acceptodds
Under review as a conference paper at ICLR 2027

Performance-driven design of machine-learned interatomic potentials

Abstract

Machine-learned interatomic potentials deliver close to density-functional accuracy at a small fraction of the cost, but they are still orders of magnitude more computationally expensive than classical force fields, which keeps most biomolecular simulation out of reach. Much of that cost is not intrinsic to the physics but is due to operations that were designed for model expressiveness and accuracy. We present the SPLine Interatomic Cluster Expansion (SPLICE), an equivariant message-passing architecture designed for inference speed. Its radial basis is a per-channel spline carrying two learned parameters, which has a closed form and is evaluated inside the message-passing kernel. We also decouple the width of the per-edge tensor product from that of the node features, so that the model capacity can grow with the least impact on inference performance, leading to a family of models along a speed and accuracy front. Trained on biomolecular data, the model family spans 0.2 to 4 μs per atom per step on a single NVIDIA H100 80GB GPU, with the fastest models handling more than a million atoms. We evaluate it on conformer energies, on physically motivated benchmarks, and on condensed-phase dynamics of a solvated protein and show on par accuracy with the state-of-the-art models.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
machine-learned interatomic potentials, atomistic simulations, equivariant graph neural networks

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_mTfPOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SPLICE, a MACE-derived interatomic potential that replaces the radial MLP used to generate tensor-product weights with a compact learnable spline parameterization. It also decouples the edge-message width from the node-trunk width. The implementation uses Loom to fuse radial evaluation, spherical harmonics, tensor products, and aggregation, avoiding materialization of a large per-edge radial tensor. Experiments demonstrate favorable speed–accuracy trade-offs on molecular datasets and include several physically motivated evaluations. The efficiency gains are promising. However, the contribution would be more compelling if it were positioned and evaluated as an architectural modification and implementation strategy for existing equivariant MLIPs, with SPLICE serving as one plugin. My main concerns are the attribution and generality of the method, the technical novelty of the hardware-aware implementation, and the limited workload coverage of the bottleneck analysis.

Critical strengths
  1. The paper addresses an important practical limitation of machine-learned interatomic potentials. The cost of energy-and-force evaluation for large molecular systems. Reducing edgewise work while preserving useful predictive accuracy is practical.
  2. The architecture and configuration details in Appendix B make the mathematical changes relatively easy to identify. The separation of edge and node widths is also a useful design choice to investigate.
Critical weaknesses

The paper overstates its novelty by presenting an architectural modification as a fundamentally new model. The main contribution should be an efficient replacement/plugin/tool for the radial MLP in existing equivariant architectures, rather than a new model.

The authors replace the function that maps radial information to the scalar weights of an equivariant tensor product. Schematically, the original computation is edge_weights = radial_MLP(radial_features), whereas SPLICE uses weights of learnable spline functions. Choosing a spline instead of an MLP for this scalar mapping is not a new mechanism. I wonder whether the same replacement principle could potentially apply to other models with distance-conditioned scalar filters, including non-equivariant architectures.

I would like to see controlled experiments that apply the modification to other MLIPs, such as NequIP. The authors can report the performance of 1) the baseline MLIP models, 2) Radial MLP with separate kernels, 3) Spline with separate kernels, 4) Spline with optimized fusion. This comparison would separate radial parameterization from fusion and test whether the benefit transfers beyond MACE.

Questions for authors

See weaknesses.

3 comments

We would like to thank the reviewer for their comments and for the constructive suggestions. We ran the experiment they proposed. We report the results below and will add them to the manuscript.

