acceptodds
Under review as a conference paper at ICLR 2027

Learning Reduced-Order Hamiltonians from Quantum Transport: A Case Study with DNA

Abstract

Deep learning methods have been applied to atomic-level structure prediction and tight-binding Hamiltonian prediction, often using graph encoding of an atomic structure as an input. In this work, graph neural networks are used to construct a model Hamiltonian from an input graph that represented the interconnected bases in double-stranded DNA, with multiple cascaded layers connecting to a reduced-order model Hamiltonian. This Hamiltonian is used to calculate the quantum transmission and density of states using the non-equilibrium Green's function (NEGF) approach. The model is trained on density of states and transmission spectra collected from ab initio data from a dataset of 2,077 transmission and DOS spectra from 520 DNA sequences. Notably, no ab-initio Hamiltonian or orbital energies entered the model or training approach; only quantum transport data was used for model supervision. The Hamiltonian model was compared to a neural-network-only direct model, which pools node embeddings and feeds them into a multi-layer perceptron to construct the transmission and DOS outputs. Four architecture parameters–number of graph layers, assigned orbitals per base, supervision method, and geometry embeddings–were subjected to a full factorial analysis of variance (ANOVA) test against seven metrics to determine significance and optimize model architecture, ablating terms that did not significantly impact the model. For in-distribution data, the validation set of 4-8 base-pair (bp) duplexes, the best Hamiltonian checkpoint reaches a 1.9x higher loss than the best direct checkpoint (0.33 vs 0.17 decades of Huber loss). For out-of-distribution data, a set of four 12-bp duplexes and four 16-bp duplexes, the best Hamiltonian checkpoint beats the best direct checkpoint in each case by 0.6-2.3 decades. This chosen Hamiltonian model also reproduces the resonant transmission bands in repeating duplexes (poly-G and poly-AT), showing promise for rapid screening of other DNA variants with similar properties.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
graph neural networks, DNA, Hamiltonian, differentiable physics, quantum transport, density functional theory

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zypoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a reduced-order, tight-binding-like Hamiltonian model for DNA charge transport. DNA duplexes are represented as graphs, with nodes for bases/contacts and edges for backbone, hydrogen-bond, and contact interactions. A graph attention network reads the graph and outputs the onsite and coupling blocks of a small Hamiltonian, with 1–2 orbitals per base. A differentiable NEGF layer then converts this learned Hamiltonian into transmission T(E)T\left(E\right), density of states, and local density of states. The model is trained only on DFT+NEGF transport observables, not directly on an ab initio Hamiltonian.

The headline claims are that the direct model is better in distribution, while the Hamiltonian model extrapolates better on longer held-out transmission spectra, and that two orbitals per base, four GNN layers, and LDOS+T supervision are preferred. The main contribution is therefore not a new ML architecture per se, but the combination of a reduced-order Hamiltonian, differentiable NEGF supervision, and DNA transport data. Validation is primarily through spectral fitting, ANOVA-selected design choices, and limited length extrapolation.

Critical strengths
  1. The differentiable NEGF layer is a conceptually useful inverse-modeling setup. The paper’s main idea is to learn a Hamiltonian only through transport observables using a differentiable NEGF readout. This is scientifically interesting because it treats Hamiltonian construction as an inverse problem constrained by physical response functions. If it worked robustly, it could provide a compact, interpretable surrogate for expensive DFT+NEGF calculations and enable reasoning about onsite energies, couplings, and contact effects. This is consequential for the review because it gives the paper a genuine physics-informed ML contribution rather than just another spectral regression model.

  2. The dataset and experimental design are unusually systematic for a domain paper. The authors generate a nontrivial DFT+NEGF dataset, run a full-factorial ANOVA over several architecture and supervision choices, use multiple seeds, and release code/data/checkpoints. The ANOVA helps expose trade-offs such as the benefit of two orbitals, the cost of four-layer message passing in terms of contact drift, and the irrelevance of geometry features for idealized B-DNA. This is consequential because reproducibility and careful ablation are strengths at ICLR, and the dataset itself may be useful to the DNA transport community even if the modeling claims are not fully convincing.

