acceptodds
Under review as a conference paper at ICLR 2027

Extending Quantum Transport Simulations to Realistic Device Dimensions through Machine-Learned Electronic Structures

Abstract

Machine-learned (ML) operator models can be trained to predict ab initio Hamiltonian matrices at significantly reduced computational cost compared to density-functional theory (DFT), extending electronic structure calculations to previously unfeasible scales and allowing for their integration with advanced technology computer aided design (TCAD) tools. Here, we introduce MALOQ, a model to train on and predict electronic structure matrices for complex materials consisting of thousands of atoms, described by large orbital basis sets, and covering a wide range of atomic elements. MALOQ extends a state-of-the-art, SO(2)-equivariant architecture with custom data-processing kernels to handle high-rank sparse Hamiltonian matrix data and a scalable edge-wise distribution to treat large atomic graphs. We demonstrate training (inference) on structures made of 3,000 (48,000) atoms, up to 384 (128) GPUs. For systems with 10k+ atoms, MALOQ produces Hamiltonian matrices 10,000x faster than standard DFT codes, at high spectral accuracy. These matrices can be input to a quantum transport (QT) solver, enabling the simulation of entirely-atomistic devices consisting of 21,600 atoms (>180k orbitals). Overall, the full MALOQ+QT pipeline outperforms the reference DFT+QT one by a factor >150x, while maintaining comparable accuracy.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Hamiltonian learning, semiconductor device modelling

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

This paper presents MALOQ, a framework for training and running MLH models on large atomic systems. Combined with a quantum transport solver, the authors simulate a MoS₂/HfO₂ transistor and report a >150× speedup over a DFT+QT pipeline.

Critical strengths
  • The problem of bridging ML Hamiltonians and device-scale quantum transport is important and relevant to TCAD.
  • The systems work is solid, with scaling experiments on up to hundreds of GPUs and micro-benchmarks for each component.
Critical weaknesses

The related issues will be elaborated in the "Questions" section.

Questions for authors
  • I am less familiar with quantum transport and device simulation, but I am familiar with the literature on machine-learned Hamiltonians. The contribution is primarily one of systems engineering and application, and it is unclear how much transferable knowledge the paper offers. The backbone is taken directly from HELM. Kernel fusion, block packing, and graph partitioning (METIS, edge-wise) are standard techniques, and distributed GNN inference for MLIPs already exists. The paper does not clearly articulate what is specific to MLH that makes these solutions non-trivial, nor does it compare against existing distributed approaches in terms of communication volume or scaling. As a result, the main novelty is first to combine these pieces for MLH at scale, which is valuable but incremental from an ML-methods perspective.
  • Large-scale DFT is indeed challenging, but linear-scaling DFT methods have also been developed. Compared with such methods, what advantages does MALOQ offer in terms of efficiency and accuracy?
  • For systems too large for DFT to provide reference results, can the authors provide any form of validation, for example against experimental data, or via indirect checks such as convergence of key observables with system size?
  • Is training on large graphs truly necessary for the systems studied here? Could the authors compare against a locality-based, small-cell training strategy (e.g., DeepH-style) to show that distributed large-graph training is genuinely required?
6 comments

Thank you for writing a thoughtful review. We have tried to structure our response below to stress that (1) Hamiltonian prediction at scale is a highly relevant application (as we submitted to the track 'applications to the physical sciences'), (2) the challenges involved in doing so are significant and fundamentally different than previously encountered by MLIPs, (3) and that our objective is to design the computational innovations to tackle them, and demonstrate how a new applications of ML in the physical sciences (specifically, semiconductor device modeling) has been enabled as a result.

(Q1) ML Hamiltonian prediction has a fundamentally different computational structure: We realize that this is important information that should have been in the main paper (and not in the appendix).

