betaFeedback

WorksPhysics and Astronomy

FACTO: fast atom-centered tensor operator for accurate and efficient foundational machine-learned interatomic potentials

Abstract

High-throughput inference remains a central challenge for equivariant machine-learned interatomic potentials. Here we introduce FACTO, an atom-centered equivariant architecture that confines evolving learned state and dense channel mixing to atoms, expressing neighbor interactions as fully fusible reductions. We complement this design with custom register-accumulated aggregation kernels and tensor-core FP32 emulation, accelerating inference while retaining FP32-level energy conservation under emulated arithmetic in molecular dynamics. Comprehensive evaluations on Matbench Discovery, MLIP Arena and MatCalc demonstrate high throughput alongside strong predictive accuracy and physical reliability. On Matbench Discovery, the four FACTO-OAM variants show top-tier performance by overall rank sum among 46 models with complete scores, advancing the accuracy-throughput Pareto frontier. On MLIP Arena, FACTO achieves state- of-the-art stability under extreme heating and compression, alongside high fidelity in hydrogen combustion. On MatCalc, FACTO outperforms the compared baselines in structural and elastic predictions while retaining competitive accuracy for energetic and thermal properties.

Introduction

The scientific reach of atomistic simulation depends on the fidelity of an interatomic potential and the cost of exploring its energy landscape. Structural relaxation and molecular dynamics require repeated force evaluations, whose cost limits system size, trajectory length and statistical sampling under a fixed computational budget. Machine-learned interatomic potentials (MLIPs) reduce this cost by learning from density-functional theory (DFT) calculations, extending materials modeling across chemical compositions and beyond the scales accessible to direct first-principles simulation[1]. When an MLIP drives relaxation or dynamics, each predicted force determines the subsequent configuration, so force errors can deflect trajectories even when static test-set errors appear small [2]. Practical atomistic simulation therefore requires joint optimization of evaluation speed and physical fidelity across the potential-energy surface.

An MLIP’s representation of local environments around atoms shapes both the interactions it can learn and the operations required to evaluate them. Equivariant neural networks encode directional information in vector and tensor features that are exchanged through pairwise edge operations in each message-passing layer during force evaluation[4, 5, 6]. Previous studies have simplified these tensor interactions and moved equivariant operations onto atoms[7, 8, 9], while complementary work on kernel fusion has reduced the storage of edge intermediates and associated data movement[10]. Yet the model architecture itself limits how much edge memory traffic can be removed: computationally intensive edge operations inevitably materialize large temporary edge tensors within each layer, while learned edge embeddings require persistent state to be carried across layers. Avoiding both bottlenecks motivates retaining geometric state in atom-wise features while making all edge operations fusible.

Here we introduce FACTO, the Fast Atom-Centered Tensor Operator, an equivariant neural network designed to reduce the computational cost and memory traffic of repeated force evaluations. Rather than evaluating explicit Clebsch–Gordan tensor products or dense projections on edges, FACTO encodes directional geometry in atom-centered tensor features and couples scalar and tensor streams through atom-wise projections and dot-product contractions. All evolving learned state resides on atoms, unlike the scalar edge embeddings propagated between layers in Geodite[9]. Once the shared geometric edge encodings are computed, each interaction block avoids both temporary edge intermediates and cross-layer edge state, leaving only lightweight, fully fusible pairwise reductions.

Beyond kernel fusion, we reduce message-aggregation overhead with custom atom-centered kernels that accumulate neighbor contributions without GPU atomic operations. Alongside memory-bound neighbor aggregation, inference throughput is governed by the dense matrix multiplications in atom-wise projections. Standard lower-precision tensor-core formats increase throughput, but their rounding errors can degrade energy and force fidelity during dynamics. We therefore adopt tensor-core FP32 emulation via a three-product bfloat16 decomposition (BF16×3\mathrm{BF16}\times3)[11] to exploit bfloat16 tensor-core throughput while recovering the leading FP32 residual bits. Together, these optimizations target the two primary bottlenecks of equivariant force evaluation: memory-bound neighbor aggregation and compute-bound atom-wise projections. Because both patterns recur across modern MLIPs, the optimizations transfer beyond FACTO.

We evaluate FACTO on Matbench Discovery, MatCalc and MLIP Arena [12, 13, 14] and benchmark large-system inference scaling and microcanonical dynamics in silicon to assess physical fidelity alongside computational cost. The FACTO-OAM family occupies the top four overall rank-sum positions on Matbench Discovery (Figure 1a). Across model sizes, it advances the accuracy–throughput frontier (Figure 1b) and maintains efficient inference as silicon system size grows (Figure 1c). On MatCalc, the family outperforms the compared baselines on structural and elastic benchmarks, with FACTO-OAM-XL also leading formation-energy accuracy while nearly eliminating off-equilibrium force softening. On MLIP Arena, it combines tight force–energy conservation with state-of-the-art survival under extreme heating, high-pressure compression and reactive hydrogen combustion. Our silicon benchmarks further confirm that pairing custom neighbor aggregation with BF16×3\mathrm{BF16}\times3 emulation yields compounding speedups for large systems, while BF16×3\mathrm{BF16}\times3 emulation retains strict-FP32-level energy fluctuations in microcanonical dynamics. Our main contributions are:

Figure 1. Matbench Discovery performance and silicon inference scaling. a, Rank sums across the Matbench Discovery combined performance (CPS), molecular-dynamics (CMDS) and diatomic (CDS) scores (lower is better). The ten lowest rank sums among the four FACTO-OAM variants and 42 public models with complete scores [3] are shown (end markers match the shared legend). b, CPS versus inference throughput for 1,000-atom silicon supercells. Numbers give the FACTO-OAM CPS values, and dashed lines trace separate Pareto frontiers for the FACTO series (blue) and leading public models (grey). c, Inference latency versus atom count for silicon supercells on logarithmic axes, comparing the four FACTO-OAM variants with ten public models. For the FACTO series, curves trace optimized inference scaling, transitioning from standard FP32 AOTI compilation at small system sizes to the fully accelerated pipeline beyond the crossover threshold (see Figure 5 and Methods for details).

  • Atom-centered equivariant architecture with fully fusible pairwise operations. FACTO confines evolving learned state and dense channel mixing to atoms and expresses directional couplings through dot-product feedback and factorized transport, allowing all within-block pairwise operations to be fused into single-pass neighbor reductions.

  • Advancing the accuracy–throughput frontier across atomistic benchmarks. Across Matbench Discovery, MatCalc and MLIP Arena, FACTO delivers high accuracy spanning crystal discovery, crystalline structural and mechanical response, off-equilibrium force fidelity and extreme dynamical stability at substantially higher throughput than existing equivariant potentials.

  • Custom atom-centered aggregation and transferable tensor-core FP32 emulation. We pair register-accumulated neighbor kernels that avoid GPU atomic additions in forward neighbor reductions with three-product bfloat16 FP32 emulation (BF16×3\mathrm{BF16}\times3) to accelerate memory-bound neighbor reductions and compute-bound atom-wise projections, with BF16×3\mathrm{BF16}\times3 preserving trajectory energy conservation in microcanonical dynamics and both optimizations transferring to future MLIPs.

Results

The FACTO architecture

We designed FACTO to keep evolving learned state and dense channel mixing on atoms, expressing neighbor interactions as fusible reductions (Figure 2a). Within each interaction block, four operations couple the scalar and tensor streams: Factorized Scalar Attention aggregates scalar neighbor information to drive both streams, Tensor-to-scalar Feedback returns accumulated angular information to the scalar update, and Geometric Tensor Injection and Tensor Transport create and propagate directional tensor features. Bias terms are omitted from the equations below for clarity.

Figure 2. FACTO architecture and atom-centered execution. a, Coupled scalar and tensor updates in block tt with scaled residual additions. Shared scalar filters ρ\rho, tensor filters ξ\xi, cutoff weights χ\chi and spherical harmonics YY are supplied along dashed arrows. b, Equivalent evaluations of one scalar-attention group using the same atom-wise projections q,k,v\mathbf{q}, \mathbf{k}, \mathbf{v} (group index gg omitted for clarity). Here ψ\psi is SiLU and ⊙\odot denotes channelwise multiplication. Factorization reduces the edge stage to the dashed, fusible neighbor sum. c, Forward aggregation for one output component. In the materialized reference, edge messages are stored and accumulated with GPU atomic additions (indivisible memory updates); in receiver-owned accumulation, they are formed and summed locally before a single ordinary output store, without a GPU atomic read–modify–write update. Neighbor indices and shared geometry remain in memory; the three neighbors are illustrative.

Feature initialization. We represent each structure as a periodic neighbor graph in which each atom ii of atomic number ZiZ_i carries scalar features hi∈RD\mathbf{h}_i \in\mathbb{R}^{D} and tensor features Tiℓ∈R(2ℓ+1)×Cℓ\mathbf{T}_i^\ell\in\mathbb{R}^{(2\ell+1)\times C_\ell} across angular orders 1≤ℓ≤L1 \le\ell\le L. The channel counts CℓC_\ell can vary with ℓ\ell; in particular, we use fewer channels at higher angular orders. We initialize the scalar state by applying a learned affine projection and the SiLU activation ψ\psi to fixed elemental descriptors, whereas the tensor state starts at zero and is populated by the first application of Geometric Tensor Injection.

Each directed edge from neighbor jj to atom ii (including periodic images) is characterized by the displacement vector rij\mathbf{r}_{ij}, distance rij=∥rij∥r_{ij} = \lVert\mathbf{r}_{ij} \rVert and unit direction r^ij=rij/rij\hat{\mathbf{r}}_{ij} = \mathbf{r}_{ij}/r_{ij}. We expand the interatomic distance relative to the interaction radius RcR_{\mathrm{c}} using fixed Gaussian smearing:

uij,n=exp⁡[−(rij/Rc−μn)22sr2],(1)u_{ij,n} = \exp\left[-\frac{\left(r_{ij}/R_{\mathrm{c}}-\mu_n\right)^2}{2s_r^2}\right], \tag*{(1)}

where the centers μn\mu_n and width srs_r are fixed. To smoothly limit the interaction range and suppress learned contributions at very short distances, we combine C3C^3-smooth polynomial cutoffs fRf_R at RcR_{\mathrm{c}} and a short-range radius RsR_{\mathrm{s}}:

fR(r)={1−84(rR)6+216(rR)7−189(rR)8+56(rR)9,0≤r<R,0,r≥R,(2)f_R(r) = \begin{cases} 1 - 84\left(\frac{r}{R}\right)^6 + 216\left(\frac{r}{R}\right)^7 - 189\left(\frac{r}{R}\right)^8 + 56\left(\frac{r}{R}\right)^9, & 0 \le r < R,\\ 0, & r \ge R, \end{cases} \tag*{(2)}

χij=fRc(rij)[1−fRs(rij)]\chi_{ij} = f_{R_{\mathrm{c}}}(r_{ij})[1-f_{R_{\mathrm{s}}}(r_{ij})].

We multiply the Gaussian basis uij\mathbf{u}_{ij} by the cutoff envelope to obtain the shared radial features ϕij=χijuij\boldsymbol{\phi}_{ij} = \chi_{ij}\mathbf{u}_{ij}. We project these features directly, without a radial multilayer perceptron or pair-type embeddings, to obtain the scalar filter ρij\rho_{ij} for Factorized Scalar Attention and the tensor filters ξijℓ\boldsymbol{\xi}_{ij}^{\ell} for Geometric Tensor Injection:

ϕij=χijuij,ρij=χijψ(Wρϕij),ξijℓ=Wξℓϕij,(3)\boldsymbol{\phi}_{ij} = \chi_{ij}\mathbf{u}_{ij}, \qquad\rho_{ij} = \chi_{ij}\psi(W_{\rho}\boldsymbol{\phi}_{ij}), \qquad\boldsymbol{\xi}_{ij}^{\ell} = W_{\xi}^{\ell}\boldsymbol{\phi}_{ij}, \tag*{(3)}

where the outer envelope factor on ρij\rho_{ij} ensures an exact zero cutoff when the linear projection includes a bias term (whereas WξℓW_{\xi}^{\ell} is bias-free). Separately from the radial filters, we evaluate real spherical harmonics Yℓm(r^ij)Y_{\ell m}(\hat{\mathbf{r}}_{ij}) from the unit direction r^ij\hat{\mathbf{r}}_{ij} to encode edge orientation. Both the radial filters and spherical harmonics are computed once per forward pass and reused across all interaction blocks, keeping radial projections outside the repeated message-passing loop.