Critical weaknesses
  1. The central motivation, which is a reduced-order Hamiltonian can faithfully describe DNA charge transport, is not established. The paper claims to produce a physically accurate reduced-order Hamiltonian that simplifies modeling of large DNA molecules. However, it never validates the learned Hamiltonian against a projection or downfolding of the DFT Hamiltonian, even though the dataset contains Fock and overlap matrices. It also does not show that 1–2 orbitals per base can capture the relevant physics. Prior work cited by the authors indicates that DNA transport involves many delocalized occupied orbitals rather than a simple frontier-orbital picture. The learned model cannot reproduce sharp localized LDOS peaks in heterogeneous sequences, and the learned onsite energies drift by a median of about 0.39 eV depending on the contact configuration. That means the “Hamiltonian” is not a clean, transferable molecular Hamiltonian. It is contact-dependent and effectively tuned to reproduce selected observables. This is consequential because it undermines the paper’s main scientific claim. Without evidence that the reduced Hamiltonian is physically meaningful or systematically connected to the DFT Hamiltonian, the model is better described as a physics-constrained spectral predictor, not a validated reduced-order electronic structure model.

  2. The empirical evidence for the key out-of-distribution advantage is weak and partly selected on the test set. The final seed and architecture were selected using the 12-bp and 16-bp held-out losses, and the ANOVA also considered those same out-of-distribution metrics. The held-out set contains only 8 records. The Hamiltonian model wins on held-out transmission, but the direct model is equal or better on held-out DOS. The abstract’s claim that the Hamiltonian model wins “in each case” is therefore too broad. There are also inconsistencies: the in-distribution gap is reported as 1.9× in the abstract and 1.7× in the main text, and the final model uses LDOS+T supervision even though Table 3 suggests DOS+T is the best main effect for several transmission metrics. These issues are consequential because the main reason to prefer the Hamiltonian model over the direct baseline is its supposed length extrapolation. If that advantage is selected on the test set, based on only 8 records, and limited to transmission while DOS behaves differently, then the central empirical claim is not robust.

Questions for authors
  1. Can you justify the reduced-order Hamiltonian physically? Since the dataset includes Fock and overlap matrices, can you compare the learned Hamiltonian to a projected/downfolded DFT Hamiltonian? Without such a comparison, what evidence supports calling the learned matrix a “physically accurate” reduced-order Hamiltonian rather than a latent variable tuned to fit spectra?
  2. How should contact drift be interpreted? The learned onsite energies shift by a median of ~0.39 eV across contact configurations. Is this intended as a physical contact perturbation, or is it an artifact of message passing from contact nodes? Can you enforce a contact-independent Hamiltonian and still fit the transport observables?
  3. What happens on a truly untouched test set? The 12-bp and 16-bp records appear to have been used for seed and architecture selection. Can you lock the architecture and seed before evaluation and report results on a genuinely held-out set? How many independent sequences and contact configurations does that set contain?
  4. Why does the final model use LDOS+T when the ANOVA tables suggest DOS+T is often best? Table 3 lists DOS+T as the best supervision main effect for several transmission metrics, yet the final model uses LDOS+T. Can you clarify the selection rule and whether the reported final model is optimal under a pre-registered criterion rather than post hoc choice?
  5. Is the Hamiltonian advantage due to physics or to the loss and readout? The direct model captures peaks better but loses on off-peak and length-dependent values. Did you try a length-aware or peak-weighted loss for the direct model, or a length-conditioned direct baseline? If the direct model matches or beats the Hamiltonian model under such adjustments, what remains of the claimed benefit of the NEGF-constrained architecture?
1 comment

