betaFeedback

WorksPhysics and Astronomy

DPA4: Pushing the Accuracy-Cost Frontier of Interatomic Potentials with EMFA SO(2) Convolution

Abstract

Machine-learning interatomic potentials now approach quantum-mechanical accuracy, but the most expressive equivariant architectures are costly to evaluate, and the leading ones depend on auxiliary denoising or direct-force pretraining. We introduce DPA4, an SE(3)-equivariant architecture spanning six size classes from 0.48 to 25 million parameters and reaching the accuracy of the strongest published models at several-fold to an order-of-magnitude higher inference throughput. Its convolution couples edge and node features across all angular degrees in an edge-local frame, and its Wigner bilinear nonlinearity is universal in its full form and, on an exact quadrature grid, equivariant to machine precision. On Matbench Discovery, DPA4 leads every ranked metric, and every variant evaluated lies on the accuracy–throughput Pareto frontier of the compliant leaderboard. DPA4-Pro attains the lowest energy error on OMat24 and lower total-energy and force errors than the strongest conservative baseline on the OMol25 composition-validation split. All variants are trained through the conservative energy-gradient path alone, made practical by a threefold-faster compiled implementation. DPA4 thus brings leaderboard-class accuracy within the routine compute budgets of molecular-dynamics and materials-screening workflows, for both inorganic crystals and organic molecules.

Introduction

Machine-learning interatomic potentials (MLIPs) are moving from case-specific models trained on dedicated density-functional-theory (DFT) data to pretrained atomistic foundation models[1], also called large atomistic models (LAMs), intended as broad DFT surrogates for simulation, materials discovery and molecular design[2]. The lineage runs from the descriptor-and kernel-based potentials of the case-specific era[3, 4, 5, 6, 7, 8] to the graph neural network models that define the LAM era[9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20].

That promise is expensive to realize: training UMA-M[18], built on the eSEN architecture[21], required 129,024 H200 GPU-hours across direct-force pretraining and conservative fine-tuning. The barrier is not confined to pretraining: the per-step expense that makes pretraining costly is paid again in fine-tuning to a target domain, and then at every step of molecular dynamics or structure relaxation in deployment. This motivates the question whether comparable accuracy can be reached at substantially lower cost, at inference as well as in training.

On the accuracy side, equivariant architectures carry directional information as first-class features that transform under SO(3)SO(3), instead of compressing it immediately into rotation-invariant features; NequIP[22], MACE[11], the Equiformer family[16, 17, 23] and eSEN[21] have shown that explicit directional features substantially improve data efficiency and benchmark accuracy. On the efficiency side, the expense of SE(3)SE(3)-equivariant models is dominated by Clebsch–Gordan tensor products, whose cost grows rapidly with angular order. eSCN showed that SO(3)SO(3)-equivariant convolutions can be reduced to equivalent edge-local SO(2)SO(2) operations[24], a strategy also used by recent high-performing models such as eSEN[21] and EquiformerV3[17]. In these constructions the edge modulates the node feature diagonally: the degree-ll edge feature multiplies only the degree-ll node feature, whereas a Clebsch–Gordan product couples every edge degree to every node degree. That restriction, and not the reduction to SO(2)SO(2), is where expressive power is lost; letting the edge couple different degrees recovers it.

Cost is not the only obstacle: the conservative training objective itself resists acceleration. Because forces are the negative gradient of the energy, the force loss differentiates through the energy model and every training step requires a double-backward pass; training stacks developed for large language models, which are tuned for single-backward gradients, do not transfer directly to this setting. Leading models such as EquiformerV3, eSEN and UMA work around this by pretraining with denoising (DeNS[25]) or direct-force prediction, objectives whose targets are obtained from a single forward pass, and only then fine-tuning with the conservative energy-gradient objective. This two-stage protocol spends an additional pretraining stage and the engineering effort to maintain two objectives, motivating an architecture in which conservative energy-gradient training is itself compiler-friendly from the start.

Here we introduce DPA4, an SE(3)SE(3)-equivariant interatomic-potential architecture built on an EMFA (Edge–node degree coupling, Multi-Focus, Attention) SO(2)SO(2) convolution and a Wigner bilinear network (WBN) nonlinearity, which achieves leading accuracy at substantially lower inference and training cost on inorganic-crystal and organic-molecule benchmarks. Its architectural innovations, co-designed for efficiency and accuracy, are (A1) a low-rank edge–node SO(2)SO(2)-equivariant degree coupling in an edge-local frame, (A2) a multi-focus design for message nonlinearity, (A3) envelope-gated attention for message aggregation — the E, MF and A of EMFA — and (A4) a Wigner bilinear network, a universal architecture for nonlinear SO(3)SO(3)-equivariant maps between irreps, provably complete in its full-frame form. All four raise generalization accuracy at low computational cost relative to standard SO(2)SO(2)-equivariant baselines; the WBN, evaluated by exact Lebedev–uniform quadrature on the rotation group, is in addition equivariant to machine precision. A shape-stable, compiler-friendly implementation of conservative energy-gradient training makes the energy-to-force path compatible with torch.compile and gives a threefold wall-clock training speedup in controlled ablations.

Across four datasets spanning inorganic crystals and organic molecules, DPA4 combines high accuracy with low inference and training cost. On Matbench Discovery[26], all five DPA4 variants evaluated on this benchmark lie on the CPS–throughput Pareto frontier among fifteen compliant models; DPA4-Pro leads every ranked metric, whereas DPA4-Plus matches the highest-CPS baseline with 11.9×11.9 \times higher saturated inference throughput (Figure 1). On OMat24[27], a materials dataset comprising 118 million DFT calculations, DPA4-Pro attains the lowest mean absolute error (MAE) in energy (9.4 meV/atom). Its force MAE is 42.7 versus 41.6 meV/Å for EquiformerV3 with Lmax⁡=6L_{\max} = 6[17], a difference of only 1.1 meV/Å; DPA4-Pro has about half as many parameters and requires one-sixth of the training compute. On the OMol25 composition-validation split[28], DPA4-Pro lowers both the total-energy and force MAEs relative to UMA-M-1.1[18], the strongest conservative baseline in the comparison. DPA4-Pro has 25.2M parameters versus 1.4B total (50M active) for UMA-M-1.1 and required over an order of magnitude less training compute than the reported UMA-M-1.0 run; the cost of the subsequent fine-tuning to UMA-M-1.1 is not reported. On SPICE-MACE-OFF[12], DPA4-Plus lowers the aggregate energy and force errors of the best published model by 39% and 36% (Table 4). All DPA4 models in these comparisons use conservative energy-gradient training alone, without auxiliary denoising or direct-force pretraining, whereas several leading baselines use one or both of these stages. DPA4 is evaluated here in the single-task, per-dataset setting; extending this backbone to multitask LAM pretraining remains a direction for future work (§ 3).

