.

Overview

This paper extends the MB-PIPNet machine-learning potential (MLP) framework, previously validated only for noncovalent systems such as water and CO₂ clusters and condensed phases, to covalently bonded molecules. The key methodological step is a fragmentation-based scheme that defines chemically meaningful monomeric units within a single covalently bonded molecule. As a proof of concept, the authors construct a potential for n-tetradecane (C₁₄H₃₀) trained on B3LYP/cc-pVDZ data and benchmark it against two alternatives: DeePMD (a standard Behler–Parrinello-type atomistic potential) and MB-PES (an atom-based many-body PIP expansion). The central claims are that monomer-based decomposition yields accuracy comparable to or better than atomistic approaches at substantially lower computational cost, and that the framework is scalable to larger molecular systems.

Methodology

MB-PIPNet decomposes the total energy into a sum over effective monomeric contributions, $E_{\text{total}} = \sum_i E_i$, where each $E_i$ is produced by a species-specific feed-forward neural network acting on a permutationally invariant polynomial (PIP) descriptor $\boldsymbol{D}_i$. The descriptor combines a one-body self term $\boldsymbol{G}_i^{(1)}$ capturing the internal geometry of the monomer with a sum of two-body environmental terms $\boldsymbol{G}_{ij}^{(2)}$ encoding interactions with neighboring monomers, both built from Morse-like variables $p_{ab} = e^{-r_{ab}/\lambda}$ symmetrized over identical atoms.

For C₁₄H₃₀, monomers are methyl (–CH₃) and methylene (–CH₂–) groups, giving A₃B and A₂B one-body symmetry classes and A₆B₂, A₅B₂, and A₄B₂ two-body classes. The PIP bases are truncated to roughly 45 terms for one-body and 200 for two-body descriptors, and two independent networks with an identical 445-15-15-1 fully connected topology (tanh activations, Levenberg–Marquardt optimization, early stopping) map descriptors to monomeric energies. Notably, no cutoff radius is applied to the environmental descriptors here, since the molecule is finite; the authors note that cutoffs can be reintroduced for extended systems.

The comparison models are trained on the same data. MB-PES expresses the energy as an atom-based many-body expansion up to four-body terms, each expanded in purified PIPs with distance-based cutoffs and switching functions; it involves 736 linear coefficients fitted by least squares. DeePMD uses the DeepPot-SE (se_e2_a) descriptor with a 6.0 Å cutoff and standard embedding/fitting network architectures.

The training set derives from 272,532 MD-sampled configurations from prior work, filtered to those within 40,000 cm⁻¹ of the global minimum (and excluding 253 configurations with incorrect atom ordering), yielding 247,211 configurations split 90:10 into training and test sets with nearly identical distributions of energy, end-to-end distance, and RMSD relative to the global minimum.

Accuracy benchmarks

On the test set, MB-PIPNet achieves an RMSE of 16.0 meV and MAE of 8.67 meV, while MB-PES reaches 12.5 meV RMSE and 4.84 meV MAE. Both dramatically outperform DeePMD, which yields 65.7 meV test RMSE — roughly a factor of four worse than MB-PIPNet despite far greater training cost. This is a strong result: the monomeric representation with compact PIP descriptors is markedly more sample-efficient than a generic atomistic descriptor on identical data.

Model	Test RMSE (meV)	Test MAE (meV)
DeePMD	65.7	40.9
MB-PIPNet	16.0	8.67
MB-PES	12.5	4.84

Property-level validation reinforces these numbers. Torsional scans about the terminal C13–C14 bond (methyl rotation) and the C12–C13 bond (ethyl rotation) show excellent agreement with direct B3LYP calculations across the relevant angular ranges, including local minima and transition states. In the high-energy, highly twisted regions of the ethyl rotation profile, MB-PIPNet actually agrees slightly better with the reference than MB-PES, despite MB-PES's lower global RMSE — evidence that the monomer-based representation is more robust for distorted conformations. Harmonic frequency analysis reproduces all low- and mid-frequency modes below ~1600 cm⁻¹ accurately, though MB-PIPNet modestly overestimates the C–H stretching frequencies near 3000 cm⁻¹, which the authors attribute to the limited PIP basis used for the monomeric descriptors. NVE molecular dynamics trajectories (100 trajectories at 5000 cm⁻¹ total energy) are stable and yield power spectra whose C–H stretching and collective low-frequency features qualitatively match the harmonic analysis.

Computational efficiency

The efficiency results are the most consequential quantitative finding. For combined energy and gradient evaluations on 100,000 geometries using a single AMD EPYC core:

Model	Energies + gradients (s)
DeePMD	1792
MB-PIPNet	240
MB-PES	1248

MB-PIPNet is approximately 7.5× faster than DeePMD and more than 5× faster than MB-PES, while simultaneously being several times more accurate than DeePMD and comparable to MB-PES. Because the cost scales linearly with the number of monomers rather than atoms times neighbor counts, this advantage should grow with system size. Combined with the accuracy table, the implication is that MB-PIPNet dominates DeePMD on both axes for this system class — a favorable Pareto position rather than a trade-off.

Limitations and open questions

Several caveats bear directly on the generality of the results. First, the demonstration is limited to a single linear alkane at the B3LYP/cc-pVDZ level; no CCSD(T)-quality reference is used, so absolute chemical accuracy relative to high-level benchmarks remains unverified. Second, the fragmentation scheme — cutting the chain into CH₃ and CH₂ units — is specific to acyclic hydrocarbons, and the authors acknowledge that applying MB-PIPNet to general covalent systems requires a physically reasonable fragmentation choice, which is nontrivial for branched, cyclic, or heteroatom-containing molecules. Third, the overestimation of high-frequency C–H stretches indicates that the current truncated PIP bases do not fully capture intramolecular distortion effects; systematic basis selection is deferred to future work. Fourth, transferability across chain lengths (e.g., training on short alkanes and predicting C₃₀H₆₂) is asserted as a possibility but not demonstrated, nor is the effect of incorporating equivariant architectures. Finally, the vibrational spectra are classical and qualitative; quantitative comparison with experiment would require path-integral or quantum dynamical treatments of nuclear quantum effects and anharmonicity.

Conclusion

This work demonstrates that monomer-based energy decomposition combined with PIP descriptors and small neural networks can be extended from noncovalent to covalently bonded systems via a fragmentation strategy. For n-tetradecane, the resulting potential matches the accuracy of an atom-based many-body PIP fit, substantially outperforms DeePMD, and delivers combined energy/force evaluations roughly five to seven times faster than either alternative. The main unresolved issues are the design of fragmentation schemes beyond linear alkanes, improved PIP bases for high-frequency modes, and empirical validation of size transferability.