Thank you for your thorough review of our work, it was helpful and inspired substantial changes described in the general response and improved clarity of which design parameters are important. Contact drift is gone by construction, [Sec. 2.2] and "physically accurate" has been changed to "an effective Hamiltonian whose base-resolved levels absorb the couplings of hundreds of basis functions". [Sec. 4] The new Hamiltonian model architecture does capture sharp LDOS peaks in heterogeneous sequences. [Sec. 4; Figs. 6-7] However, the direct model also improved substantially, and the Hamiltonian model is now lower only on the poly-G repeats. [Abstract; Sec. 4; Fig. 4] The 1.9x and 1.7x were two different quantities (selected pair versus the mean over all runs) and are now labelled. [Sec. 4] Every held-out duplex has all four contact configurations, so the set is 32 records ("8 records" in Section 2.1 was an error). [Sec. 2.1]

To address your questions directly:

  1. While yes, we can justify the physicality of our Hamiltonian matrix with the LDOS comparison, no, we do not train or validate against downfolded matrices by design. In our previous theoretical works applying NEGF to large DFT matrices, the transmission patterns could not be explained by selecting a handful of energy levels. As an alternative approach, we were interested in a characteristic Hamiltonian that fit to the spectral weights (DOS and LDOS) to reproduce the measurable quantity (transmission). With the architecture changes, the fitting of energy levels to the local DOS peaks has improved substantially, but your point remains: this is an effective Hamiltonian that is not unique and does not match up with the DFT framework. We have added references to the work of Cignoni et al. (given by Reviewer 2), who make the same argument that a projected Hamiltonian is not a unique training target and likewise construct an effective one. [Sec. 1]

  2. In the submission, we had hypothesized that contact drift was an additional fitting parameter that could represent orbital fluctuations caused by the application of the self-energy. This was disproven in the updated model: by eliminating the contacts in the input graph, the contact fluctuation was exactly zero and therefore was replaced with LDOS loss as an ANOVA metric. [Sec. 2.4; Table 2] With DOS+T supervision the contact-free model fits validation about as well (0.361 vs 0.356) and extrapolates better, so it is used in the updated submission.

  3. Our selection criteria changed in the updated submission, using only the validation transmission loss for selection, and we have generated a new testing dataset that is untouched during the ANOVA analysis for final verification, scored once under a rule written down in advance. This set contains randomly generated 16-bp duplexes (8 random sequences x 4 contact configurations, 32 records). [Sec. 2.1; App. A.3, Table 5] On it the length-aware direct model has the lower transmission loss on every duplex (mean 0.90 vs 1.36), while every Hamiltonian configuration beats every length-blind (mean-pooled) direct configuration.

  4. In the updated ANOVA, LDOS+DOS+T had a lower mean loss across all transmission metrics, but DOS+T with 2 orbitals per base was the best single configuration (main effects average over the other factors) [Sec. 3.1.1; Table 3]. We chose it by a rule written down before any held-out or untouched result existed: lowest 3-seed-mean validation transmission loss, then the seed with the lowest validation transmission loss. [Sec. 2.4; Sec. 3.2]

  5. The benefit of the Hamiltonian model is that the base dependence is explicit and measurable: this information cannot come out of the direct model. This is especially helpful when looking for resonant effects that can be exploited for molecular electronics. With a length-aware pooling method (sum pooling) added to the direct model, the Hamiltonian model losses were higher everywhere with the exception of transmission loss on the poly-G sequences. The fitting actually flipped: now the Hamiltonian model fits transmission peaks well and only drifts away from the ground truth at off resonant values, while the direct model fits across all energies. [Sec. 4; Fig. 4] We did not try a peak-weighted loss, though fitting to Lorentzians was mentioned as a future approach. [Sec. 4, future work]

We hope the contact-independent Hamiltonian, the level placement against the DFT LDOS and the untouched test set address your concerns, and that you will reconsider your score.

