REVIEW 4 major objections 5 minor 4 references
A Combined Tight Binding with Machine Learning Potential Model for Magnesium Compounds
T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash
Pith's one-line read This paper shows that replacing the pairwise repulsive term in DFTB with a MACE machine-learned many-body potential improves accuracy for magnesium compounds while retaining explicit electronic structure.
desk verdict A useful, honest extension of DFTB+MLIP to periodic Mg compounds, but the 'near-DFT' selling line outruns the actual transferability evidence. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central mechanism is the hybrid energy decomposition: the DFTB total energy is split into an electronic term (E1 + E2, from band structure and electrostatics) and a repulsive term E0, normally modeled as a pairwise spline. The paper substitutes the pairwise E0 with a MACE potential, an E(3)-equivariant message-passing neural network that captures many-body interactions. MACE is trained on the difference between DFT and DFTB electronic energies and forces, so it learns the missing physics of the repulsive term while leaving the explicit electronic structure intact.
What would settle it
Run the trained DFTB+MACE model on configurations with O–O distances near 1.28 Å, H–H near 0.77 Å, and O–H near 1.06 Å (the minimum of the pair repulsive potentials), comparing energies and forces against DFT; if errors exceed a few hundred meV/Å or the energy curve becomes unphysical, the model fails out-of-distribution. Alternatively, an MD simulation at high temperature that samples short interatomic distances should show whether the model remains stable and accurate.
Extended reading notes
Core claim
The central claim is that the conventional pairwise repulsive term in DFTB is a major source of error, since it cannot describe different coordination environments, and this term can be replaced by a many-body MACE potential trained on the residual between DFT and the DFTB electronic energy (E0 = E_DFT − (E1 + E2)). The resulting DFTB+MACE model achieves near-DFT accuracy for energies and forces in many cases, while still supplying electronic structure information that pure machine-learned potentials cannot. The paper demonstrates this for MgO and Mg(OH)2 systems: lattice constants, bulk moduli, phonon spectra, surface energies, and molecular dynamics forces all improve substantially compare
Load-bearing premise
The training set of 15,396 structures adequately samples the potential-energy surface regions encountered in applications, especially short-range repulsive configurations; the paper itself notes it may not fully represent all regions and that configurations near repulsive minima are poorly represented.
Editorial extensions
If this is right
- DFTB parameterizations can be systematically improved by learning the repulsive term with flexible machine-learned potentials, reducing the need for hand-tuned pair functions.
- The hybrid model retains the ability to describe charge redistribution, which pure machine-learned interatomic potentials lack, making it attractive for studying charge-transfer processes at scale.
- The electronic term can partially compensate for missing long-range interactions in the machine-learned potential, as seen in the bulk-water radial distribution functions.
- The approach may extend to other materials beyond magnesium, provided a sufficiently diverse training set is constructed.
- Simple MACE models with low angular momentum suffice when the electronic term already carries substantial physical information, potentially lowering computational overhead.
Reading between the lines
- A general recipe emerges: any DFTB code could be upgraded by training an MLIP on the DFT−DFTB residual, but success hinges on the machine-learned term being allowed to be many-body rather than a fixed pairwise function.
- The observed failure on out-of-distribution adsorption energies suggests that active learning or on-the-fly sampling of non-equilibrium configurations — especially near the repulsive minima — would be necessary for robust predictions in reactive or high-temperature simulations.
- The accidental agreement of the (110) surface energy with experiment hints that the MLIP might be compensating for errors in the DFTB electronic term in a non-systematic way; careful error decomposition would clarify when the correction is physically meaningful.
- The electronic structure output of the hybrid model could be used as additional features for the machine-learning correction, potentially improving charge-transfer predictions at minimal extra cost.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a hybrid density functional tight binding (DFTB) + MACE model for Mg-based compounds. In this model, the conventional pairwise repulsive potential in DFTB is replaced by a many-body MACE machine learning potential. The MACE component is trained on the difference between DFT energies/forces and the DFTB electronic energies/forces (Eqs. 5–7). The authors train on 15,396 structures (MPtrj, MatPES, and self-generated MgO/Mg(OH)2 configurations) and evaluate on static calculations (cohesive curves, phonons, surface energies, CO2 adsorption) and molecular dynamics (bulk water and water on Mg(OH)2). They compare DFTB+MACE with a restricted MACE (MACE_R), plain DFTB, and DFT. They report improved accuracy over plain DFTB, often with moderate computational overhead, while retaining an explicit electronic-structure description. The paper also acknowledges limitations: the training set may not cover all relevant potential-energy regions, and configurations near repulsive-potential minima are poorly represented.