Coupled RMS normalization. At the start of block tt, we normalize the stored states hi(t)\mathbf{h}_i^{(t)} and Ti(t,ℓ)\mathbf{T}_i^{(t,\ell)} to obtain the block inputs hi\mathbf{h}_i and Tiℓ\mathbf{T}_i^\ell used in the operations below:

νi=1D∑a=1D(hi,a(t))2,ωi,cℓ=12ℓ+1∑m=−ℓℓ(Ti,mc(t,ℓ))2,\nu_i = \frac{1}{D}\sum_{a=1}^{D}\left(h_{i,a}^{(t)}\right)^2, \qquad\omega_{i,c}^{\ell} = \frac{1}{2\ell+1}\sum_{m=-\ell}^{\ell}\left(T_{i,mc}^{(t,\ell)}\right)^2,
hi,a=γahhi,a(t)νi+ϵ,Ti,mcℓ=γcℓTi,mc(t,ℓ)νi+ωi,cℓ+ϵ,(4)h_{i,a} = \frac{\gamma_a^h h_{i,a}^{(t)}}{\sqrt{\nu_i+\epsilon}}, \qquad T_{i,mc}^{\ell} = \frac{\gamma_c^\ell T_{i,mc}^{(t,\ell)}}{\sqrt{\nu_i+\omega_{i,c}^{\ell}+\epsilon}}, \tag*{(4)}

where aa and cc index channels, mm indexes angular components, γah\gamma_a^h and γcℓ\gamma_c^\ell are learned per-channel gains, and ϵ>0\epsilon> 0 stabilizes the denominators. Including the scalar mean square νi\nu_i in the tensor denominator provides a finite normalization scale even when angular features vanish in highly symmetric local environments.

Factorized Scalar Attention. To aggregate scalar neighbor information across GG channel groups of width D/GD/G, we first apply atom-wise linear projections to obtain per-group queries qig=Wqghi\mathbf{q}_i^g = W_q^g\mathbf{h}_i on atom ii and shared keys and values kj=Wkhj\mathbf{k}_j = W_k\mathbf{h}_j and vj=Wvhj\mathbf{v}_j = W_v\mathbf{h}_j on neighbor jj. In each group gg, we couple the query gate ψ(qig)\psi(\mathbf{q}_i^g), the radial filter ρijg\rho_{ij}^g, and the key–value product ψ(kj)⊙vj\psi(\mathbf{k}_j)\odot\mathbf{v}_j. Because the query gate depends only on atom ii and the key–value product depends only on atom jj, we precompute sj=ψ(kj)⊙vj\mathbf{s}_j = \psi(\mathbf{k}_j)\odot\mathbf{v}_j on atoms and factor ψ(qig)\psi(\mathbf{q}_i^g) outside the neighbor sum (Figure 2b):

zig=∑j∈N(i)ψ(qig)⊙ρijg⊙ψ(kj)⊙vj=ψ(qig)⊙∑j∈N(i)ρijg⊙sj,sj=ψ(kj)⊙vj,(5)\begin{aligned} \mathbf{z}_i^g = \sum_{j\in\mathcal{N}(i)} \psi(\mathbf{q}_i^g)\odot\rho_{ij}^g\odot\psi(\mathbf{k}_j)\odot\mathbf{v}_j \\ &= \psi(\mathbf{q}_i^g)\odot\sum_{j\in\mathcal{N}(i)}\rho_{ij}^g\odot\mathbf{s}_j,\qquad\mathbf{s}_j = \psi(\mathbf{k}_j)\odot\mathbf{v}_j, \tag*{(5)} \end{aligned}
xi=ψ(Wx[zi1;…;ziG]),\mathbf{x}_i = \psi\left(W_x[\mathbf{z}_i^1;\ldots;\mathbf{z}_i^G]\right),

where N(i)\mathcal{N}(i) is the neighbor set of atom ii, ⊙\odot denotes channelwise multiplication and brackets denote channel concatenation. Rather than repeating ψ(kj)⊙vj\psi(\mathbf{k}_j)\odot\mathbf{v}_j and ψ(qig)\psi(\mathbf{q}_i^g) on every edge, we evaluate the key–value product sj\mathbf{s}_j and query gate ψ(qig)\psi(\mathbf{q}_i^g) once per atom and apply ψ(qig)\psi(\mathbf{q}_i^g) after aggregation, leaving only radial weighting and summation on the edges.

Tensor-to-scalar Feedback. Tensor-to-scalar Feedback incorporates accumulated angular information into the scalar stream. For each angular order ℓ\ell, we apply atom-wise channel projections UℓU_\ell and VℓV_\ell to the normalized tensor state and contract over the angular index mm via componentwise dot products:

diℓ=∑m=−ℓℓ(TiℓUℓ)m⊙(TiℓVℓ)m,\mathbf{d}^{\ell}_{i}=\sum_{m=-\ell}^{\ell}(\mathbf{T}^{\ell}_{i}U^{\ell})_{m}\odot(\mathbf{T}^{\ell}_{i}V^{\ell})_{m},
fi=Wf[di1;⋯ ;diL],\mathbf{f}_{i}=W_{\mathrm{f}}[\mathbf{d}^{1}_{i};\cdots;\mathbf{d}^{L}_{i}],
Δhi=fi⊙(W1xi)+W2xi,(6)\Delta\mathbf{h}_{i}=\mathbf{f}_{i}\odot(W_{1}\mathbf{x}_{i})+W_{2}\mathbf{x}_{i}, \tag*{(6)}

where a subscript mm selects the channel vector corresponding to angular component mm (with channel indices suppressed). Because these rotationally invariant dot products compare directional features already accumulated on each atom, both the channel projections and angular contractions scale with the number of atoms rather than edges and require no explicit Clebsch–Gordan tensor products.

Geometric Tensor Injection. Geometric Tensor Injection constructs new directional features by coupling the scalar intermediate to edge geometry. We first project this intermediate on each atom as pjℓ=Winjℓxj\mathbf{p}^{\ell}_{j}=W^{\ell}_{\mathrm{inj}}\mathbf{x}_{j} and then modulate it on edges with the precomputed radial filters and spherical harmonics:

Ji,mℓ=∑j∈N(i)(pjℓ⊙ξijℓ)Yℓm(r^ij).(7)\mathbf{J}^{\ell}_{i,m}=\sum_{j\in\mathcal{N}(i)}(\mathbf{p}^{\ell}_{j}\odot\boldsymbol{\xi}^{\ell}_{ij})Y_{\ell m}(\hat{\mathbf{r}}_{ij}). \tag*{(7)}

Because the radial filter ξijℓ\boldsymbol{\xi}^{\ell}_{ij} is precomputed once and shared across blocks, no radial network is evaluated inside the interaction block, and the remaining gather, elementwise product and neighbor sum fuse into atom-centered aggregation.

Tensor Transport. Tensor Transport propagates existing angular information between neighboring atoms. An edgewise formulation would compute invariant alignment scores by contracting incoming tensors with edge spherical harmonics and then mixing channels on every edge:

aijℓ=[∑m=−ℓℓTj,mℓYℓm(r^ij)]Walignℓ.(8)\mathbf{a}^{\ell}_{ij}=\left[\sum_{m=-\ell}^{\ell}\mathbf{T}^{\ell}_{j,m}Y_{\ell m}(\hat{\mathbf{r}}_{ij})\right]W^{\ell}_{\mathrm{align}}. \tag*{(8)}

Instead, exploiting linearity to avoid a dense projection on each edge, we project each atom’s tensor once and reuse it across all of its incident edges:

Ajℓ=TjℓWalignℓ,aijℓ=∑m=−ℓℓAj,mℓYℓm(r^ij).(9)\mathbf{A}^{\ell}_{j}=\mathbf{T}^{\ell}_{j}W^{\ell}_{\mathrm{align}},\qquad \mathbf{a}^{\ell}_{ij}=\sum_{m=-\ell}^{\ell}\mathbf{A}^{\ell}_{j,m}Y_{\ell m}(\hat{\mathbf{r}}_{ij}). \tag*{(9)}

We similarly precompute the projected tensor messages Mjℓ=TjℓWmsgℓ\mathbf{M}^{\ell}_{j}=\mathbf{T}^{\ell}_{j}W^{\ell}_{\mathrm{msg}} and scalar gates WgateℓxjW^{\ell}_{\mathrm{gate}}\mathbf{x}_{j} on atoms before forming the gated neighbor sum:

gijℓ=χij(Wgateℓxj)⊙aijℓ,g^{\ell}_{ij}=\chi_{ij}(W^{\ell}_{\mathrm{gate}}\mathbf{x}_{j})\odot\mathbf{a}^{\ell}_{ij},
Pi,mℓ=∑j∈N(i)gijℓ⊙Mj,mℓ.(10)\mathbf{P}^{\ell}_{i,m}=\sum_{j\in\mathcal{N}(i)}g^{\ell}_{ij}\odot\mathbf{M}^{\ell}_{j,m}. \tag*{(10)}

Because the alignment gate gijℓg^{\ell}_{ij} is rotationally invariant, it preserves the tensor transformation law of Mjℓ\mathbf{M}^{\ell}_{j} while leaving only angular contractions, elementwise products and neighbor reductions on edges.

Atom-centered aggregation. By eliminating dense linear operations from the three pairwise paths—Factorized Scalar Attention, Geometric Tensor Injection and Tensor Transport—we can form and accumulate each neighbor contribution directly at the receiving atom without storing intermediate edge messages (Figure 2c).

Residual updates. At the end of block tt, we update the stored scalar and tensor states with a fixed residual scale α\alpha:

hi(t+1)=hi(t)+αΔhi,\mathbf{h}^{(t+1)}_{i}=\mathbf{h}^{(t)}_{i}+\alpha\Delta\mathbf{h}_{i},
Ti(t+1,ℓ)=Ti(t,ℓ)+α(Jiℓ+Piℓ).(11)\mathbf{T}^{(t+1,\ell)}_{i}=\mathbf{T}^{(t,\ell)}_{i}+\alpha(\mathbf{J}^{\ell}_{i}+\mathbf{P}^{\ell}_{i}). \tag*{(11)}

Graph readout. After all KK interaction blocks, we concatenate the post-block scalar states as hicat=[hi(1),…,hi(K)]\mathbf{h}_i^{\mathrm{cat}} = [\mathbf{h}_i^{(1)}, \ldots, \mathbf{h}_i^{(K)}]. Let hZiso\mathbf{h}_Z^{\mathrm{iso}} denote the corresponding state for an isolated atom of element ZZ, evaluated with the same parameters and no neighbors. We apply a bias-free shared linear projection w\mathbf{w} to define per-atom energy contributions:

Ei=eZiref+δZi+λZiw⊤(hicat−hZiiso),(12)E_i = e_{Z_i}^{\mathrm{ref}} + \delta_{Z_i} + \lambda_{Z_i}\mathbf{w}^{\top}\left(\mathbf{h}_i^{\mathrm{cat}} - \mathbf{h}_{Z_i}^{\mathrm{iso}}\right), \tag*{(12)}

where eZrefe_Z^{\mathrm{ref}} is a fixed elemental reference, λZ\lambda_Z is a learned elemental scale and δZ\delta_Z is a learned elemental shift initialized to zero. The isolated-atom energy is therefore eZref+δZe_Z^{\mathrm{ref}} + \delta_Z; geometry-independent reference terms do not contribute to forces or stress. We obtain the total potential energy by summing these per-atom contributions together with a short-range screened pair-repulsion term ErepE_{\mathrm{rep}} [15] inspired by the implementation in MatGL [16]:

E=∑iEi+Erep,(13)E = \sum_i E_i + E_{\mathrm{rep}}, \tag*{(13)}

from which forces and stress follow as analytical derivatives.

Materials discovery and inference scaling

We first evaluate FACTO on Matbench Discovery [3, 12], which jointly assesses crystal stability and thermal conductivity (CPS), finite-temperature dynamical fidelity (CMDS) [17] and short-range pairwise physicality (CDS). Across the 46 ranked models, comprising the 42 public default-cohort models with complete score triplets plus the four FACTO-OAM variants, the FACTO-OAM family occupies the top four overall positions when ranked by the sum of CPS, CMDS and CDS ranks, ahead of TACE-OAM-L and ORB v3 (Figure 1a and Table 1; component metrics in Appendix Note 3). Whereas models with slightly higher static CPS, such as Prophet-OAME-MBD and EquiformerV3+DeNS-OAM, rank 26 and 24 overall because of lower dynamical and diatomic scores, the FACTO-OAM family maintains high accuracy across all three axes: FACTO-OAM-XL ranks in the top eight on every composite metric, while FACTO-OAM-S and FACTO-OAM-M take the top two positions in CMDS and ranks 3 and 4 in CDS. This accuracy is paired with an order-of-magnitude advantage in inference throughput (Figure 1b,Figure 1c). Along with Prophet-OAME-MBD, the four FACTO-OAM variants define the CPS–throughput Pareto frontier on 1,000-atom silicon (Figure 1b). At the high-accuracy end, FACTO-OAM-XL is more than 15× faster than Prophet-OAME-MBD, EquiformerV3+DeNS-OAM and eSEN-30M-OAM, and roughly 3–4× faster than GRACE-3L-OAM-L and TACE-OAM-L at comparable CPS. Across the faster tiers, FACTO-OAM-L through FACTO-OAM-S achieve throughputs of 64.88 to 149.6 call/s, outpacing lighter baselines including NequIP-OAM-L, GRACE-2L-OAM-L, ORB v3 and MACE-OAM-0 while achieving superior overall rank sums (with FACTO-OAM-L also surpassing all four in CPS and FACTO-OAM-M and FACTO-OAM-S trading a small CPS reduction for higher dynamical scores and speed). This efficiency stems from eliminating edge matrix multiplications and persistent edge state rather than from static graph pruning. At larger system sizes, we further accelerate inference with custom aggregation and tensor-core emulation kernels (§ 2.5). Beyond the smallest 8-atom cell, FACTO maintains lower latency across all evaluated system sizes (Figure 1c).