Reviewer_VQScOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper learns reduced-order Hamiltonians for DNA charge transport from transmission and density-of-states spectra. A graph attention network predicts a small Hermitian Hamiltonian, which a differentiable NEGF calculation maps to the supervised spectra. The authors provide 2,077 records from 520 short DNA sequences and evaluate length extrapolation on eight longer sequences. A factorial study examines supervision, orbital count, network depth, and geometry features. The selected Hamiltonian model performs worse on short-sequence transmission prediction but reports lower transmission losses on longer sequences than a direct spectral predictor.

Critical strengths

Learning a compact, base-level Hamiltonian from observable spectra is an interesting application. The NEGF decoder gives the learned parameters a structured connection to transport predictions.

The released dataset, underlying Fock and overlap matrices, and model checkpoints are useful resources. The factorial study with three initialization seeds systematically examines several architectural choices.

Critical weaknesses

The extrapolation claim lacks an independent evaluation. Section 3.2 uses extrapolation performance to select configurations and chooses the final seed by summing transmission losses on the validation, 12-bp, and 16-bp sets. The same eight longer sequences then support the generalization claim. Table 5 also shows that the 12-bp advantage reverses in the deeper-readout and wider-network configurations, although the 16-bp advantage remains. These findings warrant an independent test with a predetermined selection protocol and results across seeds; they do not establish the robustness implied by the selected-model results.

The additional value over relevant alternatives is insufficiently established. The only competing predictor is an in-house GNN with mean pooling and MLP readouts. The selected Hamiltonian model receives LDOS plus transmission supervision, whereas the direct model receives DOS plus transmission supervision, leaving the benefit of additional local supervision unresolved. Relevant prior work includes SchNet+H by Westermayr and Maurer [1], Electronic Excited States from Physically Constrained Machine Learning by Cignoni et al. [2], AD-NEGF by Zhouyin et al. [3], and Learning Hamiltonians for Solid-State Quantum Simulators by Pawłowski and Krawczyk [4]. The first two are not discussed, and the latter two are only briefly distinguished. Their inputs and tasks differ, but this makes careful positioning and suitably adapted comparisons necessary. The current evidence does not adequately separate the contribution of the physical decoder from supervision and architecture choices.

References

[1] Julia Westermayr and Reinhard J. Maurer. Physically inspired deep learning of molecular excitations and photoemission spectra. Chemical Science, 2021.

[2] Edoardo Cignoni, Divya Suman, et al. Electronic Excited States from Physically Constrained Machine Learning. ACS Central Science, 2024.

[3] Zhanghao Zhouyin et al. Automatic differentiable nonequilibrium Green’s function formalism: An end-to-end differentiable quantum transport simulator. Physical Review B 108, 195143, 2023.

[4] Jarosław Pawłowski and Mateusz Krawczyk. Learning Hamiltonians for solid-state quantum simulators. arXiv:2603.02889, 2026.

Questions for authors

Can the authors evaluate on new longer sequences excluded from all configuration and seed selection, reporting mean performance and variability across seeds?

Does the advantage persist against a direct predictor with explicit length information and an auxiliary LDOS head, or a simpler learned tight-binding parameterization trained with the same spectral supervision?

Can a reduced Hamiltonian predicted for one contact configuration reproduce the reference spectra of another configuration when only the NEGF contact self-energies are changed and the Hamiltonian is held fixed? This would directly test whether the learned Hamiltonian is reusable across contacts, an important aspect of its claimed physical interpretation.