Significance. If the approach is made robust, it addresses a real limitation of DFTB: the pair-potential approximation for the repulsive term cannot capture many-body effects, and MLIPs can provide a flexible correction. The manuscript combines a well-established electronic-structure method with a state-of-the-art equivariant MLIP, and the authors release their models and training data. The reported force MAEs for moderate and complex hyperparameter settings (82.0 and 66.7 meV/Å, respectively) are competitive with a pure MACE model, and there are clear improvements over plain DFTB in phonons, surface energies, and MD forces. However, the central claim of 'near-DFT accuracy' is not fully supported by the testing protocol, as the paper relies on a random validation split without a held-out test set and reports large errors on key out-of-distribution adsorption cases. The significance is therefore conditional on improved validation and dataset coverage.
major comments (4)
- [Calculation Details; Results and Discussion (Accuracy and complexity analysis)] The dataset is split only into training and validation sets; the text states 'The dataset was not explicitly divided into a separate test set. Instead, the calculation cases... were used as test set.' Hyperparameters are selected from 419 runs on the same validation split, creating a risk of selection bias. Moreover, several application cases (random-displaced fcc/bcc-MgO supercells, CO2/H2O on MgO clusters) are structurally close to the self-generated training structures, so they do not independently assess transferability. The one case the authors identify as outside the training set, CO2/MgO(110), has an adsorption energy error of about 0.75 eV (Table 2: DFTB+MACE -3.34 eV vs DFT -2.59 eV), and the (100) case error is about 1.36 eV. This does not support the Introduction's 'near-DFT accuracy' claim. The authors should add a formal held-out test set and report performance before hyperp
- [Molecular Dynamics (and Abstract)] The paper itself states that configurations near the repulsive-potential minima (O–O 1.28 Å, H–H 0.77 Å, O–H 1.06 Å) are 'poorly represented in the current training set.' Since the MACE term replaces the pair potential, these short-range regions are exactly where the learned correction must be reliable. The observed trapping in local minima during geometry optimization is a direct consequence of this under-coverage. The authors should quantify the distribution of interatomic distances in the training set relative to these minima and discuss strategies (e.g., active learning or adding repulsive-wall configurations) to ensure the model is not extrapolating in the most important short-range regime.
- [Table 2 and Static Calculations] The adsorption energies for CO2/MgO show errors of ~1 eV (DFTB+MACE: -1.43 eV for (100) vs DFT -0.07 eV; -3.34 eV for (110) vs DFT -2.59 eV). The text acknowledges this and describes the result as 'partly demonstrates transferability,' but the magnitude of the error is comparable to the energy differences that would be needed to predict adsorption preferences or reaction pathways. The paper should temper the 'near-DFT accuracy' statement in the Introduction and the Abstract's 'improved accuracy' claim should be made more precise: the improvement is systematic but the absolute accuracy for adsorption energies is still far from DFT. A discussion of how this could be remedied (e.g., fine-tuning on surface/adsorption configurations) is already present but should be integrated into the conclusions.
- [Accuracy and complexity analysis (Figure 2 and text)] The comparison between DFTB+MACE and MACE_R is informative, but the paper does not decompose the remaining error into contributions from the DFTB electronic part and from MACE. The text notes that 'when a more complex set of hyperparameters is used, the error in the DFTB electronic term begins to be apparent.' A quantitative separation of the error sources (e.g., by comparing E(1)+E(2) errors before adding MACE) would help the reader understand the limitations and guide future parameterization. This is not strictly required for the main claim, but it would strengthen the analysis.
minor comments (5)
- [Abstract and Introduction] The Abstract says 'improved accuracy relative to DFTB with a pair potential, in many cases with only a moderate increase in computational cost' — this is appropriate. However, the Introduction's 'near-DFT accuracy' is too strong given the adsorption errors and should be revised.
- [Table 1 and text] Typographical errors: 'tabel' should be 'table'; 'Quardro' should be 'Quadro'; 'dFT' appears as 'PBE-dFT' in Section 3; 'rdf' is used inconsistently with 'RDF'; 'iirps' in Section 2 should be 'irreps'.
- [Figure references] The text refers to 'figures 5-a1 and 5-a2 in the SI' and 'figure 5-b1 and 5-b2' without clear SI numbering. Please ensure SI figure labels are consistent and correctly referenced.
- [Methodology] The paper should provide more details on how the DFTB electronic energy (E(1)+E(2)) is computed for the training set, including the specific DFTB parameter set used (from reference [4]) and how the MACE correction is applied to forces and stresses (Eqs. 6–7 are stated but not derived in detail).
- [Data availability] The GitHub link and training-set availability are appreciated. It would be helpful to also provide the hyperparameter-scan results (all 419 settings) in machine-readable form, not only in the SI.
Circularity Check
No significant circularity: the residual-learning target is definitional by construction, but the accuracy claims are tested on held-out application cases, so no load-bearing step reduces to its inputs.
full rationale
The MACE correction is defined by Eqs. 5-7 as E0 = E_DFT - (E1+E2), so the sum E1+E2+E0 equals E_DFT on configurations exactly represented by the training data; that is the method's definition, not a prediction. The paper's quantitative claims rest on the held-out validation split (1535 of 15,396 configurations) and on application cases (phonons, surface energies, adsorption, water MD) not used to fit the MACE parameters. The validation MAE is an interpolation-quality estimate because hyperparameters were explored and the same dataset family is used, and the paper explicitly states no separate test set was created; this is a methodological caveat. Some application cases resemble the self-generated fcc/bcc MgO training structures, and the genuinely out-of-distribution CO2/MgO(110) case shows ~0.75-1.36 eV adsorption-energy error, which the paper reports as due to insufficient training coverage. These are accuracy/transferability limitations, not circularity. Self-citations (Refs 4, 28) supply the DFTB parameter set and prior observations but are not invoked to prove the central claim; no uniqueness theorem or ansatz is imported via self-citation.