First, we want to preface our response with a general point about applications: Many, if not most of the “big ticket” directions for Hamiltonian prediction, involve structures with a large number of atoms, from electronic materials engineering to semiconductor device modeling. However, due to inherent computational difficulties (explained below), moving to larger systems is not straightforward. The community’s efforts in this direction started with small molecule or crystalline material benchmarks, such as MD17, QH9, and crystalline MoS2 structures in the first DeepH work. Before MALOQ, the few MLH models which attempted to scale up to a large number of atoms either (1) gave up important features in the ML architecture or (2) remained limited to modest system sizes. SLEM [1] and Ref. [2] both turned to “local” approaches, with a single message-passing/interaction layer. Other models like DeepH, QHflow, UniHamGNN, or HELM, all carefully select the size of their training systems to fit within the memory of a single GPU. Practically, the challenge of generalizing to significantly larger numbers of atoms, as required for practical scientific applications, has not yet been addressed.

The reason Hamiltonian prediction poses a different set of challenges than MLIPs boils down to a key point: MLIPs are a node-wise, scalar/vector prediction problem, while ML Hamiltonians take the form of an edge-wise, tensor-prediction challenge. In MLIPs, we predict a scalar (energy) per structure or a vector (force) per atom. Conversely, in ML Hamiltonians, each edge, including self-edges, carries an orbital-pair block that decomposes into a set of irreps up to lmax = 6-8, depending on the orbital basis chosen. This means:

  1. We need to include long range interactions during training in order to generalize to arbitrary sizes during inference. If two atoms interact with each other, the corresponding nodes are connected in the underlying graph neural network, and their interaction is represented by a block in the Hamiltonian matrix. In realistic systems, atoms located 10A apart can still interact with each other (see Fig. G.1). Without having ever ‘seen’ such long-range interactions, a model cannot predict them. Meanwhile, small errors in matrix elements compound and create larger errors in the eigenvalues, so artificially truncating the matrix by imposing smaller cutoffs corrupts downstream physical quantities, e.g., transmission functions, total energies, density-of-states, or electronic currents. In practice, we can't escape the necessity of training on large graphs.

  2. ML Hamiltonian models need to consider much larger interaction cutoffs than MLIPs. This higher cutoff leads to significantly increased edge/node ratios compared to MLIPs, where typically cutoffs are restricted to 6A (UMA). For reference, we present in the table below how the edge:node ratios would scale with interaction cutoffs for one of the training structures we used for the example in Section 5:

Interaction cutoff (Å)Edges/structureAtoms (nodes)/structureEdge:node ratio
6124,9102,08860 : 1
8294,0322,088141 : 1
10579,5582,088278 : 1
12996,2722,088477 : 1
  1. Hamiltonian data contains a fixed 'lmax' and 'mmax', unlike MLIPs where these are hyperparameters. For large datasets like OMol electronic structures (see Appendix E of our paper where this data is treated), lmax goes up to 8. A single edge in a model like HELM contains 81 spherical coefficients, compared to 9 for UMA-small (or 19 for UMA-medium).

Together, these differences between ML Hamiltonian and MLIP prediction translate into massive data volumes that exceed the memory of a single GPU node, even for moderately sized systems (~2000 atoms), and a complex implementation to handle dense, edge-wise communication patterns. This is what makes the problem challenging and not comparable to what is encountered in MLIPs.

(Q1.1) The scope of our work ranges from computational innovations to real deployment in currently relevant scientific applications: We want to underline that the distributed graph implementation required for H prediction is only a sub-set of what our paper covers. MALOQ provides an integrated computational framework spanning several aspects of large-scale ML Hamiltonian learning and its deployment in scientific applications. What we present ranges from efficient distributed implementations over massive edge-wise graphs (Section 3.3), accelerated data processing pipelines over large-scale molecular datasets (Appendix F), demonstration of strong scaling performance across both large molecules (Appendix E) and much larger materials systems than have been treated in any previous work (Fig 4), connection to an advanced quantum transport modeling pipeline (Methodology in Appendix A), and an end-to-end application in an industry-relevant transistor modeling workflow (Section 5).