Image not extracted

Figure 1. Accuracy–throughput Pareto frontier for compliant Matbench Discovery models[26]. The Combined Performance Score (CPS) is plotted against saturated inference throughput for the five DPA4 variants and ten representative baselines. Marker area is proportional to the number of model parameters. For each model, saturated throughput is the largest atom-normalized throughput among the successful trials of an adaptive diamond-supercell sweep; at each system size, throughput is computed from the mean wall time of 100 evaluations of energy, forces and stress after 10 warm-up evaluations on a single NVIDIA H20 GPU. Higher values are preferred on both axes; the dashed line joins the non-dominated DPA4 variants. The horizontal arrow gives the throughput ratio between EquiformerV3+DeNS-MP and DPA4-Plus at comparable CPS. The break in the CPS axis omits the empty interval between 0.66 and 0.695. The complete benchmark protocol is given in Supplementary Section S-4.

DatasetMACE (M)MACE (L)eSEN (3.2M)eSEN (6.5M)DPA3-L24DPA4-NeoDPA4-AirDPA4-Plus
Energy MAE (meV/atom)
PubChem0.910.880.22<u>0.15</u>0.240.24<u>0.13</u>0.12
DES370K M.0.630.590.17<u>0.13</u>0.180.17<u>0.17</u>0.12
DES370K D.0.580.540.20<u>0.15</u>0.230.180.110.11
Dipeptides0.520.420.100.070.130.10<u>0.06</u>0.04
Sol. AA1.210.980.300.250.310.27<u>0.13</u>0.09
Water0.760.830.24<u>0.15</u>0.320.300.160.11
QMugs0.690.540.160.120.170.12<u>0.07</u>0.05
LWAMAE{sup}a`0.730.650.190.140.220.18<u>0.11</u>0.09
Force MAE (meV/Å)
PubChem20.5714.756.104.218.475.92<u>3.59</u>2.98
DES370K M.9.366.581.851.243.152.13<u>1.22</u>0.91
DES370K D.9.026.622.772.123.192.16<u>1.23</u>0.96
Dipeptides14.2710.193.042.004.813.10<u>1.68</u>1.25
Sol. AA23.2619.435.763.688.776.69<u>3.59</u>2.59
Water15.2713.573.882.506.894.46<u>2.47</u>1.78
QMugs23.5816.935.703.788.665.36<u>2.91</u>2.14
LWAMAE{sup}a`15.4211.663.832.585.783.89<u>2.18</u>1.64
Params2.3M6.9M3.2M6.5M4.9M1.1M5.1M8.8M
**Train time{sup}b`1014––28877.89

Table 4. SPICE-MACE-OFF small-molecule benchmark performance. The DPA4-Neo, DPA4-Air and DPA4-Plus variants are compared with MACE-OFF, eSEN and DPA3 baselines, together with parameter counts and train times. Reported values are test-set MAEs on the SPICE-MACE-OFF dataset[12]; energy MAEs are in meV/atom and force MAEs are in meV/Å. Boldface and underlining denote the best and second-best values for each dataset–target pair; values that coincide at the reported precision carry the same mark. {sup}a LWAMAE denotes the equal-weight geometric mean of subset MAEs, following the DPA3 comparison protocol{citep}20. {sup}b Train time is the aggregate accelerator time, reported in A100-equivalent GPU-days; a dash indicates that no public train-time estimate is available.

Results

The DPA4 architecture

DPA4 is a conservative SE(3)-equivariant message-passing graph neural network that maps atomic species and positions to a scalar potential energy through a Geometry-Informed Embedding (GIE) stage, NlayerN_{\mathrm{layer}} stacked equivariant interaction blocks, and an atomic energy head (Figure 2a). Each interaction block has a residual structure with two skip connections: one over an EMFA SO(2) convolution followed by an equivariant RMSNorm, and one over a second equivariant RMSNorm followed by an equivariant feed-forward network (FFN). Architectural designs A1, A2 and A3 introduced in the Introduction, namely the low-rank edge–node SO(2)-equivariant degree coupling, the multi-focus design for message nonlinearity, and the envelope-gated attention for message aggregation, operate inside the EMFA SO(2) convolution (Figure 2c). Architectural design A4 supplies the SO(3)-equivariant nonlinearity: the equivariant FFN is implemented via a WBN, a stack of Wigner bilinear blocks mapping the node feature space to itself. Forces and virials are obtained by automatic differentiation through the scalar energy.

Image not extracted

Figure 2. Overview of the DPA4 architecture. (a) The full model couples a shared edge cache, a Geometry-Informed Embedding (GIE) stage, NlayerN_{\mathrm{layer}} stacked equivariant interaction blocks (each a residual stack of an EMFA (Edge–node degree coupling, Multi-Focus, Attention) SO(2) convolution and an equivariant feed-forward network (FFN) interleaved with equivariant RMS norms), and an atomic energy head. Architectural designs A1–A3 act inside the EMFA SO(2) convolution, and the equivariant FFN is instantiated as a Wigner bilinear network (A4). (b) The shared edge cache precomputes per-edge quantities; the GIE stage initializes the l=0l=0 slice through Feature-wise Linear Modulation (FiLM) by a local-environment descriptor and the l≥1l\geq1 slices by projecting neighbor directions onto real spherical harmonics weighted by a radial-species profile. (c) The EMFA SO(2) convolution transports features into an edge-local SO(2) frame, applies the low-rank edge–node SO(2)-equivariant degree coupling (A1) followed by the multi-focus design with cross-focus competition (A2), lifts back to the global frame, and aggregates over neighbors through envelope-gated attention (A3).

A shared edge cache (Figure 2b) precomputes and reuses per-edge quantities across interaction blocks: the distance rijr_{ij} and unit direction r^ij\hat{\mathbf{r}}_{ij}, a radial-basis expansion with cutoff envelope, and per-pair species features. It also caches the Wigner-D rotation Dij\mathbf{D}_{ij} and its inverse Dij−1\mathbf{D}_{ij}^{-1} that transport features between the global and the edge-local frame. The GIE stage (Figure 2b) then injects both chemistry and geometry into the initial node representation before the first interaction block. A scalar branch combines per-species type features with an SO(3)-invariant local-environment descriptor through Feature-wise Linear Modulation (FiLM)[29], producing the initial l=0l = 0 slice. In parallel, an equivariant branch projects each neighbor direction onto real spherical harmonics and weights the projection by a radial-species profile, producing the initial l≥1l \ge1 slices. This gives the model chemical and geometric context from the outset, rather than forcing all local-environment information to be discovered through iterated message passing. The construction details are given in § 4.2.2.