1 comment

Your thorough analysis of our submitted paper is appreciated, thank you for taking the time to review our paper and for giving us a chance to improve this work. We have made substantial changes to the model architecture of both models based on your feedback, those are described in the general response. The two additional papers you cited would have enhanced our initial submission and have since been added to improve positioning. [Sec. 1] The 12-bp reversal you noticed is confirmed by the full design: against the length-aware direct model the Hamiltonian model is not lower on mixed sequences at either length, and the abstract no longer claims it. [Abstract; Sec. 4] The capacity probe was rerun on the new pair across seeds (capacity does not fix extrapolation [Sec. 3.2; App. A.4, Table 6]), and the untouched-set and capacity results now report seed-to-seed variability [App. A.3; App. A.4, Table 6]. Supervision is now matched: the selected Hamiltonian is trained on DOS+T, the direct model used for every DOS comparison is trained on DOS+T, and neither sees per-base data. [Sec. 3.2] The GNN encoder is identical in both families, so the remaining difference is the readout (NEGF versus MLP).

To respond to your questions directly:

  1. Yes we can: we have added an untouched 16-bp dataset (8 random sequences x 4 contact configurations, 32 records) as a final verification step, scored once under a rule written down in advance (one draw failed in DFT and was replaced by the next generator draw). On it the direct model has the lower transmission loss on every duplex, mean 0.90 vs 1.36, and the ordering holds at every seed. [App. A.3, Table 5] Additional figures added to the appendix [Figs. 6-7] show the Hamiltonian model locating every major LDOS peak in the chosen model, which was selected only on validation data.

  2. No, a tight-binding method was not compatible with this framework: in early development a fixed per-base tight-binding parameterization was tried, but led to onsite energies outside the 1 eV window and removed any interactions of neighboring bases that are well described by a GNN. The pooling method was modified to incorporate length awareness: the direct model ANOVA incorporated sum versus mean pooling as a variable, and the sum-pooled model is the one all comparisons now use. [Sec. 2.4; Sec. 3.2; Sec. 4] We did not add an auxiliary LDOS head: on the direct model it would be an unconstrained per-node regression with no tie to the DOS or the transmission, whereas in the Hamiltonian model all three come from one Green's function.

  3. Yes, and by modifying the input architecture of the Hamiltonian model (pruning the contact nodes and edges) we have eliminated contact dependence. [Sec. 2.2; Table 1] This has substantially improved the energy level fitting: the learned levels now line up with the DFT per-base LDOS peaks [Sec. 4; Figs. 4, 6, 7], which the submitted models mostly did not, at no cost in validation fit. As a final check, we were able to divide the contact losses by contact coupling and position: for 0.1 eV same / 0.1 eV cross / 0.6 eV same / 0.6 eV cross) the respective losses were 0.345 / 0.361 / 0.270 / 0.305. The Hamiltonian itself is identical across the four configurations and only the self-energies change, so this spread is the NEGF response to the contacts, not drift in the Hamiltonian. By comparison the submitted model drifted by more than 1 eV on some elements.

We hope the independent test set, the matched supervision and the contact-independent Hamiltonian address your soundness concerns, and that you will reconsider your score.

Reviewer_9VSsOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a novel method to learn quantum transport properties of DNA. The density of states and transmission spectra are predicted using the DNA sequence as input. Instead of predicting these properties directly, a graph neural network architecture is used to predict a matrix that is interpreted as a kind of tight-binding Hamiltonian and used to compute the downstream properties. This method is then compared to using the same architecture but trained to predict the evaluated properties directly. A large set of ablation experiments is performed to determine the optimal training settings. The models are evaluated in distribution on small DNA chains but also the generalization to longer chains is investigated.

Critical strengths

The evaluations done to decide what the best model configuration is seem very thorough.

Critical weaknesses

The proposed method does not seem to consistently improve on the introduced baseline. This seems to only be the case on the generalization to longer chains task. However also in this setting peaks are captured better by the baseline model. Evaluating the proposed method is especially difficult, because other than the newly introduced baseline no comparison to already existing methods is provided.

The problem that is solved here seems fairly specialized and the impact for the broader ICLR community may be limited.