(Q2) Linear scaling DFT is demonstrably slower here, and generally does not provide an advantage across all material systems: We used the linear scaling DFT algorithm of CP2K to recompute the Hamiltonian matrix of structures made of 2,088 and 4,176 atoms (see the table below). It has the same accuracy as DFT. As compared to the standard O(N3) solver of DFT, the time with the linear scaling algorithm increases by a factor of 4x (93 sec per iteration with linear-scaling instead of 23 sec with the O(N3) approach), which corresponds to 2.6 node hours (NHs) instead of 0.64 NHs for 100 SCF iterations (the number of iterations is model-independent in our case). Since inference on the 2,088-atom structure with MALOQ takes 2e-3 NHs, at the same accuracy, the speed up we obtain with ML compared to linear scaling DFT is in the order of 1,300x (3 orders of magnitude).

The reason why linear scaling DFT is slower than conventional DFT for our structures is that such algorithms exploit a very specific property present in large enough structures and with a significant bandgap. If these conditions are satisfied, the corresponding density matrix becomes sparse, and linear scaling algorithms can take advantage of this characteristic. In many device applications, which include metallic layers or small-gap semiconductors, linear scaling algorithms may not show any computational advantage or even break down. We will include this discussion in a revised version of the paper since this is an important point that may not be widely known.

# atomsDFT, 1 SCF iter (NH)LS-DFT, 1 SCF iter (NH)Speedup of LS-DFT over DFT
2,0880.006390.02580.25× (4.0× slower)
4,1760.02390.1600.15× (6.7× slower)

(Q3) Device-level trends in downstream transport modeling observed for smaller structures hold true for the largest ones. The MoS2/oxide structures used here are a subset of a larger applications-focused direction, and since the submission of this paper, MALOQ has been deployed on structures at the 100k atom scale on similar geometries, matching the physical dimensions of fabricated transistors under consideration as future technology nodes in the semiconductor industry. While a direct comparison with experimental data is not yet conducted, we have clear evidence that our results are sensible and interpretable.

We simulated MoS2/oxide nanoribbons with increased width w (out-of-plane axis in Fig. 6(b)) between 1.5 and 12 nm, predicted their Hamiltonian matrix with MALOQ, computed the electronic current that flows through them, and normalized this current with w. From experiments, we expect that the normalized current would increase at small w and then saturate, i.e., becomes width-independent. This is exactly the behaviour we observe in our simulations, and the length scale of saturation can be quantified and will be compared with experimental measurements. This demonstrates convergence of a 'key downstream physical observable' with increasing system size.

(Q4) Please see our response to Q1, where we explain why large-graph training cannot be circumvented.

We are interested to hear your opinions after our response, and look forward to a reply.

[1] Zhanghao Zhouyin, Zixi Gan, Shishir Kumar Pandey, Linfeng Zhang, and Qiangqiang Gu, Learning local equivariant representations for quantum operators ICLR 2025. URL https://arxiv.org/pdf/2407.06053v1

[2] Chen Hao Xia, Manasa Kaniselvan, Alexandros Nikolaos Ziogas, Marko Mladenovic, Rayen Mahjoub, Alexander Maeder, and Mathieu Luisier. Learning the electronic hamiltonian of large atomic structures, ICML 2025. URL https://arxiv.org/abs/2501.19110.

We are waiting to hear your opinions on our comments. Our responses to your original review are further summarized below:

(1) We articulated the differences between MLIPs and MLH which necessitated our work and make it non-trivial, eg the edge-wise (vs. node-wise) nature and higher spherical degree of the data.

(2) We showed that linear scaling DFT is slower than standard DFT algorithms for our atomic structures. This is not unexpected, and originates from the nature of linear scaling DFT, which takes advantage of properties (wide bandgap, sparsity) that are not generally present across all material systems.

(3) To offer the "indirect checks such as convergence of key observables with system size" suggested, we described our indirect validation of reproducing experimentally-relevant trends from scaling the transistor channel widths. To further directly validate our work against DFT, in our response to reviewer EE76 we also included a full IV-curve comparison.

(4) We clarified why the training structures cannot be made smaller without losing inference accuracy, and that this is what necessitates distributed training.

We look forward to further exchanges during the discussion period.