The EMFA SO(2) convolution (Figure 2c) transports the source node features into an edge-local SO(2) frame, constructs an equivariant per-edge message in that frame, then lifts the message back and aggregates over neighbors to update the node features. Mathematical details are given in § 4.2.3. In the local frame, the convolution applies the low-rank edge–node SO(2)-equivariant degree coupling (A1), built on the ESCN reduction[24] that replaces the costly SO(3) Clebsch–Gordan tensor product by equivariant edge-local SO(2) operations. It couples every edge degree to every node degree, as an SO(3) Clebsch–Gordan product does, in contrast to similar SO(2)-equivariant constructions in eSEN[21] and EquiformerV3[17], where the degree-ll edge feature multiplies only the degree-ll node feature. Because the SO(2) Clebsch–Gordan coefficients are trivial, they are hard-wired into the block structure of the kernel rather than tabulated; its learnable part, a function of the invariant radial–species features, is factorized at low rank across channels, improving accuracy at modest additional training cost (§ 2.7). The multi-focus design (A2) splits the hidden width into FF parallel focus streams, each processed by its own SO(2) stack and then reweighted by a cross-focus softmax competition, introducing message nonlinearity. At matched SO(2) width, this parallel-focus structure matches the accuracy of the F=1F = 1 single-focus baseline with substantially fewer parameters and lower training and inference cost (§ 2.7). Aggregation over neighbors uses an envelope-gated attention (A3) computed from the SO(3)-invariant l=0l = 0 slice, with a destination-side output gate (§ 4.2.3). This attention-weighted aggregation improves accuracy at small additional training cost relative to a plain envelope-weighted scatter sum (§ 2.7). The cutoff envelope drives edge contributions smoothly to zero as rij→rcr_{ij} \to r_{\mathrm{c}}, a requirement for stable molecular dynamics.

The equivariant FFN that follows the convolution is instantiated as a WBN (A4): a Wigner bilinear block expands the node feature into a pair of scalar functions on the rotation group, multiplies them point by point and re-expands the product, so that a single operation couples all degrees at once (§ 4.2.4). With the full range of frames retained, a stack of such blocks can approximate any transformation of the node features that respects rotations, to arbitrary accuracy (Supplementary Section S-1.8). The re-expansion integral is evaluated by a Lebedev rule on the sphere crossed with a uniform rule in the residual angle, which reaches a given algebraic order of accuracy with substantially fewer sample points than the tensor-product latitude–longitude grids used by Equiformer-family architectures [16, 17] and keeps the numerical equivariance error of the nonlinearity at machine precision (Table 5 and Supplementary Table S-1).

LLProduct gridLebedev quadrature
Rule# ptsfp64 errorfp32 errorpp# ptsfp64 errorfp32 error
28×88 \times8643.62×10−73.62 \times10^{-7}4.77×10−74.77 \times10^{-7}7262.31×10−142.31 \times10^{-14}2.38×10−72.38 \times10^{-7}
312×1212 \times121447.04×10−77.04 \times10^{-7}6.86×10−76.86 \times10^{-7}9383.58×10−143.58 \times10^{-14}3.58×10−73.58 \times10^{-7}
414×1414 \times141967.97×10−77.97 \times10^{-7}1.55×10−61.55 \times10^{-6}13745.82×10−145.82 \times10^{-14}6.56×10−76.56 \times10^{-7}
518×1818 \times183241.48×10−61.48 \times10^{-6}1.49×10−61.49 \times10^{-6}15863.22×10−143.22 \times10^{-14}6.56×10−76.56 \times10^{-7}
620×2020 \times204004.14×10−64.14 \times10^{-6}2.27×10−62.27 \times10^{-6}191467.99×10−147.99 \times10^{-14}8.35×10−78.35 \times10^{-7}
724×2424 \times245763.19×10−63.19 \times10^{-6}2.03×10−62.03 \times10^{-6}211706.86×10−146.86 \times10^{-14}8.79×10−78.79 \times10^{-7}

Table 5. Equivariance of the spherical-grid nonlinearity under random SO(3)\mathrm{SO}(3) rotations. LL is the maximum spherical-harmonic degree. Product-grid rules are (Rϕ,Rθ)(R_{\phi}, R_{\theta}) after the square-grid lift, and # pts is the total number of grid points RϕRθR_{\phi}R_{\theta}. For Lebedev quadrature, pp is the algebraic order of accuracy. fp64 and fp32 denote 64- and 32-bit floating-point arithmetic, respectively; errors are maximum absolute deviations between the two equivariance paths.

The remaining architectural component addresses close-contact physics. DPA4 decomposes the potential energy as a sum of a learned equivariant branch and an analytical short-range branch, E(Z,R)=ENN(Z,R)+EZBL(Z,R)E(\mathbf{Z},\mathbf{R}) = E^{\mathrm{NN}}(\mathbf{Z},\mathbf{R}) + E^{\mathrm{ZBL}}(\mathbf{Z},\mathbf{R}), with the analytical branch given by the Ziegler–Biersack–Littmark (ZBL) screened-Coulomb pair potential [30]. Rather than splicing the ZBL term onto the energy post hoc, DPA4 couples the two branches inside the energy model through native ZBL zone bridging: smooth bridging gates suppress the direct learned short-range pair channel so that the inner-zone pair interaction is handled exclusively by the analytical branch. This removes the unphysical switching force, proportional to the energy mismatch of the two branches, that a coordinate-dependent splicing weight would introduce. Both branches contribute to the same scalar total energy and are differentiated jointly, giving conservative forces with a smooth transition at close approach (§ 4.2.6). Inside the inner zone the learned branch contributes no force, so the short-range force is by construction the analytical ZBL force; a C–Si dimer scan confirms this and shows the force-switching artifact that an external pair correction produces instead (Supplementary Section S-5).