Questions for authors
  • Since the density of states is an extensive property, could one try a per node readout for the direct property prediction model, summing only in the end?
  • It seems like the models are chose based on validation metrics and the evaluations are performed on the same validation set. Could a separate test set be used?
  • Could any direct comparisons to literature results be added?
3 comments

Thank you for your thorough analysis and constructive review. The three main changes (validation-only selection with a new untouched test set, a contact-free Hamiltonian model, and a length-aware direct model) are summarized in the general response. Agreed: against the new length-aware direct model, the Hamiltonian model has the lower transmission loss only on the poly-G repeats, and the abstract no longer claims a general held-out win. [Abstract; Sec. 4] The method has application in molecular-electronics design, MD-trajectory analysis and inverse design, all of which need a Hamiltonian that can be produced compactly and efficiently by our method. [Sec. 4] We have also released the full dataset, all checkpoints and the untouched 16-bp set as a benchmark.

To address your questions:

  1. Yes, and this has become the single largest change in the revision. While this was initially eliminated to have a baseline that was not dependent on geometry, the resulting trained models improved so much that we have added the pooling method (mean vs sum) as a parameter to the ANOVA analysis. Interestingly enough, sum pooling lowered the transmission losses, while mean pooling lowered the DOS loss. Since the final models were chosen on validation transmission loss (updated in this version), the selected direct model is sum-pooled; for the same configuration, 16-bp transmission loss fell from 3.29 to 1.17 (3-seed means). [Sec. 2.4; Sec. 3.1.3; Sec. 4; Table 4] Following your extensivity point, an exploratory model with mean pooling for the DOS head and sum pooling for the transmission head is the best direct model on both (untouched-set transmission 0.72, DOS 0.076). [Sec. 4; App. A.3, Table 5]

  2. Yes, a separate dataset is used in the updated submission. To be clear, the validation set is separate from the training set. However, in the previous submission, the model selection used the 12- and 16-bp held out losses for final selection. With the changes to the direct model listed above, and the removal of contacts from the input graphs in the Hamiltonian model, we had fresh results to redo selection and we used the transmission validation losses (in distribution) as the only criterion for selection. [Sec. 2.4; Sec. 3.2; Table 1] Additionally, we included an untouched 16-bp dataset (8 random sequences x 4 contact configurations, 32 records) as a final verification step, scored once under a rule written down in advance; the 12- and 16-bp held-out losses stay in the ANOVA only as report-only extrapolation responses. On the untouched set the direct model has the lower transmission loss on every duplex (mean 0.90 vs 1.36). [App. A.3, Table 5]

  3. No, because no published learned model takes a DNA sequence to a transport spectrum or transport Hamiltonian and therefore no published literature number is directly comparable. Reviewer 2 did bring up two works that we had not considered and are closely related to this work, but their parameterizations are based on orbital energy matching or basis fitting from atomic coordinates, so a like-for-like number would require changing their task. While this doesn’t directly address this point, we did replace the contact invariance metric (now zero due to the architecture change) with per-base LDOS loss as an ANOVA response, which gives an additional comparison to the DFT spectrum. [Sec. 2.4; Table 2]

We hope the length-aware baseline, validation-only selection and untouched test set address your soundness concerns, and that you will reconsider your score.

Reviewer_9VSsOpenReview ↗

Response to rebuttal

Thank you for your detailed response and substantial revision of the original submission.

The addition of the new 16-bp test set satisfies my concern about selection of the best model using results obtained from the same set that is used in the final evaluation. However, the additional results obtained using the sum-pooling direct model seem to weaken the method proposed here, as the baseline now seems to consistently outperform the introduced method. Additionally, it is not clear to me why a comparison to the literature is not possible, even if models are differently parametrized. Could one not report on the same targets as Korol & Segal (2019) for example?

Overall I will keep my score, as the introduced method does not consistently outperform the baseline and concerns about the relevance to the broader ICLR audience remain.