Reviewer_RpxSOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper develops a relatively comprehensive computational framework for large-scale quantum transport simulations and demonstrates valuable progress in GPU acceleration, distributed computing, and practical applications. However, its main contributions lie in computational optimizations and application extensions of existing machine-learning architectures, rather than substantial innovations in Hamiltonian learning methods or model architectures. While these technical improvements have practical value, they do not establish broadly applicable methodological advances in machine learning. Overall, this work represents a useful contribution to computational materials science, with its primary strengths in computational efficiency and scalability. However, given ICLR's emphasis on methodological innovation and broader relevance to machine learning, I do not consider the current contributions sufficient to warrant acceptance at this conference.

Critical strengths

This paper addresses an important computational bottleneck in large-scale atomistic quantum transport simulations and has clear practical value. The authors have carried out substantial and well-executed engineering optimizations in GPU computing, data processing, and distributed graph neural network implementation, achieving good parallel efficiency and scalability.

Critical weaknesses
  1. The work offers limited methodological novelty in machine learning. MALOQ largely builds upon existing equivariant graph neural network architectures and Hamiltonian representations. Its primary technical contributions focus on GPU kernel optimization, data conversion, graph partitioning, distributed communication, and software integration. While these engineering improvements are valuable, the methodological novelty remains insufficient for ICLR.

  2. The large-scale results mainly demonstrate engineering scalability rather than advances in learning capability. The authors emphasize the ability to handle systems containing tens of thousands of atoms. However, this achievement primarily benefits from the sparsity of local Hamiltonians and efficient distributed computing. The authors should clarify what new learning capabilities the proposed framework offers over existing machine-learning Hamiltonian models.

Questions for authors
  1. The current framework uses relatively large systems for training. What are the computational costs of generating the training data and training the model? Could the model instead be trained on smaller systems and then applied to larger systems without retraining? The authors should investigate the transferability of the model across different system sizes, which is important for evaluating its practical efficiency and scalability.
Reviewer_EE76OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The work predicts DFT Hamiltonian matrices for very large structures and feeds them into a quantum transport solver. The ML model itself is the existing HELM backbone (eSCN convolutions). The contributions are engineering. Fused GPU kernels for the Wigner matrices and the matrix-to-label conversion, and splitting one large atomic graph across many GPUs with embedding exchange between them. Trained on 50 MoS$_2$/HfO$_2$ structures of 2,088 atoms, it matches DFT on one 8,352-atom test structure and is then used to simulate a 21,600-atom transistor that DFT cannot handle, with a claimed 10,000×10{,}000\times speedup in generating the Hamiltonian.

Critical strengths
  1. Real device scale. Hamiltonian prediction and quantum transport at 20k+ atoms without tiling, including the oxide layers, is a meaningful step toward atomistic device simulation.
  2. Solid engineering. Strong and weak scaling are clearly reported, and the custom kernels are much faster than the reference implementations.Moreover, the full pipeline from structure to Hamiltonian to current works, and the bond-resolved current in the full device is a nice result.
Critical weaknesses
  1. Little novelty in the distributed part. Splitting a graph across GPUs and exchanging boundary embeddings, with METIS or edge-wise partitioning, is standard in distributed GNNs and in distributed MLIPs such as SevenNet and DistMLIP. The paper's own claim is only that this has not yet been applied to Hamiltonian models.
  2. Thin accuracy evidence. All results are on one material system, with one validation and one test structure. Eigenvalue errors are 38 meV on average and 20 meV near the band gap, comparable to the thermal energy, and the transmission curves visibly differ above about 0.65 eV. The <0.1% current error comes from a single bias point, and the 21,600-atom headline device has no DFT reference at all.
  3. Not practical for most users. Compute time is short, but memory needs are large. One 3,000-atom structure needs 64 GPUs to train and 8 for inference, and the 21,600-atom device needs about 60 GH200 GPUs at once. Each new material also needs new DFT data and a new training run on 40 GPUs, so in practice the method needs a supercomputer.
  4. Missing baselines. There is no comparison with universal Hamiltonian models (NextHAM, Uni-HamGNN), which could be fine-tuned instead of training from scratch. There is none with tight-binding, DFTB or ML tight-binding (DeePTB), which are the usual route to device-scale transport on modest hardware. There is also none with simply predicting overlapping chunks separately, which the locality of the Hamiltonian allows without any GPU communication. Finally, the speedup is measured against projected cubic-scaling DFT rather than linear-scaling DFT.
Questions for authors
  1. What is new in the distributed scheme compared with existing distributed GNNs and MLIPs such as SevenNet and DistMLIP?
  2. How accurate is MALOQ across more test structures, a full I–V curve and a second material system?
  3. How many GPUs are needed at once, and can it run on a few GPUs, for example by predicting overlapping chunks separately?
  4. How does it compare in accuracy and cost with a fine-tuned universal model, and with tight-binding or DeePTB?
  5. What is the speedup against linear-scaling DFT, using measured rather than projected DFT times?
10 comments