Matbench Discovery materials benchmark

Matbench Discovery is a widely used benchmark for evaluating machine-learning models in high-throughput inorganic-crystal discovery [26]. In the compliant setting, models are trained on MPtrj [10], used to relax WBM candidate structures [33], and then evaluated by their formation-energy predictions and derived convex-hull distances. The resulting scores combine classification, regression and structural-relaxation metrics, making the benchmark a stringent test of both energy accuracy and relaxation quality. The leaderboard also reports κ\kappaSRME, a thermal-conductivity metric sensitive to the smoothness and conservativeness of the learned potential [31]. Here we focus on the compliant leaderboard to keep the training data fixed across models and more directly compare architectural differences. Non-compliant

The five DPA4 variants are trained on MPtrj in the compliant setting and compared with all compliant leaderboard entries on CPS, F1, κ\kappaSRME and RMSD, together with parameter counts and train times.

Leaderboard entries comprise all compliant models accessed before August 1, 2026.

Boldface and underlining denote the best and second-best values of each ranked metric.

leaderboard entries may additionally involve different training datasets or fine-tuning strategies and are left for future comparison.

Table 1 reports the Matbench Discovery leaderboard metrics together with the parameter count and train time of each model. Figure 1 plots the same score against saturated inference throughput. DPA4-Pro establishes the best CPS in Table 1, reaching 0.842 with an F1 score of 0.865, a κ\kappaSRME of 0.227 and an RMSD of 0.068 Å at 25.2M parameters. It is the leading compliant model in every ranked metric of the leaderboard, spanning stability classification, thermal conductivity and relaxation quality. Its CPS exceeds that of EquiformerV3+DeNS-MP[17, 25] by 0.012 (0.842 versus 0.830) while using 17% fewer parameters (25.2M versus 30.3M).

ModelCPS↑aF1↑κ\kappaSRME↓bRMSD↓bParamsTrain timecDeNSdTargetse
DPA4-Pro0.8420.8650.2270.06825.2M233✗EFSG
DPA4-Plus0.8290.847<u>0.240</u><u>0.070</u>8.8M31✗EFSG
DPA4-Air0.8160.8370.2670.0745.1M16✗EFSG
DPA4-Neo0.7820.8100.3530.0791.1M8.6✗EFSG
DPA4-Mini0.7330.7760.4960.0840.66M5.8✗EFSG
EquiformerV3+DeNS-MP<u>0.830</u><u>0.863</u>0.275<u>0.070</u>30.3M157✓EFSG
eSEN-30M-MP0.7970.8310.3400.07530.1M335✓EFSG
MatRIS-10M-MP0.7780.8470.4890.07210.4M224✓EFSGM
Eqnorm MPtrj0.7560.7860.4080.0841.31M83✗EFSG
Nequix MP PFT0.7550.7480.3070.087708k276✗EFSHG
Nequip-MP-L0.7300.7610.4660.0869.6M41✗EFSG
Nequix MP0.7290.7510.4460.085708k21✗EFSG
Allegro-MP-L0.7200.7510.5040.08218.7M57✗EFSG
DPA-3.1-MPtrj0.7180.8030.6500.0804.81M104✗EFSG
SevenNet-l3i50.7140.7600.5500.0851.17M381✗EFSG
HIENet0.7070.7770.6420.0807.51M120✗EFSG
GRACE-2L-MPtrj0.6810.6910.5260.09015.3M–✗EFSG
MACE-MP-00.6370.6690.6820.0924.69M108✗EFSG
eqV2 S DeNS0.5220.8151.6760.07631.2M228✓EFSD
ORB v2 MPtrj0.4700.7651.7260.10125.2M–✗EFSD
M3GNet0.4280.5691.4090.112228k–✗EFSG
CHGNet0.4000.6131.7170.095413k–✗EFSGM

Table 1. Matbench Discovery leaderboard performance of DPA4 and compliant baseline models.

This comparison is notable because several high-ranking baselines use additional accuracy-enhancing training stages, marked in the DeNS column of Table 1. DeNS[25] has been shown to improve equivariant force fields by denoising non-equilibrium structures, and EquiformerV3+DeNS-MP[17] further combines this strategy with direct-force pretraining.

DPA4-Pro uses neither DeNS nor direct-force training; it is trained through the conservative energy-gradient path and still surpasses the DeNS-assisted EquiformerV3 baseline in CPS. This suggests that the accuracy advantage originates in the architecture itself rather than in an auxiliary denoising or direct-force objective.

The smaller DPA4 variants extend this trade-off across model scales: all five variants evaluated on Matbench Discovery lie on the Pareto frontier in CPS and saturated inference throughput (Figure 1), and each matches or surpasses much larger baselines at a small fraction of their model size and training budget. DPA4-Plus reaches a CPS of 0.829 with 8.8M parameters, matching EquiformerV3+DeNS-MP to within 0.001 CPS while running 11.9× faster at saturated inference, using 5.1× less training compute (31 versus 157 A100-equivalent GPU-days) and 3.4× fewer parameters. DPA4-Air reaches a CPS of 0.816 with 5.1M parameters, exceeding eSEN-30M-MP[21] (0.797 CPS, 30.1M parameters) with 20.9× less training compute (16 versus 335 GPU-days) and a 5.9× smaller model. DPA4-Neo contains only 1.1M parameters but still reaches a CPS of 0.782, above MatRIS-10M-MP[34] (0.778 CPS, 10.4M parameters) with a saturated throughput higher by more than two orders of magnitude (28.0 against 0.055 atoms/ms), 26× less training compute (8.6 versus 224 GPU-days) and a 9.5× smaller model. DPA4-Mini, the smallest variant at 0.66M parameters, reaches a CPS of 0.733 at a training budget of 5.8 GPU-days, above the Nequip-MP-L and Nequix MP entries of Table 1 (0.730 and 0.729 CPS), running 23.7× and 4.3× faster at saturated inference and at 7.1× and 3.6× less training compute.

Within the DPA4 series, the accuracy–throughput trade-off of Figure 1 flattens sharply at the top end. The Mini–Neo, Neo–Air and Air–Plus steps trade 2.8×, 2.4× and 2.3× in saturated inference throughput for CPS gains of 0.049, 0.034 and 0.013, whereas the step from DPA4-Plus to DPA4-Pro costs 3.8× (5.1 to 1.4 atoms/ms) and gains only 0.013. DPA4-Plus therefore sits at the knee of the frontier: it matches the accuracy of the strongest baseline at an order of magnitude higher throughput, while DPA4-Pro converts a disproportionate share of throughput into the remaining margin at the top of the leaderboard, itself still 3.2× faster than EquiformerV3+DeNS-MP. On the training side, cost grows sublinearly with model size at the lower end, rising by only 1.9× from Neo to Air against a 4.6× increase in parameters, because neighbor-list construction, geometric preprocessing and memory traffic occupy a larger fraction of each step once the arithmetic workload no longer saturates the GPU. For perspective, the A100 has a peak FP32 throughput of 19.5 TFLOPS, whereas recent single-card workstation GPUs can exceed 100 TFLOPS. Rescaled by the peak-FLOP ratio, the 16 GPU-day budget of DPA4-Air corresponds to roughly three days on such hardware, which makes DPA4-Air and DPA4-Neo practical for high-throughput workflows.