1. The change applied to NequIP. We implemented NequIP in our code with InstaDeep's NequIP-SPICE2 configuration (5 layers, l_max 2, 5 Å cutoff). We trained it on SPICE2 with a radial MLP or with our spline, using the same recipe:

armforce MAE (meV/Å)energy MAE (meV/atom)µs/atom (24k atoms, H100)max atoms (one H100, 80 GB)
NequIP-SPICE2 as published18.50.89––
radial MLP18.21.274.62≤ 85k
spline, fused19.81.320.67 (6.9×)2.7M
spline, fused, decoupled widths15.31.020.68 (6.8×)2.4M

Overall, the model is 4.3–6.9× faster (2.7k–24k atoms) and can fit at least 28× more atoms per GPU. The benefits thus transfer to other architectures, although we have to note that for very different models an appropriate architectural change might touch very different operations, depending on the profile. To decouple the widths, we doubled every multiplicity of the node features (e.g. 128x0e instead of 64x0e) and kept the tensor product and its spline weights at the published multiplicities, with linear layers projecting between the two. The per-edge work is unchanged; the added parameters (3.0M in total, vs 1.3M for the MLP arm) act per node. With decoupled widths, the force error is 16% below the MLP's at 6.8× the speed.

2. Separating the parameterization from fusion. We used one code version, GPU and system (MACE-style model, 48k-atom water, H100), with the same tensor-product kernel in every arm.

armspeed-up, TF32 / FP32
radial MLP, separate op1 / 1
spline, separate op1.49 / 1.45
spline fused into the kernel4.58 / 4.15
+ decoupled widths, at the MLP's accuracy5.72 / 4.85

The parameterization alone gives 1.5×, moving it into the kernel is the largest step: 3.1× in total, of which 1.2× comes from the spherical harmonics and 2.6× from fusing the spline. It has to be noted that changing the architecture is what enables the fusion in the first place: a closed-form spline is cheap enough to recompute inside the kernel, so the per-edge weight tensor is never stored. A radial MLP would have to run a small dense network per edge inside the kernel, or store its per-edge output as the baseline does. The same MACE weights run about 2× slower in mace-torch with cuEquivariance than in our code (FP32, measured on two MACE models). This is not included in the 5.7× above.

3. Positioning. We agree that our main contribution in this work is a method:

  • profile the model;
  • find the dominant operation (here the radial MLP, 48–62% of GPU time);
  • replace it with one that is more amenable to acceleration or fusion.

The fused spline and the decoupled widths are what this gives for MACE-family layers. SPLICE is one instance, and NequIP shows that the change transfers. The same replacement applies in principle to non-equivariant models with distance-conditioned filters: SchNet-style continuous filters are also radial MLPs. We have not tested it, but FlashSchNet (ICML'26) already shows that fusing the filter computation pays off there. The kernel generator is not a main contribution of this paper, and every comparison above holds it fixed.

4. Bottleneck analysis. We profiled the four paper models and the MLP/spline pairs on five workloads, energy plus forces on different systems: water boxes of 2.6k, 10k and 48k atoms; a 35k-atom solvated protein; 1k gas-phase molecules.

What dominates the end-to-end evaluation depends on the model:

  • MLP models: the radial MLP takes 48–62% of GPU time in every one.
  • S, M and L after the change: the fused message-passing kernel takes 42–70% in the condensed phase.
  • XL: the dense linear layers dominate (41–54%), so at large widths they are the next bottleneck.

We will add this profile to the appendix, and can share per-workload breakdowns if useful.

Thank you for the work on modification to NequIP. It responds to the specific question well.

Thank you, we are glad the NequIP experiment answers your question.

Following your suggestion, we will change the positioning in the main paper. The abstract and introduction will present the contribution as a method (profile the model, find the dominant operation, replace it) whose result for equivariant MLIPs is a fused spline that replaces the radial MLP. SPLICE will be presented as its instance for MACE-family layers, and the NequIP results will go into the main text to show that the change transfers.

We hope our reply also covers your other points: the ablation that separates the spline from fusion, and the bottleneck profile on five workloads. Our reply to Reviewer 69Pb also compares the size of our speed-up with accepted papers that replace one dominant operation, which bears on the novelty question. If anything is still unclear, we would be glad to clarify before the discussion closes.

Reviewer_69PbOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose a new MLIP architecture that: 1) uses learnable splines instead of a radial MLP, 2) downprojects the high-dim node representations to low-dim representations used in message passing. They find that the architecture provides a favorable speed-accuracy tradeoff in an extensive benchmark against other MLIPs. Upon training with charged particles, the model also sets a new frontier for speed-accuracy in MD simulations.