Thank you for the comments! We have added some additional data, clarifications, and discussion on the weaknesses/questions mentioned in our responses below (they will be posted across multiple comments). We are looking forward to hearing your thoughts on them.

1. Novelty of distribution: As you pointed out, the few distributed solutions for GNN-based models are concerned with MLIPs and not with Hamiltonian prediction tasks. But we don’t claim that this is a simple extension of what has been done for MLIPs, there are different computational challenges between the two. Although we briefly touched on this in Section B of the Appendix, we give a more general picture below.

All of the complications boil down to this: MLIPs are a node-wise, scalar/vector prediction problem, while ML Hamiltonians are an edge-wise, tensor-prediction challenge. In MLIPs, we predict a scalar (energy) per structure or vector (force) per atom. Conversely, ML Hamiltonians are edge-wise tensor prediction problems: each edge, including self-edges, carries an orbital-pair block that decomposes into a set of irreps up to L = 6-8 (depending on basis).

Practically, that means the following: ML Hamiltonian models need to use much larger interaction cutoffs, in order to make predictions for Hamiltonian data where atoms interact with each other over distances equal or exceeding 10A. If we connect the graph to a shorter cutoff, then, at best, only matrix elements up to that distance can be learned and predicted, and this kind of truncation in H subblocks changes the eigenvalues. This high cutoff leads to significantly higher edge/node ratios compared to MLIPs, where typically cutoffs are restricted to 6A (UMA). Second, Hamiltonian data contains a fixed 'lmax' and 'mmax', unlike MLIPs where these are hyperparameters. For large datasets like OMol electronic structures (see Appendix E of our paper where this data is treated), lmax goes up to 8. A single edge in a model like HELM contains 81 spherical coefficients, compared to 9 for UMA-small (or 19 for UMA-medium).

So we can summarize the differences between H and MLIP predictions as follows: unlike MLIPs, ML Hamiltonian models (1) contain edge-dominated graphs, (2) rely on significantly larger edge-wise feature sizes, and thus (3) require distributed training, as the memory requirements exceed a single GPU node. Together, this translates into massive data volumes, even for moderately sized systems, and complex implementation to handle dense, edge-wise communication patterns. This is what makes the problem challenging. We realize that this discussion is important to understand the context of our paper, and will move it from Appendix B into the main text.

Comparison with SevenNet and DistMLIP: Please see our response above, where we tried to explain the different computational challenges between MLIPs and ML Hamiltonians. As a further note, we are familiar with both SevenNet and DistMLIP, which both treat MLIPs: SevenNet uses spatial domain decomposition, which is designed for MD where atoms move between domains. DistMLIP is not multi-node, the ICLR 2026 paper contributes a node-wise distributed MLIP implementation on up to 8 GPUs on a single node. Finally, these distribution schemes are designed for inference. With MALOQ, we show that up to 512 GPUs across multiple nodes of a supercomputer can be used to both train and infer on large atomic systems for Hamiltonian prediction.

2. We validated the accuracy of the model at length in order to establish its utility for device modeling applications. Below, we provide some additional validation data (I-V curve) and further discussion.