With the leading compliant models now separated by only small CPS margins, substantial further gains in accuracy are likely to require training data beyond the fixed MPtrj set. The compliant setting nevertheless remains the appropriate basis for the present comparison: holding the training data fixed across models attributes the observed CPS differences to the architectures themselves. Within this setting, the accuracy–cost frontier remains open in the cost direction, namely how far equal accuracy can be attained at further reduced inference and training cost.

OMat24 inorganic materials benchmark

OMat24[27] is a large-scale inorganic materials dataset of 118 million DFT calculations, obtained by non-equilibrium sampling — rattling, ab initio molecular dynamics and the The six DPA4 size classes are trained on OMat24 through the conservative energy-gradient objective and compared with published baselines.

Boldface and underlining denote the best and second-best values of each evaluation metric; values that coincide at the reported precision carry the same mark.

relaxation of rattled structures — starting from structures in the Alexandria database[36]. Its validation split has become a common reference point for single-dataset comparisons of recent equivariant architectures. Because the dataset is much larger and more diverse than MPtrj, it separates architectures by their capacity to fit a broad potential-energy surface. We therefore train the six DPA4 size classes on OMat24 with the architecture configurations of Supplementary Table S-20 and report energy, force and stress MAEs on the validation set.

Table 2 shows that DPA4-Pro attains the lowest energy MAE of all models compared, 9.4 meV/atom, with 25.2M parameters. The next most accurate models, EquiformerV2-L[16, 27] (9.6 meV/atom, 154M parameters) and UMA-L[18] (9.7 meV/atom, 700M parameters), are 6.1× and 27.8× larger. Its force MAE of 42.7 meV/Å is the second lowest, behind EquiformerV3 with Lmax⁡=6L_{\max}=6[17] (41.6 meV/Å, 49M parameters), and only the larger EquiformerV2 models reach a lower stress MAE (2.3 against 2.4 meV/Å3^{3}). All DPA4 entries predict forces as energy gradients, whereas the strongest baselines rely on direct force prediction or on direct-force pre-training followed by gradient fine-tuning. The training budgets also differ substantially: DPA4-Pro required 1,474 H100-equivalent GPU-hours, 6.2× below the 9,197 of EquiformerV3 with Lmax⁡=6L_{\max}=6.

ModelPredictiona^{a}Energyb^{b}Forceb^{b}Stressb^{b}Paramsc^{c}Train timed^{d}
MACE-Omat-Small[11, 35]Gradient17.985.93.58.2M–
MACE-Omat-Mediume^{e}Gradient16.378.43.39.1M–
EquiformerV2-S[16, 27]Direct11.049.2<u>2.4</u>31M–
EquiformerV2-MDirect10.044.82.387M–
EquiformerV2-LDirect<u>9.6</u>43.12.3154M–
EquiformerV3, Lmax⁡=4L_{\max}=4[17]Direct+Grad10.443.52.630M4,782
EquiformerV3, Lmax⁡=6L_{\max}=6Direct+Grad10.141.62.549M9,197
eSEN[21]Gradient10.747.32.630M–
UMA-S[18]Direct+Grad11.357.12.9150M (6M)46,080f^{f}
UMA-MDirect+Grad10.047.32.71.4B (50M)129,024f^{f}
UMA-LDirect9.743.52.5700M95,232f^{f}
DPA4-NanoGradient18.795.83.40.48M54
DPA4-MiniGradient14.070.72.90.66M80
DPA4-NeoGradient12.358.72.71.1M139
DPA4-AirGradient10.651.02.65.1M207
DPA4-PlusGradient9.845.62.58.8M368
DPA4-ProGradient9.4<u>42.7</u><u>2.4</u>25.2M1,474

Table 2. Performance on the OMat24 validation set.

Across the six DPA4 size classes, increasing the parameter count from 0.48M to 25.2M and the training cost from 54 to 1,474 H100-equivalent GPU-hours monotonically reduces the energy MAE from 18.7 to 9.4 meV/atom, the force MAE from 95.8 to 42.7 meV/A˚\text{\AA} and the stress MAE from 3.4 to 2.4 meV/A˚3\text{\AA}^{3}. DPA4-Nano, DPA4-Mini and DPA4-Neo occupy the compact 0.48–1.1M-parameter regime at costs of only 54–139 H100-equivalent GPU-hours; within this regime, Mini and Neo reach energy MAEs of 14.0 and 12.3 meV/atom, below the 16.3 meV/atom of the larger 9.1M-parameter MACE-Omat-Medium. DPA4-Air then matches the 30M-parameter eSEN baseline[21] in energy (10.6 versus 10.7 meV/atom) with 5.9× fewer parameters, while DPA4-Plus reaches 9.8 meV/atom and 45.6 meV/A˚\text{\AA} with 8.8M parameters, a lower energy MAE than every baseline except EquiformerV2-L and UMA-L, both more than an order of magnitude larger. At 368 H100-equivalent GPU-hours, DPA4-Plus costs 4.0× less to train than DPA4-Pro for an energy MAE only 0.4 meV/atom higher, making it, as on Matbench Discovery, the most favorable accuracy–cost balance in the series. Together with the Matbench Discovery results, this shows that the accuracy–parameter advantage of DPA4 persists when the training set is enlarged from MPtrj to OMat24.

OMol25 molecular benchmark

OMol25[28] is a molecular dataset of DFT single-point calculations at the ωB97M-V/def2-TZVPD level, spanning biomolecules, metal complexes, electrolytes and recomputed community datasets, with charge and spin states as explicit inputs. It is substantially broader in chemistry than SPICE-MACE-OFF and therefore probes transferability across molecular classes rather than accuracy within a single class. The dataset has been released in successive versions; we train on the OMol-0 release, which the dataset paper describes as containing about 100 million DFT datapoints and which is also the release used for the UMA-1.1 models[28], and we evaluate on its out-of-distribution composition validation split.