Critical strengths

The experimental section is strong; they provide a favorable speed-accuracy tradeoff. This carries over to MD simulations. This shows that despite the architectural changes, the model retains sufficient expressiveness. All of the experiments are exhaustively benchmarked against existing models.

Critical weaknesses

Both architectural innovations seem relatively simple to me. The usage of learnable splines is common in the literature along with downprojecting and up-projecting the hidden dimension. Thus, I'm not sure if this warrants top-3 ML conf novelty.

Also, it would be helpful to see longer-range MD simulations, and possibly some other targets than this one protein. I am not sure, from the provided experiments, if we can be confident the model can conduct long MD simulations.

Questions for authors

Can you constrast your empirical improvement and level of architectural novelty with other accepted top-3 ML conf papers on this topic, in the past year? I am impressed with your empirical results but also feel that the architectual innovations are quite simple. If your empirical results fall within the magnitude of improvement of typical papers on this topic, I would be more inclined towards acceptance. Further contextualization from related papers would help me make a final determination.

3 comments

We would like to thank the reviewer for their comments and concrete questions. We try to answer them below with some new simulations and data from the literature:

1. Novelty. We agree that both ingredients are simple, and neither is new on its own.

  • Prior work: Splines are used as tabulated radial functions (PACE, DP Compress) or with learned coefficients on fixed knots (UF3, KAN-based MLIPs). Narrower edge than node features appear in DimeNet++, GemNet, eSCN/EquiformerV2 and MACE-MH-1. We will cite these and correct our wording.
  • What is new: (i) a closed-form, per-channel radial basis with a trained support, evaluated inside the fused message-passing kernel, so no per-edge radial tensor is ever stored; (ii) the measured speed–accuracy map of edge width against node width.
  • The more general point: we show how analyzing the profile of a model can highlight slow operations and how the operations can be modified to improve the speed-accuracy tradeoff. From this point of view, we consider our "simple" modification of the network to be a strength rather than a weakness.

2. Our gains against accepted papers (ICLR / ICML / NeurIPS 2025–26). Speed-ups measured in the same code, with one operation or kernel changed:

paperchangespeed-upaccuracy
Gaunt TP (ICML'25)tensor product on a grid1.3×similar
E2Former (NeurIPS'25)edge→node convolution1.5× vs EquiformerV2better
Smooth Dynamic Cutoffs (ICML'26)adaptive cutoffup to 2.0×forces +5–16%
E2Former-V2 (ICML'26)sparse TP + fused kernel3.0×–
TDN (NeurIPS'25)low-rank TP3.75–4.5×forces +2.7–4.1%
SPLICEfused spline radial, decoupled widths3.3–5.3× at equal width; 5.7× at equal accuracyforces +4–7% at equal width
  • Speed: our gain is at the top of this range.
  • Most of these papers also replace one dominant operation, as we do. E2Former and TDN come with more theory than our work.
  • Memory capacity: SPLICE fits 21–44× more atoms on one GPU than its radial-MLP twin. The largest capacity gains we found in accepted papers are about 5–10× (e.g. FlashTP 4.7×).
  • Several accepted speed papers accelerate models that are not state of the art: FlashSchNet (ICML'26, 6.5× on SchNet) and FlashTP (ICML'25, 4.2–6.2× on SevenNet).
  • Accepted papers have also given up more accuracy than we do: Smooth Dynamic Cutoffs gives up to 2× for +5–16% force error (+10% on MatPES), and AllScAIP (ICML'26), which its authors present as a point on the accuracy–speed front, has 10–20% higher OMol25 force errors than eSEN-md-d (with about 20% lower energy errors).