Eigenvalue error context: The band structure of MoS2/HfO2 in Fig. G.1(d) indicates that the error between the ML and DFT bands is significantly smaller than 20 meV. Nevertheless, 20 meV in the eigenvalues is an incredibly precise prediction - more than 138 million data points, i.e., all Hamiltonian matrix entries, must be predicted to high accuracy (0.0378 meV mean absolute error in our case) in order to obtain this number. Comparing this number to the thermal energy, 25.8 meV at room temperature, is not directly meaningful. Practically, the I-V comparison (see the table below) shows directly that the eigenvalue error does not translate into large current errors.

More test structures/different material system: The MALOQ checkpoint used in this submission has since been finetuned on additional MoS2/Al2O3 structures (different oxide) as part of an upcoming applications work. We used the same number of training structures, each containing 2520 atoms, and obtain comparable accuracies (34.1 meV error in the average eigenvalues with 0.8 meV at the conduction band edge, same qualitative behaviour for the transmission function of a 2k- and 8k-atom system). If you would recommend including these results in the main paper, we would be happy to do so.

Quality of transmission curves: It is true that the differences between the ML- and DFT-based transmission functions are slightly larger at energies above 0.65 eV, as can be seen for example in Fig. 5(d). We would like to stress that (i) the agreement between both sets of curves remains appreciable, even above 0.65 eV, (ii) states at energy above 0.65 eV barely contribute to the electronic current (with a Fermi level at about 0.55 eV, states at 0.65 eV and above have a ~2% probability to be occupied and thus to carry current), and (iii) the ML- and DFT-based electronic currents we computed from these transmission curves agree very well with each other (see the Table below).

Full I-V characteristics (multiple bias points): We computed a full I-V curve with 7 bias points for a MoS2/HfO2 transistor, using the 8,352-atom system as a channel (the gate-to-source voltage was swept between 0 and 0.6 V at a drain to source voltage of 0.6 V). The average error in the electronic current for these 7 bias points, when comparing ML- and DFT-based results, is 3.3% (maximum 9.3% at V_GS = 0.5 V) (see table below) across >4 orders of magnitude. In our experience, this level of accuracy is sufficient to compare different material systems and draw meaningful conclusions from these simulations.

Gate voltage VGS (V)DFT current I_DFT (µA)ML current I_ML (µA)Error |I_DFT-I_ML|/I_DFT (%)
01.47e-051.49e-051.53
0.12.33e-042.33e-040.16
0.22.78e-032.81e-031.32
0.32.18e-022.26e-023.46
0.48.40e-028.66e-023.05
0.52.18e-012.38e-019.32
0.64.86e-015.05e-014.03

3,4,5.

GPU requirements: 8 GPU nodes with 4 GH200 GPUs each were required for memory to train on the data (2,088-atoms each) used here (2 nodes are needed for eval/inference). Our work should be placed in the context of its downstream application, ab initio quantum transport simulations of nanoscale devices, in particular transistors. These kinds of simulations are typically run on GPU clusters and supercomputers. The number of GPUs that we used for the training and inference of MALOQ here is on par with what was used for the device simulations themselves. The capability to effectively leverage hardware at this scale is one of the strengths of our work. We also want to highlight the following part of our paper: "For this system, a single inference takes 0.02 NH, instead of 200 NHs with DFT, thus fully amortizing the computational cost of dataset generation (30 NHs) and training (120 NHs)."

Overlapping chunks: It is not possible to subdivide a cell with full 3D periodicity into overlapping partitions while preserving local atomic environments. The closest attempt to do this is Xia et al [1]. Still, they could not simply consider independent chunks of materials, they had to (1) limit the architecture to a single message passing/interaction layer, and (2) augment each partition with 'virtual' boundary atoms to preserve the local atomic environments under the single-layer constraint. However, this constraint degraded the achievable accuracy: the agreement in transmission functions is only qualitative, and the error in the predicted electronic currents exceeds 50% in most cases.