Across the six DPA4 size classes, increasing the parameter count from 0.48M to 25.2M monotonically decreases the energy and force MAEs from 8.02 to 0.92 kcal/mol and from 0.776 to 0.117 kcal/mol/A˚\text{\AA}, respectively. Across the series, the training budget increases from 186 to 8,829 H100-equivalent GPU-hours from Nano through Pro. At the compact end, DPA4-Nano, DPA4-Mini and DPA4-Neo span 0.48–1.1M parameters, energy MAEs of 8.02–2.91 kcal/mol and training costs of 186–819 H100-equivalent GPU-hours.

DPA4-Air then matches the 1.45 kcal/mol energy MAE of UMA-S-1.1 at the reported precision and improves its force MAE from 0.22 to 0.189 kcal/mol/A˚\text{\AA} with 5.1M parameters, 29× fewer than the 150M total parameters of UMA-S-1.1. The strongest conservative baseline in Table 3 is UMA-M-1.1, with energy and force MAEs of 1.14 kcal/mol and 0.14 kcal/mol/A˚\text{\AA} at 1.4B total and 50M active parameters[18, 28]. DPA4-Plus matches it at 1.20 kcal/mol and 0.146 kcal/mol/A˚\text{\AA}, and DPA4-Pro surpasses it on both metrics at 0.92 kcal/mol and 0.117 kcal/mol/A˚\text{\AA}, trained with 2,521 and 8,829 H100-equivalent GPU-hours, respectively. UMA-M-1.0 required 129 GPU-hours for its two-stage training. UMA-M-1.1 was obtained by further fine-tuning this model, but the cost of that step is not reported. Non-conservative models predict energy and forces separately, free from the constraint that the forces are the gradients of the energy. Among them, GemNet-OC attains the lowest energy MAE of the table at 0.78 kcal/mol[28, 37], and eSEN-md-d. attains the lowest force MAE at 0.11 kcal/mol/A˚\text{\AA}[21, 28]. Neither, however, surpasses DPA4-Pro on energy and force simultaneously. The monotonic error reduction across the DPA4 size series extends A dash indicates that a value is unavailable.

ModelPrediction{sup}a`Cons.{sup}a`Energy{sup}b`Force{sup}b`Params{sup}c`Train time{sup}d`
MACE-OMol-L-0[11, 28]Gradient✓4.560.2552.4M{sup}e`–
GemNet-OC-r6[28, 37]Direct✗1.040.1739.1M–
GemNet-OC[28, 37]Direct✗0.780.1539.1M–
eSEN-sm-d[21, 28]Direct✗2.060.236.3M–
eSEN-sm-cons.[21, 28]Gradient✓1.770.196.3M–
eSEN-md-d[21, 28]Direct✗1.150.1150.7M–
UMA-S-1.1[18, 28]Direct+Grad✓1.450.22150M (6M)46,080{sup}g`
UMA-M-1.1[18, 28]Direct+Grad✓1.140.141.4B (50M)129,024{sup}g`
DPA4-NanoGradient✓8.020.7760.48M186
DPA4-MiniGradient✓4.970.5020.66M431
DPA4-NeoGradient✓2.910.3321.1M819
DPA4-AirGradient✓1.450.1895.1M1,549
DPA4-PlusGradient✓1.200.1468.8M2,521
DPA4-ProGradient✓<u>0.92</u><u>0.117</u>25.2M8,829

Table 3. Performance on the OMol25 OMol-0 out-of-distribution composition validation split.

Boldface and underlining denote the best and second-best values within each MAE column.

{sup}a` Prediction denotes the force-prediction pathway: Gradient obtains forces as energy gradients, Direct predicts forces directly, and Direct+Grad denotes direct-force pre-training followed by gradient fine-tuning. Cons. marks whether the evaluated forces are conservative, i.e. energy gradients: ✓ for Gradient and Direct+Grad models, ✗ for Direct models.

{sup}b` Values are mean absolute errors (MAEs), reported in kcal/mol for energy and kcal/mol/Å for forces. Values originally reported in meV and meV/Å are converted using 1 meV = 0.0230605 kcal/mol and 1 meV/Å = 0.0230605 kcal/mol/Å, respectively.

{sup}c` Params reports total parameters; parentheses give active parameters for UMA MoLE models.

{sup}d` Train time is the aggregate accelerator time, reported in H100-equivalent GPU-hours. DPA4 H20 runs are converted using the ratio of the theoretical FP32 peak throughputs of H20 and H100 GPUs. UMA runs on H200 GPUs are counted at parity, as the H200 and H100 share the same theoretical FP32 peak throughput.

{sup}e` The MACE-OMol-L-0 parameter count was computed from the released MACE-omol-0-extra-large-1024.model checkpoint distributed through the MACE repository.

{sup}f` GemNet-OC-r6 uses a 6 Å cutoff radius, whereas GemNet-OC uses the default 12 Å cutoff radius.

{sup}g` The listed UMA train times are for the multitask UMA 1.0 runs and are not OMol25-only estimates. UMA-1.1 was obtained by further fine-tuning these models, but the cost of that step is not reported.

the scaling trends observed on Matbench Discovery and OMat24 to both OMol25 energy and force.

SPICE-MACE-OFF molecular benchmark

SPICE-MACE-OFF[12] is a small-molecule benchmark for transferable organic force fields. The benchmark spans PubChem molecules, DES370K[38] monomers and dimers, dipeptides, solvated amino acids, water clusters and larger QMugs-derived molecules[39, 40]. Reference energies and forces are computed at the ω\omegaB97M-D3(BJ)/def2-TZVPPD level[12, 39]. We train the DPA4-Neo, DPA4-Air and DPA4-Plus variants on SPICE-MACE-OFF, using the same train/validation/test split as MACE-OFF[12] and DPA3[20], and report per-subset energy and force MAEs. Following DPA3[20], we also report a logarithmic weighted average MAE (LWAMAE) with equal weights, equivalent to the geometric mean of subset MAEs.

Table 4 shows that DPA4 improves molecular energy and force accuracy across the chemically diverse SPICE-MACE-OFF subsets. DPA4-Plus attains the lowest aggregate energy and force errors, with LWAMAEs of 0.09 meV/atom and 1.64 meV/Å, respectively, and is best or tied for best on all 14 subset–target pairs at the reported precision. Against the

6.5M-parameter eSEN baseline[21], this 8.8M-parameter model lowers the aggregate energy and force errors by 39% and 36%. Against DPA3-L24[20], the corresponding reductions are 61% and 72%, showing that the gain extends well beyond the inorganic-crystal benchmark.