Assumptions & free parameters
free parameters (3)
- MACE network weights =
Not enumerated; trained on 13,861 configurations
- MACE hyperparameters (moderate model: k, T, nu, Lmax, lmax) =
32, 2, 2, 1, 2
- DFTB parameter set for Mg, C, O, H (orbital radii, hopping integrals, Hubbard parameters) =
From prior work (ref 4)
assumptions (4)
- domain assumption DFTB total energy is accurate to second order in density fluctuation; higher-order terms can be absorbed into the fitted residual (Eqs. 1-4).
- domain assumption PBE-DFT is an adequate ground truth for training and evaluation.
- domain assumption The training set covers the relevant potential-energy surface, including short-range repulsive regions.
- domain assumption MACE's E(3)-equivariant architecture can represent the missing many-body repulsive term sufficiently well.
Cite this review
Pith. "Pith review of A Combined Tight Binding with Machine Learning Potential Model for Magnesium Compounds." pith.science (2026). https://pith.science/paper/AFZH37EL
@misc{pith2026260625853,
author = {Pith},
title = {Pith review of: A Combined Tight Binding with Machine Learning Potential Model for Magnesium Compounds},
year = {2026},
howpublished = {\url{https://pith.science/paper/AFZH37EL}},
note = {Machine review of arXiv:2606.25853}
}
read the original abstract
We present a model for magnesium-based systems that combines density functional tight binding (DFTB) with MACE, a machine learning interatomic potential (DFTB+MACE). In this model, the conventional repulsive potential, pair potential, is replaced by a many-body MACE potential. The MACE component of the model is trained on the difference between density functional theory (DFT) energies and forces and the corresponding DFTB values, but neglecting the pair potential contribution. Using this model we performed structural relaxation of MgO-CO2 adsorption systems, molecular dynamics calculations of water clusters and phonon spectrum calculations of stable fcc-MgO and metastable bcc-MgO structures. We compare the performance of our model with a pure MACE model and with DFT. We demonstrate that the DFTB+MACE model achieves improved accuracy relative to DFTB with a pair potential, in many cases with only a moderate increase in computational cost. In addition, it can provide electronic structures that most of the machine learning potentials cannot. The training dataset, originally developed for MACE, may not fully represent all regions of the potential surface we may encounter during simulations. Expanding the dataset for a wider potential surface is expected to further enhance predictive accuracy of DFTB+MACE model. Overall, the resulting DFTB+MACE framework enables simulations at length and time scales beyond the reach of first-principles methods while retaining an explicit description of electronic structures, making it particularly attractive for studying charge-transfer in materials.
Figures
Figures from the paper (5 more)
Reference graph
Works this paper leans on
-
[1]
Construction of tight-binding-like potentials on the basis of density-functional theory: Application to carbon.Physical Review B1995,51, 12947
(1) Porezag, D.; Frauenheim, T.; Köhler, T.; Seifert, G.; Kaschner, R. Construction of tight-binding-like potentials on the basis of density-functional theory: Application to carbon.Physical Review B1995,51, 12947. (2) Seifert, G.; Porezag, D.; Frauenheim, T. Calculations of molecules, clusters, and solids with a simplified LCAO-DFT-LDA scheme.Internation...
1996
-
[3641]
31 (33) Krack, M. Pseudopotentials for H to Kr optimized for gradient-corrected exchange- correlation functionals.Theoretical Chemistry Accounts2005,114, 145–152. (34) Deng, B.; Peichen, Z.; Jun, K.; Janosh, R.; Han, K.; Bartel, C.; Gerbrand, C. Materi- als project trajectory (MPtrj) dataset.Figshare. https://doi. org/10.6084/m9. figshare 2023,23713842, v...
arXiv 2023
-
[3730]
M.; Tavernelli, I.; Rothlisberger, U
32 (42) Doemer, M.; Liberatore, E.; Knaup, J. M.; Tavernelli, I.; Rothlisberger, U. In situ parameterisation of SCC-DFTB repulsive potentials by iterative Boltzmann inversion. Molecular Physics2013,111, 3595–3607. (43) Goyal, P.; Qian, H.-J.; Irle, S.; Lu, X.; Roston, D.; Mori, T.; Elstner, M.; Cui, Q. Molecular simulation of water and hydration effects i...
2020
-
[5641]
(8) Van den Bossche, M.; Gronbeck, H.; Hammer, B. Tight-binding approximation- enhanced global optimization.Journal of chemical theory and computation2018,14, 2797–2807. (9) Bodrog, Z.; Aradi, B.; Frauenheim, T. Automated repulsive parametrization for the DFTB method.Journal of chemical theory and computation2011,7, 2654–2664. (10) Hellström, M.; Jorner, ...
arXiv 2010
Reviewed August 2, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.