Tight binding comparison: TB is a (semi-)empirical model based on empirical parameters. It was popular in the device modeling community 10 years ago, but has been gradually replaced by ab initio methods such as DFT, which are more accurate, better transferrable to complex nanostructures, and do not rely on fitting. It generally fails at accurately describing amorphous layers, especially nearest-neighbor models, because of the inherent difficulty to fit its parameters when no clean band structure exists and strong disorder is present. Many device simulations also include metallic layers (gate contact, etc). Due to the strong localization of the underlying Löwdin orbitals, TB cannot capture the electronic properties of metals.

Speedup against linear scaling DFT: We used the linear scaling algorithm in CP2K to recompute the Hamiltonian matrix of structures made of 2,088 and 4,176 atoms (see the table below). As compared to the standard O(N3) solver of DFT, the time increases by a factor of 4x (93 sec per iteration with linear-scaling instead of 23 sec with the O(N3) approach), which corresponds to 2.6 node hours (NHs) instead of 0.64 NHs for 100 SCF iterations. Since inference on the 2,088-atom structure with MALOQ takes 2e-3 NHs, the speed up we obtain with ML is in the order of 1,300x (3 orders of magnitude).

The reason linear scaling DFT is slower than DFT for these structures is because “linear scaling” does not mean “faster”. Linear scaling DFT exploits a very specific property that appears in structures large enough and with a significant bandgap. If these conditions are satisfied, the corresponding density matrix becomes sparse. Linear scaling algorithms can exploit that. In many device applications, which include metallic layers or small-gap semiconductors, linear scaling algorithms may not show a computational advantage. We will include this discussion in the paper since this is an important point to make!

# atomsDFT, 1 SCF iter (NH)LS-DFT, 1 SCF iter (NH)Speedup of LS-DFT over DFT
2,0880.006390.02580.25× (4.0× slower)
4,1760.02390.1600.15× (6.7× slower)

[1] Chen Hao Xia, Manasa Kaniselvan, Alexandros Nikolaos Ziogas, Marko Mladenovic, Rayen Mahjoub, Alexander Maeder, and Mathieu Luisier. Learning the electronic hamiltonian of large atomic structures, 2025. URL https://arxiv.org/abs/2501.19110.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of discussion so far

We provide a summary of our discussion with reviewers at this time:

1. Reviewer EE76 has stated across their replies that all of their original points have been addressed, and generally acknowledges the significance of our work. However, the fact that MALOQ requires more than one GPU is upsetting to the reviewer. This appears to be at the origin of their focus on (1) parameterized tight-binding methods (we have many years of experience using empirical tight-binding in applications, and can state with confidence that it trades accuracy for being computationally much cheaper), (2) local message passing (this potentially trades expressivity for reduced memory requirements), and (3) CPU-only execution (the relative speedups between two optimized GPU codes is not expected to change under a CPU-only scenario). In their last comment, they indicate that they will not change their score, and will "leave the final assessment to the AC". Nevertheless, we appreciate the very extensive nature of their engagement with our work.

2. Reviewer RpxS initially felt that our work is out of scope for ICLR. They say their concerns were addressed by our justification, in short that MALOQ's capabilities don't just make existing models faster, they make new applications possible (our paper introduces a rule on the minimal cell size which must be satisfied in the training data for the model to later generalize to far larger structures). They then raised their score without raising any further points, and we interpret from this that their concerns about scope have been addressed. We are still waiting for their reply.

3. Reviewer jztg asked about the differences between MLIPs and MLH, linear scaling DFT, indirect checks, and necessity of including orbital interactions at 10A. We clarified the difference, showed that linear scaling is slower (not unexpected), provided a description of width scaling as an indirect check, and included data to show the necessity of interactions at 10A. We are now waiting for their final response.

More generally, MALOQ pushes the scale of ML Hamiltonian prediction to new frontiers, enabling ab initio device modeling with experimentally relevant dimensions. Speaking from our core expertise in the field, we can say that our results demonstrate a leap in capabilities compared to what was previously available. Our paper shows strong computational resource usage towards a practically relevant scientific problem, unprecedentedly large materials and molecular training and prediction, and a meaningful end-to-end reduction in time to solution. We appreciated having engaged and invested reviewers who have shown interest in our paper and recognized its impact.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.