The smaller variants preserve most of this accuracy at a fraction of the model size. With 5.1M parameters, DPA4-Air reaches aggregate LWAMAEs of 0.11 meV/atom and 2.18 meV/A˚2.18\,\mathrm{meV}/\text{\AA}, which are 21% and 16% below the larger 6.5M-parameter eSEN baseline[21] and 50% and 62% below DPA3-L24[20]. It ranks first or second on 12 of the 14 subset–target pairs, and on the PubChem subset its force error is the lowest of all models compared. DPA4-Neo extends the comparison to a still smaller scale: with 1.1M parameters it reaches LWAMAEs of 0.18 meV/atom and 3.89 meV/A˚3.89\,\mathrm{meV}/\text{\AA}, matching the 3.2M-parameter eSEN baseline[21] with 2.9×2.9\times fewer parameters and improving on DPA3-L24[20] by 18% in energy and 33% in force with 4.5×4.5\times fewer parameters. Beyond model size, these gains come at low training cost: DPA4-Neo, DPA4-Air and DPA4-Plus require only 7, 7.8 and 9 A100-equivalent GPU-days, respectively (Table 4), below the MACE baselines (10 and 14) and 32–41× smaller than the 288 GPU-days of DPA3-L24; no public training cost is reported for the eSEN baselines.