Table 1

ModelDatasetCPS ↑\uparrowCMDS ↑\uparrowCDS ↑\uparrowSum-of-ranks ↓\downarrowThroughput ↑\uparrow (call/s)
1,000 Si
8,000 Si
FACTO-OAM-XLOAM0.890/70.751/60.771/821/129.84
FACTO-OAM-LOAM0.883/120.773/30.781/722/264.88
FACTO-OAM-MOAM0.863/170.789/20.822/423/3117.3
FACTO-OAM-SOAM0.855/190.792/10.825/323/3149.6
Prophet-OAME-MBD[21]OAME0.912/10.578/430.572/2872/261.924
EquiformerV3+DeNS-OAM[22]OAM0.902/40.638/240.269/4371/241.833
GRACE-3L-OAM-L[23]OAM0.900/50.739/90.699/1731/710.08
TACE-OAM-L[24]OAM0.889/80.728/110.761/928/56.878
eSEN-30M-OAM[25]OAM0.888/90.625/300.648/2463/181.779
NequIP-OAM-XL[26]OAM0.886/110.623/310.796/648/1015.67
NequIP-OAM-L[26]OAM0.870/150.670/150.818/535/842.11
GRACE-2L-OAM-L[23]OAM0.865/160.759/40.694/1939/920.66
ORB v3[27]OAM0.860/180.736/100.834/129/614.29
MACE-MPA-0[28]MPA0.795/260.659/160.728/1355/1219.40
FACTO-MPA-XLMPA0.8670.7560.770—29.84
FACTO-MPA-LMPA0.8590.7750.801—64.88
FACTO-MPA-MMPA0.8470.7930.813—117.3
FACTO-MPA-SMPA0.8320.7900.807—149.6

Table 1. Matbench Discovery scores, ranks and inference throughput. Dataset codes denote training data: O = OMat24[18], M = MPtrj[1], A = Alexandria[19] and E = ELEMENTA[20] (MPA = MPtrj + Alexandria). For the 46 ranked models (four FACTO-OAM variants and 42 public models with complete scores; Figure 1a)[3], CPS, CMDS, CDS and Sum-of-ranks are reported as value/rank; the FACTO-MPA series is listed below without ranks. Throughput is measured on 1,000- and 8,000-atom silicon supercells (Figure 1b); OOM denotes out of memory. Bold values mark the best ranked and overall results in each column. Component metrics and the full public cohort are reported in Appendix Note 3.

Trained from scratch on MPtrj and subsampled Alexandria (sAlex) without OMat24 pretraining, the FACTO-MPA series also achieves strong performance across all three benchmarks at the same high throughput (Table 1, bottom block). On the same training corpus, both FACTO-MPA-XL and FACTO-MPA-L substantially outperform MACE-MPA-0 across CPS, CMDS, CDS and speed, and match or surpass OAM-pretrained baselines such as GRACE-2L-OAM-L and ORB v3 in CPS, while FACTO-MPA-M and FACTO-MPA-S closely match their OAM counterparts on CMDS and CDS. Comparing the two families indicates that, with MPA training alone, the models produce smooth molecular-dynamics trajectories and diatomic curves, whereas OMat24 pretraining primarily improves equilibrium stability and anharmonic phonon transport (reducing κSRME\kappa_{\mathrm{SRME}} from 0.2218 in FACTO-MPA-XL to 0.1560 in FACTO-OAM-XL; Appendix Table A5).

Within both families, scaling from L to XL brings a modest gain in CPS while slightly lowering CMDS and CDS, reflecting the runtime subsurface penalty at larger model sizes and higher diatomic force-reversal counts on steep repulsive walls (Appendix Tables A6 and A8). This plateau reflects both diminishing architectural returns once Tensor Transport and ℓ≤4\ell\le4 equivariant channels are active in FACTO-L and our use of a uniform training recipe across all four model sizes without size-specific hyperparameter or loss-weight tuning (Methods and Appendix Note 1). Achieving rank-1 overall standing and leading throughput under this fixed recipe underscores that FACTO’s performance stems from its atom-centered equivariant design while leaving scope for training-recipe optimization at the largest scale (Discussion).

Out-of-equilibrium stability and reactive dynamics

To probe physical reliability far from equilibrium, we evaluate the FACTO-OAM family on MLIP Arena [14, 29], where it achieves state-of-the-art performance across pairwise potentials, bulk deformation scans, extreme thermodynamic stress tests and reactive molecular dynamics (Figure 3; full diagnostics in Appendix Note 3). On static diatomic scans (Figure 3a), all four FACTO-OAM variants have low finite-grid force–energy conservation deviations of 0.01360–0.01707 eV A˚−1\mathrm{\AA}^{-1}, more than 4×4 \times smaller than those of MACE-MPA and MACE-MP(M) [28] and orders of magnitude smaller than those of ORBv2[30], CHGNet[1] and SevenNet[31] (Appendix Table A11). For volumetric crystal deformations across all 1,000 WBM structures (Figure 3b), both relaxed equation-of-state and fixed-coordinate energy–volume runtimes remain near unity (1.000–1.011) with 100% completion (Appendix Tables A14 and A15).

Figure 3. Out-of-equilibrium physicality, dynamical stability and reactive dynamics on MLIP Arena. a, Mean diatomic force reversals versus finite-grid force–energy conservation deviation on a logarithmic x-axis; the dashed line marks the ideal reversal count of one. b, Tortuosity for relaxed equation-of-state (EOS) and fixed-coordinate energy–volume (E–V) curves. Both axes are linear; values closer to one indicate smoother curves. c, Area under the heating and compression survival curves (AUC); dashed lines mark the discrete 100-bin scorer ceiling of 65/66 ≈ 0.9848 (continuous ideal of one). Coincident FACTO points (FACTO-OAM-XL, FACTO-OAM-L and FACTO-OAM-S at (0.9848, 0.9848)) are shown with nested symbols at their measured coordinates. d, Water yield versus absolute potential-energy reaction-enthalpy residual in hydrogen combustion on a logarithmic x-axis. For CHGNet and M3GNet, the published combustion endpoints are from partial trajectories. Public baseline values are from the MLIP Arena leaderboard[29]; full FACTO-OAM diagnostics are reported in Appendix Note 3.

This smooth potential energy surface translates directly into trajectory stability in the molecular-dynamics stress benchmark (Figure 3c), which measures the area under the finite-energy survival curve (AUC) during heating ramps from 300 to 3,000 K and compression ramps up to 100 GPa. Most public baselines exhibit substantially reduced trajectory survival under high-pressure compression: SevenNet, MACE-MP(M), M3GNet[33] and CHGNet fall to compression AUCs of 0.1550–0.3998, ORBv2 drops from 0.9848 in heating to 0.7725 in compression, and MACE-MPA (0.9635 and 0.8986) and MatterSim[34] (0.9511 and 0.9207) also fall short of the benchmark’s discrete 100-bin ceiling (65/66 ≈ 0.9848). By contrast, all four FACTO-OAM variants cluster at the upper-right of the plot, with FACTO-OAM-XL, FACTO-OAM-L and FACTO-OAM-S reaching the 0.9848 ceiling in both heating and compression AUC and FACTO-OAM-M recording 0.9799 and 0.9848, surpassing every compared baseline in joint thermal and barostat robustness across all 400 trajectories (Appendix Table A12).

This dynamical stability extends to the 1 ns (2,000,000-step) reactive hydrogen combustion benchmark on 384 atoms (128 H2_2 + 64 O2_2; Figure 3d). Among the public baselines, trajectories evaluated with CHGNet and M3GNet terminate prematurely (26.30% and 22.70% completion) with large enthalpy residuals (> 55 kcal mol−1^{-1}), whereas the ORBv2 trajectory reaches completion without ignition (0% water yield and a 227.5 kcal mol−1^{-1} residual). With MatterSim and SevenNet, water yields are moderate (75.00% and 71.09%), with residuals of 13.38 and 16.70 kcal mol−1^{-1}. By contrast, for all four FACTO-OAM variants, the reactive trajectories reach completion in the high-yield, low-residual regime alongside MACE-MPA and MACE-MP(M), achieving 85.94–90.10% water yield, with FACTO-OAM-M reaching the highest yield (90.10%) among all compared models and FACTO-OAM-L reaching 88.80%. Across the series, we also observe small reaction-enthalpy residuals (1.878–5.288 kcal mol−1^{-1}, down to 1.878 kcal mol−1^{-1} for FACTO-OAM-L) and negligible center-of-mass drift (0.001247–0.001764 Å), while sustaining rollout throughputs of 45.41–157.7 steps s−1^{-1} (Appendix Table A13).

Crystalline elastic, thermal and force response

We next evaluate crystalline property prediction on MatCalc [13, 35, 36], comparing the FACTO-OAM series against five published baselines (NequIP-OAM-XL and NequIP-OAM-L [4, 26], Allegro-OAM-L [6, 26], MatterSim-5M [34] and UMA-s [37]) across equilibrium structure, energetics and continuum elasticity (Figure 4a–Figure 4d; full values in Appendix Note 3). Despite operating at substantially higher inference throughput than competing equivariant models (Table 1), the FACTO-OAM family matches or surpasses the baselines across these observables. All four FACTO-OAM variants have lower errors than every published baseline in relaxed CrystalNN structural-fingerprint distance (0.1959–0.2291 versus 0.2670–0.3220; Figure 4a) as well as Voigt–Reuss–Hill bulk-modulus (11.05–12.58 versus 12.70–16.90 GPa; Figure 4c) and shear-modulus (12.04–13.54 versus 14.10–21.50 GPa, led by FACTO-OAM-L at 12.04 GPa; Figure 4d) MAEs. On formation energy (Figure 4b), FACTO-OAM-XL achieves the lowest error among all compared models (22.35 meV atom−1^{-1}), while FACTO-OAM-M, FACTO-OAM-L and FACTO-OAM-S (24.73–26.06 meV atom−1^{-1}) perform on par with Allegro-OAM-L, UMA-s, NequIP-OAM-XL and NequIP-OAM-L (24.70–26.60 meV atom−1^{-1}) and substantially ahead of MatterSim-5M (40.50 meV atom−1^{-1}).

Figure 4. Crystalline-property accuracy and off-equilibrium force amplitudes. a–e, CrystalNN structural-fingerprint distance, formation-energy MAE, bulk- and shear-modulus MAEs, and 300 K harmonic heat-capacity MAE; lower values indicate smaller errors. f, Mean proportional slope between predicted and DFT force components; the dashed line marks the ideal value of one. Filled circles, diamonds, squares and triangles denote FACTO-OAM-XL, FACTO-OAM-L, FACTO-OAM-M and FACTO-OAM-S, respectively; open circles denote published baselines. Top annotations give the FACTO-OAM-XL values. Each panel uses a separately labeled linear scale. Published baseline values are taken from ref. 37. Complete property values are reported in Appendix Note 3.

Comparable accuracy extends to vibrational thermodynamics and off-equilibrium force calibration (Figure 4e,Figure 4f). For 300 K harmonic heat capacity across 1,170 crystals (Figure 4e), where finite-displacement force constants are sensitive to small fluctuations in predicted forces, FACTO-OAM-XL (4.397 J mol−1^{-1} K−1^{-1}) and FACTOOAM-L (4.801 J mol−1 K−14.801\ \mathrm{J\,mol^{-1}\,K^{-1}}) perform comparably to Allegro-OAM-L (4.380 J mol−1 K−14.380\ \mathrm{J\,mol^{-1}\,K^{-1}}) and NequIP-OAM-L (4.490 J mol−1 K−14.490\ \mathrm{J\,mol^{-1}\,K^{-1}}), while trailing UMA-s (3.600 J mol−1 K−13.600\ \mathrm{J\,mol^{-1}\,K^{-1}}) and NequIP-OAM-XL (4.020 J mol−1 K−14.020\ \mathrm{J\,mol^{-1}\,K^{-1}}). Conversely, on high-energy WBM configurations (Figure 4f), FACTO-OAM-XL (0.9861) and FACTO-OAM-L (0.9847) closely approach the ideal force-calibration slope of unity, nearly eliminating systematic force softening. Even the fastest variants, FACTO-OAM-S and FACTO-OAM-M, which run 2.785–9.546× faster than NequIP-OAM-L and NequIP-OAM-XL on 1,000-atom silicon (Table 1), outperform both NequIP-OAM models on structural relaxation and elastic moduli while matching their formation-energy and force-slope accuracy.