3. Longer MD on more systems. We ran ns-scale MD of the charge-aware SPLICE-C models on two proteins and a carbohydrate. For the sake of time, we have restricted the largest/longest of these runs to the smaller models. Protocol: Langevin at 298 K, 0.5 fs time step, every run ending in 20 ps of NVE.

system (atoms)modelslengthoutcome
DHFR, AMBER benchmark (23,558)C-S, C-M, C-L (3 seeds), C-XL1 nsall complete; no bond breaks; Cα RMSD 1.2–1.7 Å; NVE drift −0.1 to 4.4% kT/ns
DHFRC-S10 nscomplete, no divergence; Cα RMSD 1.8 Å (max 2.3 Å); drift 3.2% kT/ns
ubiquitin, PDB 1UBQ (17,248)C-S, C-M, C-L (3 seeds), C-XL1 nsall complete; no bond breaks; Cα RMSD 1.6–2.1 Å; drift 1.3–3.1% kT/ns
ubiquitinC-S10 nscomplete, no divergence; Cα RMSD 1.7 Å (max 2.7 Å); drift 2.4% kT/ns
cellulose in water, AMBER benchmark (408,609)C-S1 nscomplete; no bond breaks or proton transfers; drift 9.9% kT/ns; 1.1 ns/day on one GB200 GPU

C-XL, and C-L on DHFR, are retrained with a lower learning rate (same architecture and data), which made them more stable in MD: they complete every DHFR and TYK2 run. Their Eastman errors are 1.11 (C-L; submitted 1.10) and 0.84 kcal/mol (C-XL; submitted 0.86). We will release the retrained models and use them in the paper.

For scale:

  • Conservative foundation MLIPs drift by 16–70% kT/ns on LAMBench.
  • MLIPAudit's pass threshold corresponds to about 64% kT/ns.

We would be glad to give per-run details or run further systems you have in mind.

Reviewer_69PbOpenReview ↗

Reviewer Response

Thank you; this is a strong response. I raise my score to a 3.

Thank you for your reply and for raising your score. Your questions on long MD and on how our gains compare with recent work will improve the revised paper.

Reviewer_5yEROpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper propose a new family of MLIPs, SPLICE, that designed based on the inference speed. A new SPLine expansion is proposed as the core model design. They trained a family of models that trace the speed accuracy curve for SPICE 2 dataset. They also trained on OMol dataset and used a few benchmark to evaluate the model.

Critical strengths
  1. The performace driven design makes sense for MLIP development, also follows the general ML community
  2. Good and careful benchmarks on throughput
Critical weaknesses
  1. For the OMol model, there are no energy or force MAE results, and there are no results from the official evals of the OMol25 paper. All evals of the OMol trained model are external evals. I want to see the MAE number and results from the OMol Leaderboard. The "on par with SOTA" claim is too borad as the evalations shown are specific to the model. The lead on MLIPAudit is only comparing with a handful of models train on the same dataset, not SOTA.
  2. it's unclear where does the speed gain come from, there are any factors changing and there's no clean comparison (for the 14x faster than MACE-OMOL-0 and 25x faster than UMA-s claims throughout the paper) There was the new compiler, TF32 linear, new archticure, JAX backend, etc. The claim could be missleading that the model arch is the main gain, but it's unkown just from the data presented. The L model also failed on conservative test, which is concerning. It might be caused by the TF32, and if that's the case, the model cannot be trusted to run at the current precision, and there are no comparison of the full FP32 speed.
  3. Training cost not reported
Questions for authors
  1. why don't you show the MAE from OMol?
  2. please show a clean ablation for the speed gain
5 comments

First of all we would like to thank the reviewer for their comments. We ran some further experiments to address the weaknesses that were highlighted. We report our findings below for convenience, and will edit the paper as mentioned.

1. OMol25 MAE and the official evaluations. We report below our SPLICE-OMol models, scored on the validation and on the seven official OMol25 evaluation tasks (fairchem's code, scored by the official leaderboard, geometric mean over the tasks):

modelval E (kcal/mol)val F (kcal/mol/Å)official evals: error rel. to MACE-OMOL-0 (geom. mean, 7 tasks)ms/step, 2,661 atoms, H100
SPLICE-OMol-115.231.2163.854.4
SPLICE-OMol-210.050.8632.905.4
SPLICE-OMol-310.150.7462.746.6
SPLICE-OMol-46.930.5491.959.0
SPLICE-OMol-55.270.3931.5519.6
SPLICE-OMol-64.990.3581.4229
SPLICE-OMol-73.860.3221.2559
MACE-OMOL-04.240.2521847

Validation errors are on a representative subsample: a uniform random ~75k of the 2.76M structures of the official validation split (sampling error of the MAE about 2%). We ran the released MACE-OMOL-0 on the same structures (published full-split values 4.56 / 0.252).

SPLICE-OMol-7 thus has about 1.25× MACE-OMOL-0's error on OMol25 tasks at 14× its speed. We can then analyze the error per task (MAE against DFT in kcal/mol):

taskcompares structures withMACE-OMOL-0SPLICE-OMol-7ratio
ligand strainthe same charge and spin0.1830.1540.84
conformersthe same charge and spin0.1560.1591.02
distance scaling (long range)the same charge and spin5.655.230.93
protonationcharges differing by 10.5791.041.80
IE/EAcharges differing by 17.8112.291.57
ligand–pocketdifferent charges in 75% of complexes6.8610.231.49
spin gapdifferent spins10.1214.161.40
geometric mean, same charge and spin0.93
geometric mean, charge or spin changes1.56

Our best model reaches MACE-OMOL-0's error where charge and spin stay the same (0.93×), but not where they change (1.56×).

The leaderboard keeps the DFT references private, so we cannot compare with DFT sample by sample. We can compare with MACE-OMOL-0 itself: we ran the released checkpoint on the same inputs.

tasksamples|OMol-7 − MACE-OMOL-0| (kcal/mol), mean (median)half of the total comes from the worst
ligand strain1,3730.15 (0.08)11% of samples
conformers7,3090.15 (0.08)11%
protonation2,1001.64 (0.53)2.6%
IE/EA7,7496.28 (4.22)17%
ligand–pocket, all5,0115.19 (2.27)10%
ligand–pocket, ligand and pocket neutral1,2641.07 (0.78)19%
ligand–pocket, a partner charged3,7476.57 (3.41)12%
spin gap5,4937.43 (5.18)19%

Most of the differences come from comparing structures of different total charge. Ligand–pocket clearly shows this: when ligand and pocket are both net-neutral, the two models agree to 1.07 kcal/mol; when a partner is charged they differ by 6.6. Since two models' MAEs on a subset differ by at most their mean absolute difference, the neutral complexes account for at most 0.27 of the 3.37 kcal/mol gap in MAE: at least 92% of it comes from complexes with a charged partner. Single charged structures are not the problem: on OMol25 validation, OMol-7's energy error on closed-shell charged structures equals MACE-OMOL-0's (0.98×), while open-shell structures are weaker (1.23×), which adds to the spin-gap and IE/EA errors. Additionally, the differences can be heavy-tailed: on protonation, half of the total comes from 2.6% of the samples, mostly relaxations that end in different minima, while on the other tasks it is spread over 10–19% of the samples (24% for Gaussian differences).

On the official evaluations, the remaining gap therefore lies with the total charge and spin, which enter SPLICE-OMol as one embedding added to every atom's input features, separate from the spline and the fused kernel. Our target application, MD at one fixed total charge, never takes energy differences between charge states. We will add this analysis to the paper and state the limitation.

In measuring this we also noticed a systematic issue in the published Eastman conformer benchmark scoring of MACE-OMOL-0: the published 0.897 kcal/mol is inflated by FP32 round-off. The model sums its atomic reference energies (~70 keV for a 100-atom molecule) in FP32. In FP64 it scores 0.671, purely removing round-off error (FP32 with references set to zero reproduces FP64). SPLICE-OMol-7 (0.797) is therefore at 1.19× MACE-OMOL-0's error on Eastman too. We will correct our statement that it surpasses MACE-OMOL-0.

Accordingly, we will remove the statements about being "on par with SOTA". The claim we support is a speed–accuracy trade-off at fixed training data, as above. We would also like to point out that our architecture search started from the original MACE architecture and changed only a small number of components, chosen for speed. MACE-OMOL-0 keeps lower validation force errors (0.252 vs 0.322 kcal/mol/Å) at 14× our cost per step, which is the trade-off we claim. Our MLIPAudit comparison uses the same training data by design, because the claim is about the architecture. The general-purpose models in the same figure (UMA-S, MACE-POLAR, MACE-OFF, MACE-MP) were trained on more data, and we do show how they lead SPLICE on several tests.

2. Ablation for the speed gains The ablation below uses one code version, one H100 and one 48,384-atom water box. Compiler, kernels and precision are the same in every arm, so only the architecture changes, one piece at a time.

stepµs/atom, TF32speed-up over the radial MLP in the same precision, TF32 / FP32
radial MLP1.811 / 1
spline instead of the MLP, outside the kernel1.221.49 / 1.45
spline fused into the message-passing kernel0.394.58 / 4.15
+ decoupled edge/node widths (force error 23.0 vs the MLP's 23.6 meV/Å)0.325.72 / 4.85

Comparing to the same MACE architecture, our model gives 5.7× at matched accuracy (4.9× in full FP32). The largest step is fusion at 3.1×. This fusion is what the spline makes possible: it is cheap enough to evaluate inside the kernel, so no per-edge weight tensor is ever stored. A radial MLP would have to run a small dense network per edge inside the kernel, or store its per-edge output as the baseline does. So, as argued in the manuscript, the architecture change is what enables the acceleration in the first place. The end-to-end ratios against other codes (14× vs MACE-OMOL-0, 25× vs UMA-S) compare complete models as distributed, so they also include the implementation. Our stack runs the same MACE weights about 2× faster than mace-torch with cuEquivariance (FP32; this is not included in the 5.7×).

As for SPLICE-L's conservation test: TF32 is not the cause.

  • Only one MLIPAudit system fails, a solvated peptide with ions. It fails at random in both precisions: 2 of 10 seeds in TF32 and 2 of 10 in full FP32 (all matmuls in FP32, checked in the compiled graph).
  • Each failure starts with the N-terminal NH3+ proton stretching to 1.3–1.9 Å. That is a chemistry error of this charge-unaware model.
  • The paper's 0.75 is one such seed. With the default seed on the current code, the test scores 1.00 in both precisions, and completed runs drift by 1–5% of the test's threshold.
  • In long NVE runs of solvated proteins (17k–42k atoms), the charge-aware SPLICE-C models drift by at most 4.4% of kT per ns (with C-L and C-XL retrained for stability, see our reply to Reviewer 69Pb).
  • TF32 does not change force accuracy: against DFT, the force MAE moves by at most 0.1% for all four models (2,000 SPICE2 validation structures; FP32 matches FP64 to 0.002 meV/Å). Appendix B ("Numerical precision") estimates +0.13–0.29% for S, M and L from the TF32–FP32 force differences; we will add this direct measurement there.

3. Training cost. Every model trains on one node.

  • SPICE2 models: 8–55 h (about 30–430 GPU-hours, on 4× GB200 or 8× H100).
  • OMol25 models: 16 h to about 8 days (about 130–1,500 GPU-hours).

We will add this as a column of the training table.

We tried to keep the answer reasonably short, but we are happy to give more detail on any point: per-task tables, per-arm timings, or seeds.

As the discussion period closes tomorrow, we would like to check whether our response addressed your concerns. In short:

  • OMol25: validation errors and all seven official evaluations, scored by the leaderboard. SPLICE-OMol-7 has 1.25× MACE-OMOL-0's error at 14× its speed, and matches it where charge and spin stay the same (0.93×).
  • Claims: we withdraw "on par with SOTA" and claim a speed–accuracy trade-off at fixed training data. We will also correct the Eastman comparison with MACE-OMOL-0.
  • Speed ablation: with code, kernels, compiler and precision fixed, the architecture gives 5.7× at matched accuracy (4.9× in full FP32).
  • Conservation test: TF32 is not the cause. The one failing system fails in 2 of 10 seeds in both TF32 and full FP32, and TF32 changes force errors by at most 0.1%.
  • Training cost: 8–55 h per SPICE2 model and 16 h to about 8 days per OMol25 model, on one node.

If any point is still unclear, or there is a result you would like to see, we would be glad to answer before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.