We agree that the direct model is a better spectral predictor and we are up front about that in the abstract: the transmission properties themselves are better predicted by a neural-network only direct method, which extrapolate better on longer duplexes with the exception of the poly-G repeats. This result in itself is interesting because it shows that the reduced-order Hamiltonian loses information that could otherwise be parameterized in the MLPs, and length awareness is all that was necessary for the NN-only method to extrapolate to larger systems. This is what makes it relevant: we have shown that the Hamiltonian operator can be learned through the differentiable NEGF layer, with a penalty created by the constraints of the Hamiltonian reduction.

The motivation for producing the direct model was to have a comparable model to the Hamiltonian case because none of the prior works provided a dataset or architecture matching our task. For Korol & Segal, the training conductances were generated from a fixed literature tight-binding parameterization, so comparing against their reported errors would compare datasets rather than methods. The closest like-for-like test is training on tight-binding-generated data, where our Hamiltonian model recovers the generating parameters (Demo.ipynb).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Reviewer Response Summary

Thank you all for your thorough review of this work. Each reviewer independently touched on the same three issues: selection involving the “held-out data”, the contact dependence of the Hamiltonian model, and a baseline (the direct model) that was blind to length. In response to each of these issues, we have made the following architecture changes:

  • Changing the selection criteria and adding a new testing data set
  • Eliminating contact dependence in the Hamiltonian model
  • Exploring sum pooling in the direct model ANOVA

These are described in detail:

Selection criteria was modified to only use the transmission loss across the validation data. Since the “held-out” data was used in the ANOVA to quantify the extrapolation ability of the model, a new untouched dataset of 8 random 16-bp duplexes (32 records) was computed and scored once.

The Hamiltonian model architecture was modified to remove contact nodes and edges from the input graph. Since the contact data is used in the NEGF framework already, this did not remove any information from the model and ensured that the generated Hamiltonian was independent of the contact location and coupling. Since this makes the “contact variance” zero for all cases, this ANOVA metric was replaced with the per-base LDOS loss. As an added benefit, this model created Hamiltonian onsite energies that were better fitted to the LDOS peaks across all duplexes.

In the direct model, sum pooling was added as an option and was compared against mean pooling across all metrics in the direct model ANOVA analysis. Sum pooling was found to lead to lower extrapolation losses for the transmission, with the off-resonant transmission values matching the DFT dataset closely, while mean pooling was found to reduce DOS losses in the held-out data. As a final exploratory test, applying a separate head for each model (mean pool for DOS, sum pool for transmission) synergistically lowered losses across DOS and transmission. This was demonstrated in the untouched dataset: for mean pool / sum pool / split pool the transmission losses were 3.71/0.73/0.72 and DOS losses were 0.084/0.101/0.076.

We think the revision is a more grounded and more useful paper than the submission, and we are grateful for the push. All changes are marked in red in the revised PDF; OpenReview also provides a diff against the submission.

Summary of revisions:

  • Model selection on validation data only: Sec. 2.4, Sec. 3.2, Table 1.
  • Untouched 16-bp test set (8 duplexes, 32 records): Sec. 2.1, Sec. 4, App. A.3 (Table 5, Figs. 6-7).
  • Contact-free Hamiltonian model: Sec. 2.2, Table 1; per-base LDOS loss replaces contact variance as a response: Sec. 2.4, Table 2.
  • Pooling (mean, sum) as a factor of the direct-model ANOVA: Sec. 2.4, Sec. 3.1, Sec. 3.1.3, Table 4, Sec. 4.
  • Exploratory split-pooled direct model: Sec. 4, Table 5.
  • Capacity probe rerun on the new pair: Sec. 3.2, App. A.4 (Table 6).
  • Positioning against Westermayr & Maurer and Cignoni et al.: Sec. 1.
  • Revised claims: Abstract, Sec. 4.
  • Released data and checkpoints, including the untouched set: Reproducibility statement.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.