Accelerating FACTO inference

Because FACTO separates matrix-multiplication-free neighbor reductions from dense atom-wise channel mixing, its inference can be accelerated through two orthogonal optimizations in addition to PyTorch ahead-of-time Inductor (AOTI) compilation[38]. First, although pointwise edge operations are fused across the unified energy, force and stress graph in AOTI, standard scatter reductions still incur global-memory traffic and contention from GPU atomic additions. Our custom atom-centered forward kernels avoid materializing intermediate edge messages and eliminate GPU atomic additions to the aggregation outputs: neighbor contributions are accumulated in registers, and each output component is written once with an ordinary store (Figure 2c). Neighbor indices, shared geometry and final outputs still require global-memory accesses. Second, to accelerate the dense atom-wise linear projections in both the forward energy evaluation and backward differentiation for forces and stress while retaining near-single-precision numerical fidelity, we apply tensor-core FP32 emulation via a three-product bfloat16 decomposition (BF16×3) that recovers the leading FP32 residual terms[11] (Appendix Note 2).

Targeting memory-bound neighbor aggregation and compute-bound node projections together yields compounding speedups as the simulated system grows (Figure 5a,Figure 5b). On the smallest diamond-silicon supercells (N≤216N \le216 atoms), kernel-launch and operand-splitting overheads outweigh the savings from custom aggregation and BF16×3. Once the workload saturates the GPU (N≥1,000N \ge1{,}000 atoms), combining both optimizations achieves the lowest latency at every system size, with custom aggregation also reducing latency on its own for N≥4,096N \ge4{,}096 atoms. For the wide (D=2,048D = 2{,}048, L=6L = 6) FACTO-XL backbone, where dense atom-wise projections dominate runtime, BF16×3 alone delivers a 1.339× speedup at 21,952 atoms, and adding custom receiver accumulation increases the combined speedup to 1.441× (reducing calculator-call latency from 546.3 to 379.1 ms).

Figure 5. Faster large-system evaluation with controlled energy fluctuations. a, FACTO-XL inference latency for diamond-silicon supercells. b, Speedup relative to plain FP32 AOTI. The combination of custom aggregation and FP32 emulation (labeled BF16×3) is compared with each optimization alone; the annotation marks the peak combined speedup (1.441× at 21,952 atoms). c,d, Signed total-energy drift and detrended RMS fluctuation for three 500 ps trajectories per arithmetic mode using FACTO-OAM-M, 64-atom silicon and a 1 fs timestep. Each point represents an independent trajectory. Numbers in c give the mean absolute drift across the three trajectories for each arithmetic mode. Short horizontal bars and numbers in d show the geometric means. Under FP32 emulation (BF16×3), fluctuation amplitudes remain close to those of FP32, whereas they are substantially increased under TF32. Total-energy traces are shown in Appendix Fig. A1.

Tensor-core FP32 emulation avoids the dynamical degradation caused by standard reduced-precision tensor-core arithmetic (Figure 5c,Figure 5d). Across three 500 ps microcanonical (NVE) simulations of 64-atom silicon with FACTO-OAM-M (Δt=1\Delta t = 1 fs), standard TF32 arithmetic inflates the geometric-mean detrended total-energy fluctuation by 7.424× relative to strict FP32 (31.74 versus 4.276 μ\mueV atom−1^{-1}) and increases the mean absolute energy drift by 15.0× (0.1727 versus 0.01151 μ\mueV atom−1^{-1} ps−1^{-1}; Appendix Table A4 and Appendix Fig. A1). By recovering the leading residual bits of each matrix product, BF16×3 restores the geometric-mean detrended energy fluctuation to the strict-FP32 level (4.141 versus 4.276 μ\mueV atom−1^{-1}) while keeping the mean absolute drift in the 10−210^{-2} μ\mueV atom−1^{-1} ps−1^{-1} regime (0.02695 μ\mueV atom−1^{-1} ps−1^{-1}). Tensor-core FP32 emulation thus provides a transferable path to accelerate dense equivariant projections while preserving the energy conservation required for long-horizon molecular dynamics.

Discussion

We introduced FACTO, an equivariant interatomic potential that confines evolving learned state and dense channel mixing to atoms while factoring pairwise interactions into lightweight, fully fusible neighbor reductions. By pairing this atom-centered architecture with custom register-accumulated neighbor kernels and three-product bfloat16 tensor-core FP32 emulation (BF16×3), we accelerate both memory-bound edge aggregation and compute-bound node projections while retaining strict-FP32-level energy fluctuations under BF16×3 in microcanonical dynamics. Across Matbench Discovery, MLIP Arena and MatCalc, the FACTO family advances the accuracy–throughput frontier, combining leading positions in the overall Matbench Discovery rank-sum comparison and state-of-the-art dynamical and reactive stability on MLIP Arena with competitive crystalline-property predictions at substantially higher throughput than existing equivariant potentials.

Several limitations suggest directions for future work. First, wide atom-wise projections and high angular orders (DD up to 2,048 and LL up to 6) increase parameter counts and per-atom memory footprints in our largest models, restricting single-GPU evaluation of FACTO-L and FACTO-XL to smaller maximum atom counts than FACTO-S and FACTO-M (Figure 1c). Parameter sharing, low-rank node projections or multi-GPU domain decomposition will be needed to extend the largest backbones to larger systems. Second, computational constraints limited training-recipe exploration. We used the same optimizer hyperparameters across model scales, concatenated sAlex and MPtrj without dataset reweighting, adopted a two-stage conservative OAM pipeline without auxiliary denoising objectives, and fixed the energy, force and stress loss weights at 1 : 1 : 1. Systematic recipe exploration should further improve accuracy at a fixed inference cost. Third, scaling from FACTO-L to FACTO-XL yields diminishing returns: modest gains in static discovery, formation energy and bulk modulus alongside a slight increase in shear-modulus error at higher latency. Distinguishing training-recipe saturation from the intrinsic limits of semi-local message passing remains an open challenge for foundation interatomic potentials.

Looking forward, co-designing equivariant architectures, arithmetic schemes and hardware-level execution opens promising avenues for atomistic modeling at unprecedented scales. By eliminating the memory bottlenecks of pairwise edge tensors and preserving numerical fidelity under hardware-accelerated emulation, the atom-centered paradigm provides a natural foundation for multi-GPU spatial domain decomposition, making million-atom simulations and long-timescale investigations of defect kinetics, crack propagation and phase transitions accessible with quantum accuracy. Beyond crystalline materials, extending these efficient equivariant operators to multiphase interfaces, amorphous electrolytes and complex chemical reaction networks could substantially broaden their physical domain. Ultimately, pairing high-throughput foundational potentials with autonomous active-learning workflows and robotic synthesis offers a practical path toward accelerating scientific discovery across diverse materials spaces.

Methods

Model specification

We implement FACTO using the model settings in Appendix Table A1. Tensor Transport is omitted in the small and medium backbones (FACTO-S and FACTO-M) and enabled in the large and extra-large backbones (FACTO-L and FACTO-XL), whereas Geometric Tensor Injection and Tensor-to-scalar Feedback are retained in all four. Only scalar features are updated in the final interaction block. To construct periodic neighbor graphs, we include periodic images of atoms within 6 Å and exclude zero-distance self-interactions. The neighbor limit is given in Appendix Table A1.

We initialize scalar features by applying an affine projection and SiLU to fixed elemental descriptors. For the Gaussian radial basis, we use centers μn=(n−1)/(Nr−1)\mu_n = (n - 1)/(N_r - 1) for n=1,…,Nrn = 1, \ldots, N_r, with Nr=32N_r = 32 for FACTO-S and Nr=64N_r = 64 for FACTO-M, FACTO-L and FACTO-XL. Real spherical harmonics are normalized according to the integral convention ∫S2∣Yℓm∣2 dΩ=1\int_{S^2} \lvert Y_{\ell m}\rvert^2 \,\mathrm{d}\Omega= 1.

Training data and optimization

We train FACTO-MPA from random initialization on the MPtrj dataset released with CHGNet [1] and subsampled Alexandria (sAlex). For FACTO-OAM, we first pretrain the model on OMat24 and then finetune it on the MPtrj–sAlex mixture [18, 19]. During finetuning, we retain the pretrained architecture and matching parameters but restart the optimizer and learning-rate schedule. We concatenate MPtrj and sAlex without additional dataset balancing.

Fitting and validation partitions are kept separate during OMat24 pretraining. In MPA training and OAM finetuning, the validation partitions are included in the training set and used to monitor training loss. We use the final exponential-moving-average (EMA) checkpoint for evaluation. Structures are randomly permuted and packed by atom count up to a per-device limit; larger structures are omitted. Appendix Note 1 specifies partitions, batching and stage hyperparameters.

Fixed elemental references are subtracted before evaluating the energy loss and restored in physical predictions. Isolated-atom references are used in MPA training and OAM finetuning, whereas OMat24 references are used in pretraining. For BB structures with atom counts NbN_b, let A\mathcal{A} denote the set of atoms not marked as fixed. With prediction-minus-target differences denoted by Δ\Delta, the objective is

L=wELE+wFLF+wσLσ,\mathcal{L}=w_E\mathcal{L}_E+w_F\mathcal{L}_F+w_\sigma\mathcal{L}_\sigma,
LE=1B∑b=1B∣ΔEb∣Nb,LF=1∣A∣∑i∈A∥ΔFi∥2,\mathcal{L}_E=\frac{1}{B}\sum_{b=1}^{B}\frac{|\Delta E_b|}{N_b},\qquad\mathcal{L}_F=\frac{1}{|\mathcal{A}|}\sum_{i\in\mathcal{A}}\|\Delta\mathbf{F}_i\|_2,
Lσ=19B∑b=1B∑u,v=13∣Δσb,uv∣.(14)\mathcal{L}_\sigma=\frac{1}{9B}\sum_{b=1}^{B}\sum_{u,v=1}^{3}|\Delta\sigma_{b,uv}|. \tag*{(14)}

We evaluate the loss terms in eV for energy, eV A˚−1\mathring{\mathrm{A}}^{-1} for force and eV A˚−3\mathring{\mathrm{A}}^{-3} for stress. We average absolute energy errors per atom across structures, force-vector error norms across free atoms and absolute errors across all nine stress components. In distributed training, loss reductions are normalized by the global counts of structures, free atoms and stress components. Loss weights and shared optimizer settings are given in Appendix Table A3.

We use Muon[39] for hidden weight matrices and AdamW[40] for the elemental input projection, embeddings, energy head and non-matrix parameters, similar to the hybrid optimizer setup in DPA4 [41]. In Muon, Nesterov momentum, Newton–Schulz orthogonalization and matrix-size scaling [42] are combined, and updates are damped using an exponential moving average of momentum–gradient alignment[41, 43] (Appendix Note 1). Biases, normalization gains, embeddings and elemental readout scales and shifts are excluded from weight decay. In each training stage, we use a learning-rate schedule with linear warmup, a constant-rate phase and cosine decay, together with gradient clipping and an exponential moving average of model parameters.

Benchmark definitions and comparisons

Matbench Discovery. We use the benchmark’s default combined performance score (CPS), molecular-dynamics score (CMDS) and diatomic score (CDS)[3]. CPS combines stability-classification F1 on unique WBM prototypes[32], StructureMatcher RMSD[44] and thermal-conductivity symmetric relative mean error[12, 45]. CMDS combines discrepancies in vibrational spectra, angular distributions and pressure distributions from DynaMat trajectories with runtime [17]. CDS combines energy/force accuracy, diatomic geometry, curve shape and runtime. Appendix Note 3 reports the component subscores and the full public comparison cohort.

For Figure 1a and Table 1, we summarize overall standing by the sum of CPS, CMDS and CDS ranks. For each metric, we rank models with complete CPS, CMDS and CDS scores in the public default cohort accessed on 22 October 2026 together with the four FACTO-OAM variants. We order the sums of the three ranks from smallest to largest. The resulting comparison contains 42 public models and four FACTO-OAM variants (46 complete models). Table 1 reports the same 46-model ranks alongside unranked FACTO-MPA scores.

MatCalc. We evaluate equilibrium formation energy and structural-fingerprint distance (972 materials[32]), Voigt–Reuss–Hill bulk and shear moduli (3,953 materials[46]), harmonic heat capacity at 300 K (1,170 materials[47, 48]), and force-softening slope (979 materials)[13, 35, 36, 49]. Formation energies are evaluated relative to elemental references, and the dimensionless structural-fingerprint distance is computed from CrystalNN environment descriptors[50]. We compute mean absolute errors (MAEs) in the predicted properties across all successful evaluations. Full property values and baseline comparisons are reported in Appendix Note 3.

MLIP Arena. We compare diatomic, equation-of-state, energy–volume, heating/compression and hydrogen-combustion diagnostics[14, 29]. We evaluate conservation deviation by comparing predicted forces with numerical energy derivatives; tortuosity measures excess curve variation, and the force-flip count has an ideal value of one. The area under the stability curve (AUC) measures finite-energy survival (with a discrete 100-bin scorer ceiling of 65/66 ≈0.9848\approx0.9848 for complete survival), and the combustion residual serves as an operational potential-energy diagnostic. Appendix Note 3 reports the complete set of diagnostics across these tasks.

Inference measurements and numerical fidelity

Acceleration ablation. We compare plain FP32 AOTI, custom aggregation, BF16×\times3 and their combination for the FACTO-XL model. For each system size, we perform five warmup calls followed by 20 timed evaluations.

Multi-model speed comparison. Fourteen implementations (the four FACTO backbones and ten public baselines) are timed at the Atomic Simulation Environment (ASE) [51] calculator-call boundary using silicon supercells with 8n38n^3 atoms, n=1,…,20n = 1, \ldots, 20. We report mean latencies and standard deviations from direct ASE calculator-call measurements. For FACTO, we report plain FP32 AOTI throughputs at 1,000 atoms (Figure 1b and Table 1) and throughputs obtained with custom aggregation and BF16×\times3 at 8,000 atoms (Table 1). The scaling curves in Figure 1c switch from plain FP32 AOTI to custom aggregation with BF16×\times3 at each backbone’s measured crossover size (N≥512N \ge512, 1,728, 2,744 and 4,096 atoms for FACTO-XL, FACTO-L, FACTO-M and FACTO-S, respectively). Runtime dispatch overhead is excluded. OAM and MPA models of the same size have identical architectures and inference latencies. For each public model, we use the fastest available implementation in the speed comparison. Model-specific precision, acceleration backends, compilation settings and software versions for the public comparators are listed in Appendix Table A16.

Silicon NVE trajectories. To isolate the effect of matrix-multiplication arithmetic on trajectory conservation, we evaluate strict FP32, BF16×\times3 and TF32 using the same FACTO-OAM-M checkpoint with custom aggregation disabled. For each arithmetic mode, we perform three independent runs of a 64-atom diamond-silicon supercell with a lattice constant of 5.43 Å. In each run, we equilibrate the system for 2 ps at a target temperature of 300 K using a Langevin thermostat with a friction coefficient of 0.01 fs−1^{-1} under the corresponding arithmetic mode, then collect 500 ps of fixed-cell velocity-Verlet NVE dynamics at a 1 fs timestep. Energy is sampled every ten steps, including both endpoints.

For atom count NN and Ns=50,001N_{\mathrm{s}} = 50{,}001 samples of total energy per atom uk=(Epotential,k+Ekinetic,k)/Nu_k = (E_{\mathrm{potential},k} + E_{\mathrm{kinetic},k})/N at time tkt_k, we fit uk=a+btku_k = a + bt_k by ordinary least squares and report the signed drift bb and the detrended root-mean-square (RMS) fluctuation:

RMSdetrended=[1Ns∑k=1Ns(uk−a−btk)2]1/2.(15)\mathrm{RMS}_{\mathrm{detrended}} = \left[\frac{1}{N_{\mathrm{s}}}\sum_{k=1}^{N_{\mathrm{s}}}(u_k-a-bt_k)^2\right]^{1/2}. \tag*{(15)}

Both quantities are evaluated over all NsN_{\mathrm{s}} samples, and we summarize the three detrended RMS values for each mode by their geometric mean. We plot total-energy changes without subtracting the fitted drift.

Author contributions

S.L. led the project and contributed to all aspects of the work, including designing and implementing the model and acceleration methods, conducting the experiments and writing the manuscript. C.H. implemented acceleration methods and the training and evaluation pipeline. S.G. designed an earlier version of the model. Y.Y. implemented acceleration methods. T.Z., Y.X., K.Y. and S.X. contributed to discussions. W.Y. contributed to model design and project administration.

Appendix

Model and training specifications

Appendix Tables A1–A3 specify the architecture and training hyperparameters.

Benchmark-model hyperparameters. CℓC_{\ell} is the number of channels for angular order ℓ\ell, starting at ℓ=1\ell= 1. Tensor Transport denotes the optional transport increment; when it is disabled, the tensor stream, Geometric Tensor Injection and Tensor-to-scalar Feedback are retained. Only scalar features are updated in the final block.

SettingXLLMS
Scalar width DD2,0481,024768384
Attention groups GG4463
Tensor channels (C1,…,CL)(C_1, \ldots, C_L)(512, 256, 128, 64, 32, 16)(256, 128, 64, 32)(128, 64, 32, 16)(128, 64, 32, 16)
Maximum angular order LL6444
Radial basis size NrN_r64646432
Tensor Transport enabledYesYesNoNo
Interaction blocks KK5
Learned cutoff RcR_c6 Å
Short-range radius RsR_s1 Å
Repulsive cutoff RrepR_{\mathrm{rep}}6 Å
Maximum neighbors per atom65,536
Gaussian width srs_r2.5/(Nr−1)2.5/(N_r - 1) (fixed)
Residual increment scale α\alpha40−1/240^{-1/2} (fixed)

Table A1.

Stage-specific training hyperparameters. The three stages are MPA training from scratch, OMat24 pretraining and OAM finetuning. The same stage recipes apply to all four FACTO backbones (S, M, L and XL). The atom limits are global.

SettingMPA trainingOMat24 pretrainingOAM finetuning
DataMPtrj + sAlexOMat24MPtrj + sAlex
InitializationRandomRandomPretrained checkpoint
Maximum epochs1002040
Maximum atoms32,00064,00016,000
Muon base LR2×10−32 \times10^{-3}2×10−32 \times10^{-3}3×10−43 \times10^{-4}
AdamW base LR5×10−45 \times10^{-4}5×10−45 \times10^{-4}7×10−57 \times10^{-5}
Weight decay10−210^{-2}10−210^{-2}3×10−43 \times10^{-4}
Final decay fraction fdf_d0.350.500.35
TF32 matmul in trainingNoYesNo

Table A2.

SettingValue
Loss weights (wE,wF,wσw_E, w_F, w_\sigma)(1,1,1)(1, 1, 1)
Muon momentum0.99; Nesterov enabled; no momentum warmup
Newton–Schulz iterations5
Update multiplier γopt\gamma_{\mathrm{opt}}0.18
Alignment dampingEnabled; smin⁡=0.1s_{\min}=0.1, τ=2\tau=2, βa=0.9\beta_a=0.9
AdamW (β1,β2)(\beta_1,\beta_2) / ϵ\epsilon(0.9,0.95)(0.9, 0.95) / 10−1010^{-10}
Warmup fraction0.01 of total steps
Initial / final LR multiplier0.2 / 0.01
Gradient clipping value10
Parameter EMA decay0.999
torch.compile in trainingYes

Table A3. Training settings shared across the three stages. Weight decay in Appendix Table A2 applies to the eligible Muon and AdamW groups. The learned elemental energy scale and shift are excluded from weight decay. EMA denotes exponential moving average.

Elemental descriptors and references. The fixed elemental descriptor has 31 components encoding scaled elemental properties and chemical-group indicators. MPtrj isolated-atom DFT energy references are used in MPA training and OAM finetuning, whereas OMat24 references are used in pretraining. We obtain the MPtrj values from Kaplan’s isolated-atom energy dataset [52] and the OMat24 values from the FAIRChem repository [53].

Muon alignment damping. We apply continuous damping to Muon [39] updates using smoothed momentum–gradient cosine alignment, similar to the deterministic alignment damping in DPA4 [41] and the alignment-based update modulation in Magma [43]. For a matrix parameter, let gtg_t be its gradient and let Mt=μMt−1+gtM_t=\mu M_{t-1}+g_t, M0=0M_0=0, be its momentum buffer. The Nesterov input to Newton–Schulz is gt+μMtg_t+\mu M_t. We calculate

ζt=βaζt−1+(1−βa)max⁡(0,⟨Mt,gt⟩Fmax⁡(∥Mt∥F,ϵa)max⁡(∥gt∥F,ϵa)),δt=smin⁡+(1−smin⁡)ζtτ,(16)\begin{aligned} \zeta_t &= \beta_a\zeta_{t-1}+(1-\beta_a)\max\left(0,\frac{\langle M_t,g_t\rangle_F}{\max(\lVert M_t\rVert_F,\epsilon_a)\max(\lVert g_t\rVert_F,\epsilon_a)}\right),\\ \delta_t &= s_{\min}+(1-s_{\min})\zeta_t^\tau, \tag*{(16)} \end{aligned}

with ζ0=1\zeta_0=1 and ϵa=10−12\epsilon_a=10^{-12}. We multiply the orthogonalized update by δtγoptmax⁡(nrow,ncol)\delta_t\gamma_{\mathrm{opt}}\sqrt{\max(n_{\mathrm{row}},n_{\mathrm{col}})} before applying the learning rate (similar to the matrix-size scaling in refs. [41, 42]), where nrown_{\mathrm{row}} and ncoln_{\mathrm{col}} are the numbers of rows and columns of the parameter matrix. Alignment is evaluated in FP32. Decoupled weight decay is applied separately and is not multiplied by δt\delta_t.

Learning-rate schedule. We use a warmup–stable–decay (WSD) schedule [54]. For TT optimizer steps, the warmup lasts Tw=⌈0.01T⌉T_w=\lceil0.01T\rceil steps and the final cosine decay lasts Td=⌈fdT⌉T_d=\lceil f_dT\rceil, where fdf_d is the stage-specific fraction in Appendix Table A2. Between these phases, the learning-rate multiplier is one. During warmup, the multiplier at optimizer step tt is 0.2+0.8t/Tw0.2+0.8t/T_w; during decay, it is

0.01+0.992[1+cos⁡(πt−(T−Td)Td)].(17)0.01+\frac{0.99}{2}\left[1+\cos\left(\pi\frac{t-(T-T_d)}{T_d}\right)\right]. \tag*{(17)}

The scheduler is updated at each optimizer step over the configured training horizon.

Dataset partitions and validation reuse. In MPA training and OAM finetuning, we concatenate the configuration-level MPtrj fitting partition with sAlex without dataset reweighting. In OMat24 pretraining, we use the official training, validation and test splits. Atom-count batching is used across all stages.

Appendix Note 2: Acceleration and numerical fidelity

Three-product bfloat16 FP32 emulation (BF16×3). Let B\mathcal{B} round an FP32 matrix to BF16. For operands X and W, we split off the leading residual terms[11]:

Xhi=B(X),Xlo=B(X−Xhi),Whi=B(W),Wlo=B(W−Whi).(18)\begin{aligned} \mathrm{X}_{\mathrm{hi}} &= \mathcal{B}(\mathrm{X}), &\quad\mathrm{X}_{\mathrm{lo}} &= \mathcal{B}(\mathrm{X}-\mathrm{X}_{\mathrm{hi}}), \\ \mathrm{W}_{\mathrm{hi}} &= \mathcal{B}(\mathrm{W}), &\quad\mathrm{W}_{\mathrm{lo}} &= \mathcal{B}(\mathrm{W}-\mathrm{W}_{\mathrm{hi}}). \tag*{(18)} \end{aligned}

Residual subtraction and accumulation are performed in FP32, and three tensor-core products are used to approximate the original multiplication:

XW≃XhiWhi+XhiWlo+XloWhi.(19)\mathrm{XW} \simeq\mathrm{X}_{\mathrm{hi}}\mathrm{W}_{\mathrm{hi}}+\mathrm{X}_{\mathrm{hi}}\mathrm{W}_{\mathrm{lo}}+\mathrm{X}_{\mathrm{lo}}\mathrm{W}_{\mathrm{hi}}. \tag*{(19)}

The residual–residual product is omitted. We apply this FP32 emulation to matrix multiplications during energy evaluation and differentiation.

ArithmeticRunT0T_0 (K)Drift (μeV atom−1 ps−1\mu\mathrm{eV}\ \mathrm{atom}^{-1}\ \mathrm{ps}^{-1})RMS (μeV atom−1\mu\mathrm{eV}\ \mathrm{atom}^{-1})
FP321265.8−0.0072234.250
FP322247.8−0.020293.957
FP323303.5−0.0070174.648
BF16×31265.80.040724.079
BF16×32247.8−0.015553.950
BF16×33303.5−0.024584.406
TF321266.10.173928.98
TF322247.8−0.171231.18
TF323303.30.172935.40

Table A4. Energy conservation for all silicon trajectories. Each 64-atom diamond-Si trajectory runs for 500 ps with a 1 fs timestep. T0T_0 is the instantaneous temperature after 2 ps of equilibration at a 300 K target. Drift is the fitted total-energy slope; RMS is calculated after subtracting that line. Both quantities are evaluated over all samples using FACTO-OAM-M.

Figure A1. Total-energy changes in all silicon NVE runs. Panels a–c show three independently prepared runs for FP32 (dashed), BF16×3 (solid) and TF32 (dash-dotted). The vertical axis is [Etotal(t)−Etotal(0)]/N[E_{\mathrm{total}}(t)-E_{\mathrm{total}}(0)]/N, where total energy is potential plus kinetic energy and N=64N=64. Each trajectory spans 500 ps with a 1 fs timestep using FACTO-OAM-M. No fitted linear drift is subtracted. Traces are unsmoothed at 0.2 ps spacing; drift and detrended RMS in Appendix Table A4 use all samples at 0.01 ps spacing.

Appendix Note 3: Benchmark measurements and comparison sources

Matbench Discovery

ModelCPSF1RMSDκSRME\kappa_{\mathrm{SRME}}κSRE\kappa_{\mathrm{SRE}}
FACTO-OAM
XL0.8900.92190.059800.15600.06490
L0.8830.91780.061580.17330.07037
M0.8630.90360.065330.22840.1225
S0.8550.89640.066340.24250.1167
FACTO-MPA
XL0.8670.91010.065500.22180.09135
L0.8590.89880.063550.24150.1003
M0.8470.89150.065740.27530.1271
S0.8320.88340.065430.32960.1902

Table A5. CPS and its subscores for FACTO models. All quantities are dimensionless. F1 is evaluated on unique WBM prototypes; RMSD is evaluated at a symmetry tolerance of 10−210^{-2}.

Table A6 CMDS and its subscores for FACTO models. Errors are normalized percentage discrepancies. StS_t is the runtime subscore.

ModelCMDSvDOS (%)ADF (%)Pressure (%)Time (10310^3 s)StS_t
FACTO-OAM
XL0.75122.601.31346.2917.910.8038
L0.77322.781.44748.4110.810.9478
M0.78923.011.36846.567.7711.000
S0.79223.461.37744.977.5331.000
FACTO-MPA
XL0.75623.291.40744.5117.410.8118
L0.77523.241.42447.4910.780.9484
M0.79322.991.45845.187.7781.000
S0.79023.651.52345.327.4331.000

Table A6.

Table A7 Diatomic PBE-reference errors. Energy and well depth are in eV, force in eV/Å, and distances in Å.

ModelEnergyForceWall distanceBond lengthWell depth
FACTO-OAM
XL0.66570.63380.078880.23840.6193
L1.0650.70170.088880.28681.223
M0.74990.63230.067090.18310.7709
S0.79150.68960.067550.14690.7332
FACTO-MPA
XL0.89040.92820.098320.19100.6883
L0.82600.87520.091170.14840.6382
M0.84390.81680.11550.13280.6937
S0.89661.1770.081290.15180.6342

Table A7.

Table A8 Diatomic physicality and runtime. Flips and tortuosity are dimensionless.

ModelJump (eV)FlipsTortuosityTime (s)CDS
FACTO-OAM
XL0.42663.5291.01491.430.771
L0.19882.4601.02766.950.781
M0.16522.6781.01053.030.822
S0.22492.5401.01152.210.825
FACTO-MPA
XL0.39002.8391.02387.840.770
L0.21182.6441.01566.660.801
M0.15082.4251.01752.280.813
S0.15732.3101.00951.950.807

Table A8.

Public modelCPSCMDSCDS
Prophet-OAME-MBD0.9120.5780.572
TECE-OAM-RRA-1.0[55]0.9080.6300.270
EquFlashV20.9070.6510.644
EquiformerV3+DeNS-OAM0.9020.6380.269
GRACE-3L-OAM-L0.9000.7390.699
PET-OAM-XL0.8980.6340.388
TACE-OAM-L0.8890.7280.761
eSEN-30M-OAM0.8880.6250.648
EquFlash0.8880.6780.328
NequIP-OAM-XL0.8860.6230.796
MatRIS-10M-OAM[56]0.8770.5810.678
SevenNet-Omni-i120.8730.6260.696
NequIP-OAM-L0.8700.6700.818
GRACE-2L-OAM-L0.8650.7590.694
ORB v30.8600.7360.834
DPA-4.0.1-Pro-MPtrj0.8400.5980.123
Allegro-OAM-L0.8400.6020.828
GRACE-2L-OAM0.8370.7510.527
EquiformerV3+DeNS-MP0.8300.6220.413
DPA-3.1-3M-FT0.8020.6470.533
eSEN-30M-MP0.7970.5930.417
MACE-MPA-00.7950.6590.728
MatRIS-10M-MP0.7780.5930.264
AlphaNet-v1-OAM0.7690.6530.739
MatterSim v1 5M0.7670.6590.717
GRACE-1L-OAM0.7610.7490.552
Eqnorm MPtrj0.7560.6390.711
Nequix MP PFT0.7550.6160.565
NequIP-MP-L0.7300.6290.573
Nequix MP0.7290.6030.561
Allegro-MP-L0.7200.5560.747
SevenNet-l3i50.7140.6510.533
HIENet0.7070.6530.649
GRACE-2L-MPtrj0.6810.6980.355
BAM-MP-DeNS0.6720.5640.708
MACE-MP-00.6370.6330.627
eqV2 M0.5580.5950.283
ORB v2 MPA0.5280.7500.678
eqV2 S DeNS0.5220.6150.258
ORB v2 MPtrj0.4700.7270.667
M3GNet0.4280.5930.747
CHGNet0.4000.5540.297

Table A9. Public Matbench Discovery comparison at 19 October 2026. The 42 scored default-cohort models are ordered by CPS[3].

Figure A2. Effect of training data and model size on Matbench Discovery scores. Panels a–c show CPS, CMDS and CDS for OAM (solid circles) and MPA (dashed squares). All scores are higher-is-better. Training data are defined in Appendix Note 1; component measurements are reported in the accompanying tables.

MatCalc

ModelddEfE_fKVRHK_{\mathrm{VRH}}GVRHG_{\mathrm{VRH}}CVC_VForce slope
FACTO-OAM
XL0.195922.3511.0512.394.3970.9861
L0.211725.3811.7112.044.8010.9847
M0.210224.7312.5613.494.9780.9798
S0.229126.0612.5813.545.6540.9758
Published baselines
NequIP-OAM-XL0.279025.4012.7014.104.0200.9790
NequIP-OAM-L0.281026.6014.9014.904.4900.9730
Allegro-OAM-L0.287024.7012.7014.704.3800.9690
MatterSim-5M0.322040.5016.9015.805.4600.9730
UMA-s0.267025.0014.9021.503.6000.9860

Table A10. Crystalline-property errors and force calibration. Formation energy is in meV/atom, elastic moduli in GPa and CVC_V in J mol−1^{-1} K−1^{-1}. Structural distance dd and force slope are dimensionless; the ideal slope is one. Published baselines are from ref. 37.

MLIP Arena

DiagnosticOAM-XLOAM-LOAM-MOAM-S
Conservation (eV/A˚\text{\AA})0.017070.013600.016850.01426
ρ(E\rho(E, repulsion)−0.9823-0.9823−0.9915-0.9915−0.9970-0.9970−0.9985-0.9985
ρ(F\rho(F, descending)−0.7081-0.7081−0.9666-0.9666−0.9053-0.9053−0.9052-0.9052
Tortuosity1.0201.0231.0101.011
Energy jump (eV)0.013740.0069050.0069750.007406
Force flips4.2192.7083.0622.833
ρ(E\rho(E, attraction)0.85620.54030.69850.7496
ρ(F\rho(F, ascending)0.64910.56430.62730.7387

Table A11. Homonuclear-diatomic behavior on MLIP Arena. Spearman rank correlations ρ\rho, flips and tortuosity are dimensionless.

DiagnosticOAM-XLOAM-LOAM-MOAM-S
Heating AUC0.98480.98480.97990.9848
Compression AUC0.98480.98480.98480.9848
Heating exponent0.55790.53400.42250.2854
Compression exponent0.64790.64420.52270.3860

Table A12. Heating and compression stability on MLIP Arena. AUC integrates finite-energy survival (discrete 100-bin ceiling 65/66≈0.984865/66 \approx0.9848); exponents give the power-law scaling of steps per second with atom count.

DiagnosticOAM-XLOAM-LOAM-MOAM-S
Enthalpy residual (kcal/mol)3.8141.8783.6195.288
COM drift (A˚\text{\AA})0.0013300.0017640.0013750.001247
Throughput (steps s−1\text{s}^{-1})45.4173.89123.4157.7
Water yield (%)86.9888.8090.1085.94

Table A13. Reactive hydrogen combustion. The enthalpy residual is Arena’s potential-energy diagnostic; COM denotes center of mass.

DiagnosticOAM-XLOAM-LOAM-MOAM-S
Derivative flips1.0391.0481.0471.041
Tortuosity1.0061.0091.0071.006
ρ(E\rho(E, compression)−0.9978-0.9978−0.9986-0.9986−0.9981-0.9981−0.9995-0.9995
ρ(ΔE\rho(\Delta E, compression)0.99530.99650.99640.9968
ρ(E\rho(E, tension)0.98470.98430.98110.9828

Table A14. Relaxed equation-of-state diagnostics. Cells give the mean; ρ\rho denotes Spearman’s rank correlation. All quantities are dimensionless.

DiagnosticOAM-XLOAM-LOAM-MOAM-S
Derivative flips1.0061.0081.0081.011
Tortuosity1.0111.0001.0011.001
ρ(E, compression)\rho(E,\ \mathrm{compression})−0.9983-0.9983−0.9983-0.9983−0.9983-0.9983−0.9983-0.9983
ρ(ΔE, compression)\rho(\Delta E,\ \mathrm{compression})0.99950.99991.0000.9999
ρ(E, tension)\rho(E,\ \mathrm{tension})0.99600.99910.99870.9989

Table A15. Fixed-coordinate energy–volume diagnostics. Cells give the mean; ρ\rho denotes Spearman’s rank correlation. All quantities are dimensionless.

Third-party inference settings

ModelExecution and accelerationSoftware and revision
Prophet-OAME-MBDOEQ tensor-product kernels enabled; graph compilation disabled.Prophet dcfdc0b; PyTorch 2.8.0; OEQ 0.7.0
EquiformerV3+DeNS-OAMConservative gradient checkpoint; eager inference; graph compilation, AMP and TF32 disabled.EquiformerV3 a7300c58; PyTorch 2.7.1+cu128; e3nn 0.5.6
GRACE-3L-OAM-LTensorFlow/XLA; dense uniform aggregation; UQ disabled.TensorPotential 0.6.1; TensorFlow 2.16.2; tf-keras 2.16.0
NequIP-OAM-XLAOTInductor + OEQ; compiled .nequip.pt2 model loaded for inference.NequIP 0.19.1; PyTorch 2.8.0; OEQ 0.7.0
NequIP-OAM-LAOTInductor + OEQ; compiled .nequip.pt2 model loaded for inference.NequIP 0.19.1; PyTorch 2.8.0; OEQ 0.7.0
TACE-OAM-LOEQ enabled; float32 evaluation.TACE 0.2.1; PyTorch 2.8.0; OEQ 0.5.4
eSEN-30M-OAMConservative gradient inference in eager mode; AMP and TF32 disabled.fairchem-core 1.10.0; PyTorch 2.4.1; e3nn 0.6.0
ORB v3Conservative; internal graph network compiled with torch.compile (default mode, dynamic shapes); fixed brute-force neighbor search.orb-models 0.5.5; PyTorch 2.8.0
MACE-MPA-0Medium checkpoint; cuEq enabled; float32 evaluation.MACE 0.3.16; PyTorch 2.8.0; cuEq 0.11.0
GRACE-2L-OAM-LFP32 export supporting dense uniform aggregation; TensorFlow/XLA; UQ disabled.TensorPotential 0.6.1; TensorFlow 2.16.2; tf-keras 2.16.0

Table A16. Third-party inference and acceleration settings. Settings are for the ten public comparators in the main speed comparison. All weights are stored in FP32. OEQ and cuEq denote OpenEquivariance and cuEquivariance, respectively; AMP and UQ denote automatic mixed precision and uncertainty estimation.

Matbench Discovery score definitions

We use the default Matbench Discovery scorer and the public model cohort accessed on 22 October 2026 [3]. With [x]01=min⁡(1,max⁡(0,x))[x]_{0}^{1} = \min(1,\max(0,x)), the combined performance score is

CPS=0.5 F1+0.1[1−r/0.15]01+0.4(1−κSRME/2).(20)\mathrm{CPS} = 0.5\,\mathrm{F1} + 0.1\left[1-r/0.15\right]_{0}^{1} + 0.4(1-\kappa_{\mathrm{SRME}}/2). \tag*{(20)}

Here, F1 measures stability classification on unique WBM prototypes, rr is the dimensionless StructureMatcher RMSD at a symmetry tolerance of 10−210^{-2}, and κSRME\kappa_{\mathrm{SRME}} is the thermal-conductivity symmetric relative mean error[12, 45]. The SRME resolves mode-level errors, whereas κSRE\kappa_{\mathrm{SRE}} compares total thermal conductivities; both are dimensionless. The public comparison includes active models that predict forces: 43 registered models, of which 42 have complete CPS, CMDS and CDS values.

For molecular dynamics, let eve_v, eae_a and epe_p be normalized percentage discrepancies in vibrational density of states, angular distributions and pressure distributions across the 17 DynaMat trajectories (each 80,000 steps of 0.25 fs) [17]. The score is

CMDS=0.3(1−ev/100)+0.2(1−ea/100)+0.3(1−ep/100)+0.2St,St=[1−log⁡(t/9000)log⁡(300000/9000)]01,(21)\begin{aligned} \mathrm{CMDS} &= 0.3(1-e_v/100) + 0.2(1-e_a/100) \\ &\quad+ 0.3(1-e_p/100) + 0.2S_t, \\ S_t &= \left[1-\frac{\log(t/9000)}{\log(300000/9000)}\right]_{0}^{1}, \tag*{(21)} \end{aligned}

where tt is total rollout time in seconds and StS_t is the runtime subscore reported in Appendix Table A6. For eve_v, the Wasserstein-1 distance between normalized spectra is divided by the reference spectral standard deviation after clipping the reference tail at 0.999 cumulative power. For eae_a, the Wasserstein-1 distance is divided by the reference distribution’s distance to a structureless sin⁡θ\sin\theta background. These two ratios are capped at one and converted to percentages. The pressure error epe_p measures histogram non-overlap: 0% for identical distributions and 100% for disjoint distributions.

The diatomic score CDS combines four pillars with weights 4/9 for energy/force accuracy, 2/9 for geometry, 2/9 for physical curve shape and 1/9 for runtime. For a quantity xx with ideal floor aa and baseline bb, the linear subscore is [b−xb−a]01\left[\frac{b-x}{b-a}\right]_0^1; for runtime, we instead use [log⁡(b/x)log⁡(b/a)]01\left[\frac{\log(b/x)}{\log(b/a)}\right]_0^1. Appendix Table A17 gives the component weights and normalization endpoints. The weighted sum of the pillar scores is multiplied by 1−nexcluded/871-n_{\mathrm{excluded}}/87; no elements are excluded for FACTO, and 79 PBE reference curves are used for reference-relative errors.

PillarQuantityWeightFloorBaseline
AccuracyEnergy MAE5/904 eV
Force MAE4/905 eV/Å
GeometryWall-radius MAE1/200.5 Å
Bond-length error1/300.8 Å
Well-depth error1/603.5 eV
PhysicalityEnergy jump2/503 eV
Force reversals2/513.5
Tortuosity1/512
SpeedRuntime (logarithmic)150 s5,000 s

Table A17. CDS component weights and normalization endpoints. Weights are normalized within each pillar.

MatCalc observable conventions

The equilibrium, elasticity, phonon and force-softening collections contain 972, 3,953, 1,170 and 979 materials (9,308 high-energy configurations), respectively, corresponding to the WBM equilibrium, binary Materials Project elasticity, binary Alexandria phonon and WBM high-energy collections [49]. Equilibrium observables are formation-energy MAE and dimensionless CrystalNN structural-fingerprint distance; elastic moduli use the Voigt–Reuss–Hill average. We report harmonic heat capacity at 300 K using the benchmark’s Phonopy cell-molar convention [13, 35, 36], and force calibration as the mean proportional-fit slope between predicted and DFT force components across materials (ideal value one). Baseline values in Appendix Table A10 are taken from Table 3 of ref. [37].

MLIP Arena diagnostic definitions

Public MLIP Arena values are from the 27 March 2026 comparison [14, 29]. On homonuclear diatomics and volumetric deformation scans (Appendix Tables A11, A14 and A15), conservation deviation compares predicted forces with numerical energy derivatives on the sampled grid, tortuosity measures excess curve variation relative to a single-well curve, and ρ\rho denotes Spearman’s rank correlation coefficient. In Appendix Table A12, heating and compression AUCs integrate finite-energy trajectory survival using the benchmark’s 100-edge discrete bin scorer (which has a maximum achievable value of (97+1/2)/99=65/66≈0.9848(97+1/2)/99=65/66\approx0.9848 when all trajectories survive to the final step), and the heating and compression exponents αspeed\alpha_{\mathrm{speed}} are obtained by fitting steps per second to a power law proportional to N−αspeedN^{-\alpha_{\mathrm{speed}}}, where NN is the atom count. In Appendix Table A13, the combustion enthalpy residual is computed from the potential-energy change divided by the stoichiometric yield of 128 water molecules and converted using 23.0 kcal mol−1^{-1} per eV relative to −68.3078-68.3078 kcal mol−1^{-1}.

References

  1. [1]Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J. Bartel, and Gerbrand Ceder. CHGNet as a pretrained universal neural network potential for charge-informed atomistic modelling. Nature Machine Intelligence, 5(9):1031–1041, 2023. doi: 10.1038/s42256-023-00716-3.DOI
  2. [2]Xiang Fu, Zhenghao Wu, Wujie Wang, Tian Xie, Sinan Keten, Rafael Gomez-Bombarelli, and Tommi Jaakkola. Forces are not enough: Benchmark and critical evaluation for machine learning force fields with molecular simulations. Transactions on Machine Learning Research, 2023. doi: 10.48550/arXiv.2210.07237.DOI
  3. [3]Matbench Discovery contributors. Matbench Discovery: combined-score implementation and default model roster. https://github.com/janosh/matbench-discovery, 2026. Version of 19 October 2026, accessed 22 October 2026.
  4. [4]Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature Communications, 13(1):2453, 2022. doi: 10.1038/s41467-022-29939-5.DOI
  5. [5]Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Advances in Neural Information Processing Systems, volume 35, pages 11423–11436, 2022. URL https://proceedings.cc/paper_files/paper/2022/hash/4a36c3baf11ed9f34615b81edb5bbc-Abstract-Conference.html.
  6. [6]Albert Musaealian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, and Boris Kozinsky. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications, 14(1):579, 2023. doi: 10.1038/s41467-023-36329-y.
  7. [7]Guillem Simeon and Gianni De Fabritiis. TensorNet: Cartesian tensor representations for efficient learning of molecular potentials. In Advances in Neural Information Processing Systems, volume 36, pages 37334–37353, 2023. doi: 10.52202/075280-1623.DOI
  8. [8]Yaolong Zhang and Hua Guo. Node-equivariant message passing for efficient and accurate machine learning interatomic potentials. Chemical Science, 17(7):3793–3803, 2026. doi: 10.1039/D5SC07248D.DOI
  9. [9]Thiago Reschützegger, Sarp Aykent, Gabriel Jacob Perin, Bruno Henrique Nunes, Flaviu Cipcigan, Rodrigo Neumann Barros Ferreira, Mathias Steiner, and Fabian L. Thiemann. Equivariant interatomic potentials without tensor products. arXiv preprint arXiv:2601.15492, 2026. doi: 10.48550/arXiv.2601.15492.DOI
  10. [10]Seung Yul Lee, Hojoon Kim, Yutack Park, Dawoon Jeong, Seungwu Han, Yeonhong Park, and Jae W. Lee. FlashTP: Fused, sparsity-aware tensor product for machine learning interatomic potentials. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors, Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 33143–33156. PMLR, 2025. URL https://proceedings.mlr.press/v267/lee25l.html.
  11. [11]Greg Henry, Ping Tak Peter Tang, and Alexander Heinecke. Leveraging the bfloat16 artificial intelligence datatype for higher-precision computations. arXiv preprint arXiv:1904.06376, 2019. doi: 10.48550/arXiv.1904.06376.arxiv.org/abs/1904.06376
  12. [12]Janosh Riebesell, Rhys E. A. Goodall, Philipp Benner, Yuan Chiang, Bowen Deng, Gerbrand Ceder, Mark Asta, Alpha A. Lee, Anubhav Jain, and Kristin A. Persson. A framework to evaluate machine learning crystal stability predictions. Nature Machine Intelligence, 7(6):836–847, 2025. doi: 10.1038/s42256-025-01055-1.DOI
  13. [13]Aaron D. Kaplan, Runze Liu, Ji Qi, Tsz Wai Ko, Bowen Deng, Janosh Riebesell, Gerbrand Ceder, Kristin A. Persson, and Shyue Ping Ong. A foundational potential energy surface dataset for materials. arXiv preprint arXiv:2503.04070, 2025. doi: 10.48550/arXiv.2503.04070.DOI
  14. [14]Yuan Chiang, Tobias Kreiman, Christine Zhang, Matthew Kuner, Elizabeth Weaver, Ishan Amin, Hyunsoo Park, Yunsung Lim, Jihan Kim, Daryl Chrzan, Aron Walsh, Samuel Blau, Mark Asta, and Aditi Krishnapriyan. MLIP Arena: Advancing fairness and transparency in machine learning interatomic potentials via an open, accessible benchmark platform. In Advances in Neural Information Processing Systems, volume 38, 2025. doi: 10.52202/085713-4410.
  15. [15]James F. Ziegler, Jochen P. Biersack, and Ulf Littmark. The Stopping and Range of Ions in Matter, volume 1. Pergamon, New York, 1985.DOI
  16. [16]Tsz Wai Ko, Bowen Deng, Marcel Nassar, Luis Barroso-Luque, Runze Liu, Ji Qi, Atul C. Thakur, Adesh Rohan Mishra, Elliott Liu, Gerbrand Ceder, Santiago Miret, and Shyue Ping Ong. Materials graph library (MatGL), an open-source graph deep learning library for materials science and chemistry. npj Computational Materials, 11(1):253, 2025. doi: 10.1038/s41524-025-01742-y.DOI
  17. [17]Mikołaj J. Gawkowski, Nongnuch Arthit, Silvia Bonfanti, Abhijeet Sadashiv Gangan, Hendrik H. Heenen, Joseph Kioseoglu, Ivor Lončarić, Hemanadhan Myneni, Janosh Riebesell, Mariana Rossi, Matthias Rupp, Jonathan Schmidt, Shubham Sharma, Benjamin X. Shi, Antoni Wadowski, Lukas Hörmann, and Venkat Kapil. Dyna-Mat: End-to-end benchmarking of foundation machine learning interatomic potentials in finite-temperature ensembles. arXiv preprint arXiv:2607.03433, 2026. doi: 10.48550/arXiv.2607.03433.DOI
  18. [18]Luis Barros-Luque, Muhammed Shuaibi, Xiang Fu, Brandon M. Wood, Misko Dzamba, Meng Gao, Ammar Rizvi, Matt Uyttendaele, C. Lawrence Zitnick, and Zachary W. Ulissi. The open materials 2024 (OMat24) inorganic materials dataset and models. Nature Computational Science, 6(6):642–652, 2026. doi: 10.1038/s43588-026-00996-w.
  19. [19]Jonathan Schmidt, Tiago F. T. Cerqueira, Aldo H. Romero, Antoine Loew, Fabian Jäger, Hai-Chen Wang, Silvana Botti, and Miguel A. L. Marques. Improving machine-learning models in materials science through large datasets. Materials Today Physics, 48:101560, 2024. doi: 10.1016/j.mtphys.2024.101560.DOI
  20. [20]Kairos Materials. ELEMENTA: Toward unbiased scaling of materials data from first principles. Technical report, Kairos Materials, July 2026. URL https://www.kairosmaterials.com/papers/ELEMENTA.pdf.DOI
  21. [21]Kairos Materials. Prophet: Atomistic foundation models for materials simulation and discovery. https://github.com/kairosmaterial/prophet, 2026. Research preview, accessed 10 October 2026.
  22. [22]Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen, Alexandre Duval, Sam Walton Norwood, and Tess Smidt. EquiformerV3: Scaling efficient, expressive, and general SE(3)-equivariant graph attention transformers. arXiv preprint arXiv:2604.09130, 2026. doi: 10.48550/arXiv.2604.09130.DOI
  23. [23]Yury Lysogorskiy, Anton Bochkarev, and Ralf Drautz. Graph atomic cluster expansion for foundational machine learning interatomic potentials. npj Computational Materials, 12(1):114, 2026. doi: 10.1038/s41524-026-01979-1.DOI
  24. [24]Zemin Xu, Wenbo Xie, and P. Hu. Spectral/spatial tensor atomic cluster expansion with universal embeddings in Cartesian space. arXiv preprint arXiv:2509.14961, 2025. doi: 10.48550/arXiv.2509.14961.DOI
  25. [25]Xiang Fu, Brandon M. Wood, Luis Barroso-Luque, Daniel S. Levine, Meng Gao, Misko Dzamba, and C. Lawrence Zitnick. Learning smooth and expressive interatomic potentials for physical property prediction. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors, Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pages 17875–17893. PMLR, 2025. URL https://proceedings.mlr.press/v267/fu25h.html.arxiv.org/abs/2502.12147
  26. [26]Chuin Wei Tan, Marc L. Descoteaux, Mit Kotak, Gabriel de Miranda Nascimento, Seán R. Kavanagh, Laura Zichi, Menghang Wang, Aadit Saluja, Yizhong R. Hu, Tess Smidt, Anders Johansson, William C. Witt, Boris Kozinsky, and Albert Musaelian. High-performance training and inference for deep equivariant interatomic potentials. arXiv preprint arXiv:2504.16068, 2025. doi: 10.48550/arXiv.2504.16068.DOI
  27. [27]Benjamin Rhodes, Sander Vandenhaute, Vaidotas Šimkus, James Gin, Jonathan Godwin, Tim Duignan, and Mark Neumann. Orb-v3: atomistic simulation at scale. arXiv preprint arXiv:2504.06231, 2025. doi: 10.48550/arXiv.2504.06231.
  28. [28]Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena, Dávid P. Kovács, Janosh Riebesell, Xavier R. Advincula, Mark Asta, Matthew A Waylon, William J. Baldwin, Fabian Berger, Noam Bernstein, Arghya Bhowmik, Filippo Bigi, Samuel M. Blau, Vlad Carare, Michele Ceriotti, Sanggyu Chong, James P. Darby, Sandip De, Flaviano Della Pia, Volker L. Deringer, Rokas Elijošius, Zakariya El-Machachi, Edwin Fako, Fabio Falcioni, Andrea C. Ferrari, John L. A. Gardner, Mikołaj J. Gawkowski, Annalena Genreith-Schriever, Janine George, Rhys E. A. Goodall, Jonas Grandel, Clare P. Grey, Petr Grigoriev, Shuang Han, Will Handley, Hendrik H. Heenen, Kersti Hermansson, Cheuk Hin Ho, Stephan Hofmann, Christian Holm, Jad Jaafar, Konstantin S. Jakob, Hyunwook Jung, Venkat Kapil, Aaron D. Kaplan, Nima Karimitarı, James R. Kermode, Panagiotis Kourlis, Namu Kroupa, Jolla Kullgren, Matthew C. Kuner, Domantas Kuryla, Guoda Lieponiute, Chen Lin, Johannes T. Margraf, Ioan-Bogdan Magdău, Angelos Michaelides, J. Harry Moore, Aakash A. Naik, Samuel P. Niblett, Sam Walton Norwood, Niamh O’Neill, Christoph Ortner, Kristin A. Persson, Karsten Reuter, Andrew S. Rosen, Louise A. M. Rosset, Lars L. Schaaf, Christoph Schran, Benjamin X. Shi, Eric Sivonxay, Tamás K. Stenczel, Christopher Sutton, Viktor Svahn, Thomas D. Swinburne, Jules Tilly, Cas van der Oord, Santiago Vargas, Eszter Varga-Umbrich, Tejs Vegge, Martin Vondrák, Yangshuai Wang, William C. Witt, Thomas Wolf, Fabian Zills, and Gábor Csányi. A foundation model for atomistic materials chemistry. The Journal of Chemical Physics, 163(18):184110, 2025. doi: 10.1063/5.0297006.DOI
  29. [29]MLIP Arena contributors. MLIP Arena: task data and ranking implementation. https://huggingface.co/spaces/atomind/mlip-arena, 2026. Version of 27 March 2026, accessed 9 October 2026.
  30. [30]Mark Neumann, James Gin, Benjamin Rhodes, Steven Bennett, Zhiyi Li, Hitarth Choubisa, Arthur Hussey, and Jonathan Godwin. Orb: a fast, scalable neural network potential. arXiv preprint arXiv:2410.22570, 2024. doi: 10.48550/arXiv.2410.22570.
  31. [31]Yutack Park, Jaesun Kim, Seungwoo Hwang, and Seungwu Han. Scalable parallel algorithm for graph neural network interatomic potentials in molecular dynamics simulations. Journal of Chemical Theory and Computation, 20(11):4857–4868, 2024. doi: 10.1021/acs.jctc.4c00190.
  32. [32]Hai-Chen Wang, Silvana Botti, and Miguel A. L. Marques. Predicting stable crystalline compounds using chemical similarity. npj Computational Materials, 7(1):12, 2021. doi: 10.1038/s41524-020-00481-6.DOI
  33. [33]Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science, 2(11):718–728, 2022. doi: 10.1038/s43588-022-00349-3.DOI
  34. [34]Han Yang, Chenxi Hu, Yichi Zhou, Xixian Liu, Yu Shi, Jielan Li, Guanzhi Li, Zekun Chen, Shuizhou Chen, Claudio Zeni, Matthew Horton, Robert Pinsler, Andrew Fowler, Daniel Zügner, Tian Xie, Jake Smith, Lixin Sun, Qian Wang, Lingyu Kong, Chang Liu, Hongxia Hao, and Ziheng Lu. MatterSim: A deep learning atomistic model across elements, temperatures and pressures. arXiv preprint arXiv:2405.04967, 2024. doi: 10.48550/arXiv.2405.04967.DOI
  35. [35]MatCalc contributors. MatCalc: benchmark classes and property calculators. https://github.com/materialyzeai/matcalc, 2026. Accessed 8 September 2026.
  36. [36]Bowen Deng, Yunyeong Choi, Peichen Zhong, Janosh Riebesell, Shashwat Anand, Zhuohan Li, KyuJung Jun, Kristin A. Persson, and Gerbrand Ceder. Systematic softening in universal machine learning interatomic potentials. npj Computational Materials, 11(1):9, 2025. doi: 10.1038/s41524-024-01500-6.DOI
  37. [37]Seán R. Kavanagh, Chuin Wei Tan, Menghang Wang, Marc L. Descoteaux, Gabriel de Miranda Nascimento, Ulrik Unneberg, Laura Zichi, Francesco Libbi, Norma Rivano, Austin Glover, Vivek Bharadwaj, Anders Johansson, William C. Witt, Albert Musaeelian, and Boris Kozinsky. Fast and accurate foundation models for equivariant machine-learned interatomic potentials. arXiv preprint arXiv:2607.28461, 2026. doi: 10.48550/arXiv.2607.28461. Version 1; Table 3.
  38. [38]Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burkovski, Geeta Chauhan, Anjali Chourdia, Will Constable, Alban Desmaison, Zachary DeVito, Elias Ellison, Will Feng, Jiong Gong, Michael Gschwind, Brian Hirsh, Sherlock Huang, Kishteej Kalambarkar, Laurent Kirsch, Michael Lazos, Mario Lezcano, Yanbo Liang, Jason Liang, Yinghai Lu, C. K. Luk, Bert Maher, Yunjie Pan, Christian Puhrsch, Matthias Reso, Mark Saroufim, Marcos Yukio Sraichi, Helen Suk, Michael Suo, Phil Tillett, Eikan Wang, Xiaodong Wang, William Wen, Shunting Zhang, Xu Zhao, Keren Zhou, Richard Zou, Ajit Mathews, Gregory Chanan, Peng Wu, and Soumith Chintala. 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, pages 929–947, 2024. doi: 10.1145/3620665.3640366.DOI
  39. [39]Keller Jordan, Yuchen Jin, Vlado Boza, Jiacheng You, Franz Cesista, Laker Newhouse, and Jeremy Bernstein. Muon: An optimizer for hidden layers in neural networks. https://kellerjordan.github.io/posts/muon/, 2024. Published 8 December 2024.
  40. [40]Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7.
  41. [41]Tiancheng Li, Wentao Li, Anyang Peng, Jianming Xue, Linfeng Zhang, Duo Zhang, and Han Wang. DPA4: Pushing the accuracy-cost frontier of interatomic potentials with EMFA SO(2) convolution. arXiv preprint arXiv:2606.02419, 2026. doi: 10.48550/arXiv.2606.02419.DOI
  42. [42]Jingyuan Liu, Jianlin Su, Xingcheng Yao, Zhejun Jiang, Guokun Lai, Yulun Du, Yidao Qin, Weixin Xu, Enzhe Lu, Junjie Yan, Yanru Chen, Huabin Zheng, Yibo Liu, Shaowei Liu, Bohong Yin, Weiran He, Han Zhu, Yuzhi Wang, Jianzhou Wang, Mengnan Dong, Zheng Zhang, Yongsheng Kang, Hao Zhang, Xinran Xu, Yutao Zhang, Yuxin Wu, Xinyu Zhou, and Zhilin Yang. Muon is scalable for LLM training. arXiv preprint arXiv:2502.16982, 2025. doi: 10.48550/arXiv.2502.16982.DOI
  43. [43]Taejong Joo, Wenhan Xia, Cheolmin Kim, Ming Zhang, and Eugene Ie. On surprising effectiveness of masking updates in adaptive optimizers. arXiv preprint arXiv:2602.15322, 2026. doi: 10.48550/arXiv.2602.15322.DOI
  44. [44]Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffrey Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L. Chevrier, Kristin A. Persson, and Gerbrand Ceder. Python materials genomics (pymatgen): A robust, open-source Python library for materials analysis. Computational Materials Science, 68:314–319, 2013. doi: 10.1016/j.commatsci.2012.10.028.DOI
  45. [45]Balázs Póta, Paramvir Ahlawat, Gábor Csányi, and Michele Simoncelli. Thermal conductivity predictions with foundation atomistic models. Nature Communications, 17(1):10566, 2026. doi: 10.1038/s41467-026-76391-w.DOI
  46. [46]Maarten de Jong, Wei Chen, Thomas Angst, Anubhav Jain, Randy Notestine, Anthony Gamst, Marcel Sluiter, Chaitanya Krishna Ande, Sybrand van der Zwaag, Jose J. Plata, Cormac Toher, Stefano Curtarolo, Gerbrand Ceder, Kristin A. Persson, and Mark Asta. Charting the complete elastic properties of inorganic crystalline compounds. Scientific Data, 2(1):150009, 2015. doi: 10.1038/sdata.2015.9.DOI
  47. [47]Antoine Loew, Dewen Sun, Hai-Chen Wang, Silvana Botti, and Miguel A. L. Marques. Universal machine learning interatomic potentials are ready for phonons. npj Computational Materials, 11(1):178, 2025. doi: 10.1038/s41524-025-01650-1.DOI
  48. [48]Atsushi Togo, Laurent Chaput, Terumasa Tadano, and Isao Tanaka. Implementation strategies in phonopy and phono3py. Journal of Physics: Condensed Matter, 35(35):353001, 2023. doi: 10.1088/1361-648X/acd831.DOI
  49. [49]Materialyze.AI. MatCalc-Bench datasets. https://huggingface.co/datasets/materialyze/matcalc-bench, 2026. Dataset version of 26 April 2026, accessed 8 September 2026.
  50. [50]Nils E. R. Zimmermann and Anubhav Jain. Local structure order parameters and site fingerprints for quantification of coordination environment and crystal structure similarity. RSC Advances, 10(10):6063–6081, 2020. doi: 10.1039/C9RA07755C.DOI
  51. [51]Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E. Castelli, Rune Christensen, Marcin Dulak, Jesper Friis, Michael N. Groves, Björk Hammer, Cory Hargus, Eric D. Hermes, Paul C. Jennings, Peter Bjerre Jensen, James Kermode, John R. Kitchin, Esben Leonhard Kolsbjerg, Joseph Kubl, Kristen Kaasbjerg, Steen Lysgaard, Jón Bergmann Maronsson, Tristan Maxson, Thomas Olsen, Lars Pastewka, Andrew Peterson, Carsten Rostgaard, Jakob Schiøtz, Ole Schütt, Mikkel Strange, Kristian S. Thygesen, Tejs Vegge, Lasse Vilhelmsen, Michael Walter, Zhenhua Zeng, and Karsten W. Jacobsen. The atomic simulation environment—a Python library for working with atoms. Journal of Physics: Condensed Matter, 29(27):273002, 2017. doi: 10.1088/1361-648X/aab0ee.DOI
  52. [52]Aaron D. Kaplan. Total energies of neutral isolated atoms. https://github.com/esoteric-ephemera/isolated_atomic_energies/blob/28d3ce280d0a94d9b6926b573ea5f56dcc0635d/atomic_energies.csv, 2024. Repository version of 11 December 2024, accessed 31 October 2026.
  53. [53]FAIR-Chem contributors. FAIRChem. https://github.com/facebookresearch/fairchem/tree/977a80328f2be44649b414a9907a1d6ef21f81e95, 2025. Repository version of 23 April 2025, accessed 31 October 2026.
  54. [54]Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganyu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, Xinrong Zhang, Zheng Leng Thai, Kaihuo Zhang, Chongyi Wang, Yuan Yao, Chenyang Zhao, Jie Zhou, Jie Cai, Zhongwu Zhai, Ning Ding, Chao Jia, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. MiniCPM: Unveiling the potential of small language models with scalable training strategies. arXiv preprint arXiv:2404.06395, 2024. doi: 10.48550/arXiv.2404.06395.DOI
  55. [55]Zemin Xu, Wenbo Xie, and P. Hu. Edge cluster expansion with radial rotary attention for interatomic potentials. arXiv preprint arXiv:2607.10664, 2026. doi: 10.48550/arXiv.2607.10664.DOI
  56. [56]Yuanchang Zhou, Siyu Hu, Xiangyu Zhang, Hongyu Wang, Guangming Tan, and Weile Jia. MatRIS: Toward reliable and efficient pretrained machine learning interatomic potentials. arXiv preprint arXiv:2603.02002, 2026. doi: 10.48550/arXiv.2603.02002.DOI

Paper details

vlnc:W_9k5tspxg
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 9

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