Training and inference efficiency

References

  1. [1]Yuan, E. C.-Y. et al. Foundation models for atomistic simulation of chemistry and materials. Nature Reviews Chemistry 10, 212–230, DOI: 10.1038/s41570-025-00793-5 (2026).DOI
  2. [2]Merchant, A. et al. Scaling deep learning for materials discovery. Nature 1–6 (2023).
  3. [3]Behler, J. & Parrinello, M. Generalized neural-network representation of high-dimensional potential-energy surfaces. Physical review letters 98, 146401 (2007).
  4. [4]Bartók, A. P., Payne, M. C., Kondor, R. & Csányi, G. Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons. Physical review letters 104, 136403 (2010).
  5. [5]Schütt, K. et al. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. Advances in neural information processing systems 30 (2017).arxiv.org/abs/1706.08566
  6. [6]Unke, O. T. & Meuwly, M. Physnet: A neural network for predicting energies, forces, dipole moments, and partial charges. Journal of chemical theory and computation 15, 3678–3693 (2019).DOI
  7. [7]Zhang, L., Han, J., Wang, H., Car, R. & E, W. Deep potential molecular dynamics: a scalable model with the accuracy of quantum mechanics. Physical review letters 120, 143001 (2018).DOI
  8. [8]Wang, H., Zhang, L., Han, J. & E, W. Deepmd-kit: A deep learning package for many-body potential energy representation and molecular dynamics. Computer Physics Communications 228, 178–184 (2018).DOI
  9. [9]Chen, C. & Ong, S. P. A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science 2, 718–728 (2022).DOI
  10. [10]Deng, B. et al. Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nature Machine Intelligence 5, 1031–1041 (2023).arxiv.org/abs/2302.14231
  11. [11]Batatia, I., Kovacs, D. P., Simm, G., Ortner, C. & Csányi, G. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields. Advances in Neural Information Processing Systems 35, 11423–11436 (2022).arxiv.org/abs/2206.07697
  12. [12]Kovács, D. P. et al. Mace-off23: Transferable machine learning force fields for organic molecules. arXiv preprint arXiv:2312.15211 (2023).
  13. [13]Yang, H. et al. Mattersim: A deep learning atomistic model across elements, temperatures and pressures. arXiv preprint arXiv:2405.04967 (2024).arxiv.org/abs/2405.04967
  14. [14]Neumann, M. et al. Orb: A fast, scalable neural network potential (2024). 2410.22570.arxiv.org/abs/2410.22570
  15. [15]Rhodes, B. et al. Orb-v3: atomistic simulation at scale (2025). 2504.06231.arxiv.org/abs/2504.06231
  16. [16]Liao, Y.-L., Wood, B., Das, A. & Smidt, T. Equiformerv2: Improved equivariant transformer for scaling to higher-degree representations. arXiv preprint arXiv:2306.12059 (2023).arxiv.org/abs/2306.12059
  17. [17]Liao, Y.-L. et al. Equiformerv3: Scaling efficient, expressive, and general (SE(3))-equivariant graph attention transformers. arXiv preprint arXiv:2604.09130 (2026).arxiv.org/abs/2604.09130
  18. [18]Wood, B. M. et al. Uma: A family of universal models for atoms. arXiv preprint arXiv:2506.23971 (2025).arxiv.org/abs/2506.23971
  19. [19]Zhang, D. et al. Dpa-2: a large atomic model as a multi-task learner. npj Computational Materials 10, 293 (2024).DOI
  20. [20]Zhang, D. et al. A graph neural network for the era of large atomistic models. arXiv preprint arXiv:2506.01686 DOI: 10.48550/arxiv.2506.01686 (2025).arxiv.org/abs/2506.01686
  21. [21]Fu, X. et al. Learning smooth and expressive interatomic potentials for physical property prediction. arXiv preprint arXiv:2502.12147 DOI: 10.48550/arxiv.2502.12147 (2025).DOI
  22. [22]Batzner, S. et al. SE(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications 13, 2453, DOI: 10.1038/s41467-022-29939-5 (2022).DOI
  23. [23]Liao, Y.-L. & Smidt, T. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. arXiv preprint arXiv:2206.11990 (2022).arxiv.org/abs/2206.11990
  24. [24]Passaro, S. & Zitnick, C. L. Reducing SO(3) convolutions to SO(2) for efficient equivariant GNNs. arXiv preprint arXiv:2302.03655 DOI: 10.48550/arXiv.2302.03655 (2023).DOI
  25. [25]Liao, Y.-L., Smidt, T., Shuaibi, M. & Das, A. Generalizing denoising to non-equilibrium structures improves equivariant force fields. arXiv preprint arXiv:2403.09549 (2024).arxiv.org/abs/2403.09549
  26. [26]Riebesell, J. et al. A framework to evaluate machine learning crystal stability predictions. Nature Machine Intelligence DOI: 10.1038/s42256-025-01055-1 (2025).DOI
  27. [27]Barroso-Luque, L. et al. Open materials 2024 (omat24) inorganic materials dataset and models. arXiv preprint arXiv:2410.12771 (2024).
  28. [28]Levine, D. S. et al. The open molecules 2025 (omol25) dataset, evaluations, and models. arXiv preprint arXiv:2505.08762 (2025).arxiv.org/abs/2505.08762
  29. [29]Perez, E., Strub, F., de Vries, H., Dumoulin, V. & Courville, A. FiLM: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32, DOI: 10.1609/aaai.v32i1.11671 (2018).DOI
  30. [30]Ziegler, J. F., Biersack, J. P. & Littmark, U. The Stopping and Range of Ions in Solids (Pergamon Press, New York, 1985).
  31. [31]Póta, B., Ahlawat, P., Csányi, G. & Simoncelli, M. Thermal conductivity predictions with foundation atomistic models. arXiv preprint arXiv:2408.00755 (2024).arxiv.org/abs/2408.00755
  32. [32]Koker, T., Gangan, A., Kotak, M., Marian, J. & Smidt, T. PFT: Phonon fine-tuning for machine learned interatomic potentials. arXiv preprint arXiv:2601.07742 (2026).arxiv.org/abs/2601.07742
  33. [33]Wang, H., Botti, S. & Marques, M. Predicting stable crystalline compounds using chemical similarity. npj Computational Materials 7, 12, DOI: 10.1038/s41524-020-00481-6 (2021).DOI
  34. [34]Zhou, Y. et al. MatRIS: Toward reliable and efficient pretrained machine learning interatomic potentials. arXiv preprint arXiv:2603.02002 (2026).arxiv.org/abs/2603.02002
  35. [35]Batatia, I. et al. Cross learning between electronic structure theories for unifying molecular, surface, and inorganic crystal foundation force fields. arXiv preprint arXiv:2510.25380 (2025).arxiv.org/abs/2510.25380
  36. [36]Schmidt, J. et al. Machine-learning-assisted determination of the global zero-temperature phase diagram of materials. Advanced Materials 35, 2210788 (2023).
  37. [37]Gasteiger, J. et al. Gemnet-OC: Developing graph neural networks for large and diverse molecular simulation datasets. Transactions on Machine Learning Research (2022).
  38. [38]Donchev, A. G. et al. Quantum chemical benchmark databases of gold-standard dimer interaction energies. Scientific Data 8, 55, DOI: 10.1038/s41597-021-00833-x (2021).DOI
  39. [39]Eastman, P. et al. Spice, a dataset of drug-like molecules and peptides for training machine learning potentials. Scientific Data 10, 11 (2023).
  40. [40]Isert, C., Atz, K., Jiménez-Luna, J. & Schneider, G. Qmugs, quantum mechanical properties of drug-like molecules. Scientific Data 9, 273 (2022).
  41. [41]Larsen, A. H. et al. The atomic simulation environment—a python library for working with atoms. Journal of Physics: Condensed Matter 29, 273002 (2017).
  42. [42]NVIDIA Corporation. NVIDIA cuEquivariance (2025). https://docs.nvidia.com/cuda/cuequivariancе/.
  43. [43]Ansel, J. et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, DOI: 10.1145/3620665.3640366 (2024).DOI
  44. [44]Jain, A. et al. The materials project: A materials genome approach to accelerating materials innovation. APL Mater (2013).DOI
  45. [45]Najibi, A. & Goerigk, L. The nonlocal kernel in van der waals density functionals as an additive correction: An extensive analysis with special emphasis on the b97m-v and ωb97m-v approaches. Journal of Chemical Theory and Computation 14, 5725–5738 (2018).DOI
  46. [46]Grimme, S., Antony, J., Ehrlich, S. & Krieg, H. A consistent and accurate ab initio parametrization of density functional dispersion correction (dft-d) for the 94 elements h-pu. The Journal of chemical physics 132 (2010).DOI
  47. [47]Grimme, S., Ehrlich, S. & Goerigk, L. Effect of the damping function in dispersion corrected density functional theory. Journal of computational chemistry 32, 1456–1465 (2011).
  48. [48]Weigend, F. & Ahlrichs, R. Balanced basis sets of split valence, triple zeta valence and quadruple zeta valence quality for h to rn: Design and assessment of accuracy. Physical Chemistry Chemical Physics 7, 3297–3305 (2005).
  49. [49]Rappoport, D. & Furche, F. Property-optimized gaussian basis sets for molecular response calculations. The Journal of chemical physics 133 (2010).DOI
  50. [50]Smith, D. G. et al. Psi4 1.4: Open-source software for high-throughput quantum chemistry. The Journal of chemical physics 152 (2020).DOI
  51. [51]Zhang, L. et al. End-to-end symmetry preserving inter-atomic potential energy model for finite and extended systems. Advances in Neural Information Processing Systems 31 (2018).arxiv.org/abs/1805.09003
  52. [52]Wang, H., Guo, X., Zhang, L., Wang, H. & Xue, J. Deep learning inter-atomic potential model for accurate irradiation damage simulations. arXiv preprint arXiv:1904.00360 (2019).DOI
  53. [53]Wen, K. et al. Understanding warmup-stable-decay learning rates: A river valley loss landscape perspective (2024). 2410.05192.arxiv.org/abs/2410.05192
  54. [54]Jordan, K. et al. Muon: An optimizer for hidden layers in neural networks. https://kellerjordan.github.io/posts/muon/ (2024).
  55. [55]Liu, J. et al. Muon is scalable for llm training (2025). 2502.16982.
  56. [56]Joo, T., Xia, W., Kim, C., Zhang, M. & Ie, E. On surprising effectiveness of masking updates in adaptive optimizers (2026). 2602.15322.arxiv.org/abs/2602.15322
  57. [57]Varshalovich, D. A., Moskalev, A. N. & Khersonskii, V. K. Quantum Theory of Angular Momentum (World Scientific, Singapore, 1988).
  58. [58]Dao-AILab. gram-newton-schulz: Fast polar decomposition for muon. https://github.com/Dao-AILab/gram-newton-schulz (2026). Accessed 24 May 2026.
  59. [59]Musaelian, A. et al. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications 14, 579 (2023).
  60. [60]Park, Y., Kim, J., Hwang, S. & Han, S. Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations. Journal of chemical theory and computation 20, 4857–4868 (2024).DOI
  61. [61]Batatia, I. et al. A foundation model for atomistic materials chemistry. arXiv preprint arXiv:2401.00096 (2024).arxiv.org/abs/2401.00096
  62. [62]Li, P. et al. Large-scale ab initio simulations based on systematically improvable atomic basis. Computational Materials Science 112, 503–517 (2016).arxiv.org/abs/1503.00097

Paper details

vlnc:W_pgwukwuu
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents