betaFeedback

WorksPhysics and Astronomy

Scalable Foundation Interatomic Potentials via Message-Passing Pruning and Graph Partitioning

Abstract

Atomistic foundation models (AFMs) have great promise as accurate interatomic potentials, and have enabled data-efficient molecular dynamics simulations with near quantum mechanical accuracy. However, AFMs remain markedly slower at inference and are far more memory-intensive than conventional interatomic potentials, due to the need to capture a wide range of chemical and structural motifs in pre-training datasets requiring deep, parameter-rich model architectures. These deficiencies currently limit the practical use of AFMs in molecular dynamics (MD) simulations at extended temporal and spatial scales. To address this problem, we propose a general workflow for accelerating and scaling AFMs containing message-passing architectures. We find that removing low-contribution message-passing layers from AFM backbones serves as an effective pruning method, significantly reducing the parameter count while preserving the accuracy and data-efficiency of AFMs. Once pruned, these models become more accessible for large scale simulations via a graph-partitioned, GPU-distributed strategy, which we implement and demonstrate within the AFM fine-tuning platform MatterTune. We show that this approach supports million-atom simulations on both single and multiple GPUs, and enables task-specific large-scale simulations at nanosecond timescales with AFM-level accuracy.

Introduction

Atomistic molecular dynamics (MD) simulations capture the evolution of atomic configurations over time and allow for the mapping of microscopic interactions onto macroscopic observables, making them indispensable for uncovering reaction mechanisms, predicting material behavior, and guiding the design and screening of new materials [1, 2, 3, 4, 5, 6, 7]. A central challenge in computational materials science lies in extending the spatial and temporal limits of these atomistic simulations, thereby enabling the study of increasingly complex phenomena such as phase transitions, defect migration, or chemical reactions in realistic environments. Achieving such scales requires models that can evaluate the potential energy surface (PES) of a given system with both high efficiency and fidelity.

Ab initio methods, such as density functional theory (DFT) and coupled cluster (CC) approaches [8, 9, 10], are capable of highly accurate evaluation of the PES, providing high-fidelity results with no parameterization needed. However, the steep computational scaling of these methods with respect to system size severely limits their practical applicability, restricting simulations to relatively short timescales and small system sizes. While they remain the gold standard for benchmarking and for training parameterized models, their direct use in large-scale or long-timescale molecular dynamics remains computationally prohibitive. This limitation has motivated the search for alternative strategies that retain near-ab initio accuracy while significantly reducing computational cost. [11, 12, 13, 14, 15].

Machine learning interatomic potentials (MLIPs) [16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26] are a rapidly growing class of methods that overcome this computational bottleneck by serving as an efficient alternative to first-principles calculations. Through the training of parameterized models, MLIPs can reduce the evaluation cost of the PES to linear or even sublinear scaling, accelerating PES predictions by several orders of magnitude while maintaining near-quantum accuracy. Among these, graph neural networks (GNNs) are a widely used architecture in MLIP models, with the capability to effectively learn long-range and high-order atomic interactions in a data-driven manner, particularly for structurally diverse and compositionally complex systems. The use of these MLIP models to accelerate MD simulations has been extensively studied, and mature interfaces with MD simulation packages have been established in practice [19, 27, 28, 20, 29, 30].

Recently, the concept of Atomistic Foundation Models (AFMs) has been proposed [31, 32, 33, 34, 35, 36, 37, 38, 39] as a class of models with broad or "universal" applicability across the periodic table and with comparable or sometimes better accuracy than bespoke MLIPs. These models are constructed via pre-training on large, diverse atomistic datasets, learning high-quality, general-purpose material embeddings that transfer across a broad chemical and structural space. Early studies [35, 40, 41, 42] show that AFMs can be further fine-tuned on only a handful of task-specific data and still match or even surpass the bespoke MLIPs in terms of accuracy, significantly lowering training costs for MLIPs and greatly enabling widespread use of these models in simulations.

However, despite their superior accuracy and data-efficiency, AFMs remain difficult to deploy in production-level MD simulations, especially those spanning nanoseconds and involving more than tens of thousands of atoms [43, 44, 45, 46]. To capture the complex chemical and geometric relationships present in large, element-wide training sets, AFMs generally rely on sophisticated, high-capacity architectures to maximize expressivity; for instance many message-passing layers are introduced in the models to better capture medium- and long-range interactions, further inflating model size and complexity. For example, the MatterSim V1 [35] model employs 3-4 message-passing layers, JMP [34] uses 4-6 layers, and ORB V2 [36] incorporates up to 15 message-passing layers. These design choices lead to two practical bottlenecks: (i) AFMs run markedly slower than conventional MLIPs and cannot keep pace with the millions of timesteps required for nanosecond-scale MD; (ii) their much larger memory footprint at inference restricts the maximum structure size that can be processed on a single GPU.

To overcome the slow inference and high memory demand of AFMs and thereby position them as capable replacements for conventional MLIPs in MD simulations, we propose an acceleration workflow that satisfies two general requirements: (i) that it should be reasonably architecture agnostic, applying broadly to most AFMs rather than being custom-tailored to a single architecture; and (ii) that it should impose only minimal modifications to existing AFM implementations, ensuring flexible and straightforward adoption.

Our approach is inspired by the over-smoothing phenomenon observed in graph neural networks: as message passing grows deeper, node features tend to homogenize and the incremental information gained at each layer diminishes [47, 48, 49, 50]. Motivated by these insights, we compared the node embeddings produced at every message-passing layer in three representative AFMs: MatterSim, JMP, and ORB, and verified that layers at different depths contribute unevenly to the final representation. The early layers provide most of the useful signal, whereas the deepest layers add relatively little. Based on this finding, we designed a pruning scheme that keeps the first KK message-passing layers and removes the other low-contribution ones, yielding a lightweight version of an AFM. Fine-tuning experiments and accuracy analyses show that the pruned AFMs, even when only one or two layers are retained, preserve the accuracy and data-efficiency superiority of AFMs.

This pruning strategy, which we term Message-Passing Depth Pruning (MPDP), not only reduces the parameter count of AFMs and thus their inference costs, it also makes the model better suited to the second stage of our proposed acceleration workflow: graph-partitioned multi-GPU parallelism. A central challenge in enabling multi-GPU parallelization of potential models via graph partitioning lies in preserving the effective receptive field, which is the spatial extent over which information can be aggregated, of every atom. For a model with local cutoff rcr_c and MM layers of message passing, the effective receptive field cutoff is consequently extended to become reff=M×rcr_{\mathrm{eff}} = M \times r_c. One approach to parallelization relies on shared buffers or explicit inter-GPU communication to propagate node messages across devices, thereby maintaining global information flow. This does, however, introduce a communication overhead that grows approximately linearly with the number of message-passing layers, which can be costly on hardware without high-bandwidth interconnects such as NVLink. An alternative approach utilizes the so-called halo method, where each subgraph is pre-expanded to contain all neighbors inside the receptive field at the time of partitioning. However, as the effective receptive field expands roughly cubically with the message-passing depth MM, this strategy can lead to substantial computational and memory overhead. However, by pruning the layers of the model with the MPDP approach, we can reduce the original depth MM—which is typically large—to a much smaller value KK. This substantially alleviates the overhead associated with the halo-based graph partitioning scheme, thereby significantly improving its efficiency with a small loss in accuracy. Beyond this, the halo method also offers several additional advantages: it completely eliminates the need for inter-GPU communication during inference, and its implementation is straightforward without requiring any modifications to the underlying model architecture. Taken together, these properties make the halo-based parallelization scheme particularly well-suited to our pruned models, and we adopt this strategy in the present work.

By integrating the MPDP method with graph-partitioned multi-GPU parallelism, we have designed an acceleration workflow applicable to any AFM that relies on deep message passing. The procedure first applies MPDP, retaining only the first KK message-passing layers, then finetunes the pruned model, and finally partitions the graph to either run on a single GPU in serial, or distribute inference across multiple GPUs for an additional speedup. It is worth noting that a number of recent studies have explored multi-GPU acceleration of MLIPs and AFMs. chemtrain-deploy [51] adopts the aforementioned halo-based spatial domain decomposition of LAMMPS with a focus on JAX-defined [52] models. Meanwhile, DistMLIP [53] employs direct atomic graph partitioning with explicit inter-GPU communication at each message-passing layer. Our approach differs in the inclusion of the Message-Passing Depth Pruning (MPDP) strategy, which substantially reduces model depth and synergizes well with the halo-based partitioning strategy, enabling efficient and communication-free parallelization. We furthermore integrate this parallelism with AFM finetuning to enable nanosecond-scale MD simulations with AFM-level quality of systems containing tens of thousands or more atoms, making the previously cumbersome AFMs powerful engines for realistic large-scale MD.

Results

In the following, we demonstrate the utility of the proposed method in three illustrative case studies: (i) simulating quenching of molten Li3_3PO4_4 phosphate electrode, (ii) modeling phase segregation during the melt–quench process of the mixed silicon-silica system, and (iii) exploring metal-support interactions on a oxide-supported metal catalyst. These simulations span from tens of thousands to approximately half a million atoms with trajectory lengths ranging from 50 ps to several nanoseconds. With pruned, partition-parallel AFMs in place, all simulations can be completed within practical wall-clock budgets while yielding physically consistent structural and dynamical behavior.

Melt–Quench Formation of Amorphous Li3_3PO4_4 and Li-Ion Diffusion

Li3_3PO4_4 is widely investigated as a solid-state electrolyte due to its high Li-ion conductivity and electrochemical stability making it a promising replacement for liquid electrolytes in Li-ion batteries [54, 55, 56, 57]. A prior work by Musaelian et al. [25] demonstrated the training of a lightweight Allegro model on a provided dataset containing Li3_3PO4_4 melting and quenching trajectories, which serves as a representative benchmark for speed and scalability in massively parallel MD simulations.

Following the same dataset splitting method in [25], we sampled 5 000 frames from melting and quenching trajectory separately to construct a 10 000-structure training set (hereafter denoted 10k-sample). To highlight the superior data efficiency of AFMs, we additionally constructed another ultra-small training set by selecting only the first 100 frames from the melting trajectory (denoted 100-sample). We fine-tuned the AFMs MatterSim, MACE, JMP, and ORB on both the 10k-sample and 100-sample datasets and report the results in Table 1. From these results we observe that, for a given AFM backbone, fine-tuning on the 10k-sample versus the 100-sample dataset yields only modest performance differences. This highlights a key advantage of AFMs over conventional MLIPs, namely their ability to leverage pretrained representations to transfer across chemical and structural variations without relying on large training datasets. Furthermore, most AFM variants fine-tuned on as few as 100 samples can already match or surpass the accuracy of the Allegro baseline trained on the full 10k-sample set and, of particular note, the 100-sample dataset consists only the first 100 frames of the melting trajectory and contains no quenching structures. This demonstrates that, with virtually no prior coverage of the target configuration space, it is possible to obtain a model of competitive accuracy using only a minimal amount of training data. In conventional AFMs, this comes with a significant trade-off in much slower simulation speeds and size constraints. We subsequently show that these limitations can be overcome through graph partitioning and parallelization.

10k-sample100-sample
MAEE_E (meV/atom)MAEF_F (meV/Å)MAEE_E (meV/atom)MAEF_F (meV/Å)
Allegro1.773.40––
MatterSimβ V1-1MMP×10.656.838.573.17
MatterSimβ V1-1MMP×20.537.9310.660.17
MatterSimβ V1-1MMP×3*0.332.604.350.57
JMP-SMP×10.621.07111.673.26
JMP-SMP×20.614.32120.138.68
JMP-SMP×4*0.411.1784.229.78
ORB-V3-OmatMP×10.332.60137.01259.50
ORB-V3-OmatMP×31.224.949.368.05
ORB-V3-OmatMP×5*1.019.9811.273.38
MACE-M-OmatMP×10.4631.7653.4551.87
MACE-M-OmatMP×2*0.1920.9614.9425.86

Table 1. Fine-tuning performance of various AFMs on Li3_3PO4_4 and comparison with the Allegro baseline

We then applied varying degrees of MPDP to each AFM, denoting a variant that retains only the first KK message-passing layers as MP ×K\times K. The pruned models were also fine-tuned on both datasets, with results also summarized in Table 1. Remarkably, we found that, in most cases, pruning the message-passing stack lead to an unacceptable loss in accuracy; even in the extreme case The Allegro baseline MAEs are taken from [25]. MP×KK denotes a pruned model retaining the first KK message-passing layers. An asterisk superscript MP×M∗M^* indicates that the original backbone natively has MM message-passing layers; i.e., the model is unpruned. The designation 10k-sample refers to models trained on the complete AIMD dataset containing 10 000 structures, whereas 100-sample denotes models trained solely on the first 100 frames of the melting trajectory.

where only a single message-passing layer is retained, AFMs fine-tuned on the 100-sample dataset can achieve a performance comparable to the Allegro baseline. These results partially substantiate our hypothesis that AFMs suffer from over-smoothing and contain deep message-passing layers with marginal contribution, thus supporting the proposed pruning strategy.

To further understand why pruning works, we analyzed the per-layer embeddings of the AFMs. We uniformly sampled 100 Li3_3PO4_4 structures from the 10k-sample dataset and collected the node embeddings produced after each successive message-passing layer for MatterSim, ORB, and JMP. On these per-layer embeddings we computed mean feature shift and centered kernel alignment (CKA) between every layer ii and layer i+1i+1 as shown in Figure 2. The results for MatterSim and ORB exhibit a clear depth-dependent attenuation of feature updates: successive layers induce progressively smaller shifts and reduced diversity, indicating gradual convergence of the node representations, which is a sign of over-smoothing. Similar trends are not as pronounced for the JMP model, but the results in Table 1 indicate that the same pruning strategy is still applicable to JMP.

Image not extracted

Figure 2. (a). Mean feature shift between successive message-passing layers for MatterSimV1-5M, ORB-V3-Omat, and JMP-S on Li3_3PO4_4. Smaller values indicate weaker updates and greater feature stability. (b). Linear CKA between successive message-passing layers for the same models. Values approaching 1.0 indicate higher similarity of node features across layers. (c). Radial distribution functions (RDF) of the quenched Li3_3PO4_4 structure containing 421 824 atoms. The red RDF curve was obtained from an MD trajectory driven by MatterSimV1-1M-MP×\times1 model trained on the 100-sample dataset, whereas the black curve corresponds to the AIMD results [25].

The pruned models substantially reduce parameter count and architectural complexity; when coupled with multi-GPU parallel acceleration this removes prior system-size bottlenecks and enables large-scale simulations. We evaluated the MD throughput achievable from several pruned variants under single and multi-GPU execution on Li3_3PO4_4 systems of varying sizes as shown in Table 2. We observed that pruning the models led to a substantial speedup compared to their unpruned counterparts. On top of this, multi-GPU parallelization can provide further acceleration for MD simulations, yielding more than 60% speedup for MatterSim V1-1M and over 50% for MACE-M-Omat.

192 atoms41 472 atoms421 824 atoms943 296 atoms5 184 000 atoms
GPU×1\mathrm{GPU}\times1GPU×1\mathrm{GPU}\times1GPU×1\mathrm{GPU}\times1GPU×2\mathrm{GPU}\times2GPU×4\mathrm{GPU}\times4GPU×1\mathrm{GPU}\times1GPU×1\mathrm{GPU}\times1
MatterSimV1-1MMP×1\mathrm{MP}\times1245.95102.8658.6333.1823.2253.8368.65
MP×3\mathrm{MP}\times3374.34374.44408.00204.08118.12××
ORB-V3-OmatMP×1\mathrm{MP}\times1212.4094.4177.1065.4060.67175.611161.92
MP×5\mathrm{MP}\times5448.36××××××
JMP-SMP×1\mathrm{MP}\times1264.642937.86×××××
MP×4\mathrm{MP}\times4482.24××××××
MACE-M-OmatMP×1\mathrm{MP}\times1322.29139.89111.5370.5251.41113.46957.66
MP×2\mathrm{MP}\times2533.121102.78955.37500.82276.86××

Table 2. MD step time (μs/(atom×step)\mu\mathrm{s}/(\mathrm{atom}\times\mathrm{step})) of various AFMs on Li3PO4\mathrm{Li}_{3}\mathrm{PO}_{4} structures of different sizes and GPU counts (NVIDIA A100-40GB PCIe)

Furthermore, Table 2 highlights another advantage of our method: since predictions on the partitioned subgraphs are completely independent, the approach can also process the subgraphs sequentially on a single GPU. This feature is particularly useful for large-scale simulations in resource-constrained situations. Here we show that simulations of at least 5 million atoms can be performed even on a single GPU using method, far exceeding the size constraints of currently available GNN-based MLIPs (by comparison, Allegro can simulate between 200,000 to 500,000 atoms on a single 80GB A100 GPU, while larger GNNs such as MACE struggle beyond 100,000 atoms [51]).

Reported values represent the average molecular dynamics step time in microseconds per atom per step. All benchmarks were conducted on NVIDIA A100-40 GB PCIe GPUs using the ASE implementation of Langevin dynamics. The symbol “×” indicates that even after partitioning the structure into more than 1000 subgraphs, the model could not fit within the 40 GB GPU memory, suggesting that the model is unable or unsuitable to perform MD simulations at this scale.

Considering the trade-off between accuracy and speed, we selected the MatterSimV1-1M-MP×1 model fine-tuned on the 100-sample dataset to reproduce the Li3PO4\mathrm{Li}_{3}\mathrm{PO}_{4} quenching experiment. The simulation employed the Langevin thermostat implemented in ASE [58], cooling the melt from 3000K to 600K over 50ps. From the resulting trajectory we computed the radial distribution function (RDF) of Li3PO4\mathrm{Li}_{3}\mathrm{PO}_{4} at 600K and compared it with the AIMD reference as shown in Figure 2. The RDF matches the AIMD reference in both peak positions and relative amplitudes, while likewise lacking significant correlations past 5Å, indicating correct short and medium range structure without spurious ordering. This confirms that AFM pruned by our MPDP method not only maintains accuracy metrics in terms of energies and forces, but also preserves key physical observables critical for MD reliability.

Nanoscale Segregation Phenomenon in the Mixed Silicon-silica System

SiO2_2 is known to lack a stable crystalline phase under ambient conditions and will readily disproportionate. Recent DFT-based explorations of possible ordered phases of homogeneous SiO have shown that a mixture of crystalline Si and SiO2_2 is enthalpically more stable than SiO [59]. Subsequently, Erhard et al. [60] employed an active learning workflow, sampling MD trajectories to construct a SiOx_x dataset and trained an ACE model, and was able to reproduce the segregation process of SiO into crystalline Si and SiO2_2 via MD simulations. The modelling of the segregation process there involved lengthy melting and quenching simulations of SiO and is more complex than the Li3_3PO4_4 quench. We therefore chose to reproduce this simulation to assess the robustness of the pruned AFMs under extreme, out-of-distribution conditions.

To highlight the data-efficiency and robustness of AFMs, we randomly sampled 2% of the active learning data from [60] to construct a small training set. We then applied varying degrees of pruning to MatterSim as we did before. After fine-tuning each variant on the 2% subset, we evaluated them on the separate test sets provided in [60] and compared the performance with the ACE-Complex benchmark model, which was trained on all of the available data; the results are summarized in Table 3. The single-layer variant, MatterSimV1-1M-MP×\times1, attains force-prediction accuracy comparable to the ACE-Complex benchmark on most test sets, whereas the two-layer variant, MatterSimV1-1M-MP×\times2, surpasses the benchmark on all test sets. A caveat here is that the energy predictions of MatterSim fine-tuned on the 2% subset are less accurate than the forces. We observe the same trend in Table 1: AFMs fine-tuned on a small subset of the full dataset tend to exhibit larger energy deviations. We attribute this primarily to sampling-induced bias in the training energy distribution. However, in the context of the MD simulations, we find that these slightly higher energy errors did not materially affect the trajectories.

ACE-ComplexMatterSim–V1–1M MP×\times1MatterSim–V1–1M MP×\times2
RMSEE_ERMSEF_FRMSEE_ERMSEF_FRMSEE_ERMSEF_F
SiO2_2 crystals0.90.05120.600.08146.360.04
a-SiO2_2 (CHIK-MD)2.20.19115.600.2930.800.16
a-SiO2_2 (GAP-MD)4.60.10227.450.1233.760.05
a-SiO2_2 (ACE-MD)3.20.18228.320.1133.420.06
a-SiO2_2 surfaces4.70.16141.630.1534.510.10
a-Si51.50.26417.810.24386.620.20
a-SiOx_x38.00.43214.290.42262.980.32
high-p a-SiO2_24.60.2435.840.2767.190.20

Table 3. Energy and force root mean square error (RMSE) of various pruned AFMs fine-tuned on 2% of the Si-O dataset and their comparison with the ACE baseline, reported in units of [meV] and [meV/Å]

All AFMs reported in this table are fine-tuned with 2% of the entire dataset. The ACE-Complex model and its RMSE results are taken from the best-performing model reported in [60].

For the segregation simulations, we chose the MatterSimV1-1M-MP×\times1 model and performed MD simulations with ASE. A 43 904-atom Si/SiO2_2 mixture was first heated in an NVT ensemble to 4000 K and held for 10 ps, then equilibrated for 100 ps at 2300 K in an NPT ensemble at zero external pressure to ensure complete melting and mixing. The system was subsequently quenched to 300 K under zero pressure NPT using four cooling rates ranging from 101210^{12} K/s to 101310^{13} K/s, followed by a 10 ps NPT hold at ambient temperature. The resulting structures are visualized in Figure 3a. The simulations reproduce the nanoscale segregation in which distinct a-Si and a-SiO2_2 regions appear and the average Si-rich grain size increases systematically with decreasing quench rates. We further calculated the structure factor S(q)S(q) of the room-temperature configuration generated at a quench rate of 5×10135 \times10^{13} K/s and compared it with the corresponding experimental data as shown in Figure 3b. These results show the pruned AFMs can fully reproduce the observed phenomena even with 2% of the training data.

Image not extracted

Figure 3. (a). Visualization of SiO structures quenched with cooling rates ranging from 101210^{12} K/s to 101310^{13} K/s (top left to bottom right: 1×10131 \times10^{13}, 5×10125 \times10^{12}, 2×10122 \times10^{12}, 1×10121 \times10^{12} K/s). (b). Structure factor S(q)S(q) of room-temperature SiO obtained from the simulation quenched at 5×10125 \times10^{12} K/s, compared with the experimental reference [61]. The structure factor was computed using code provided in [60]

Metal–Support Interaction and Suboxide Encapsulation of Pt Nanocluster on TiO2 Surface

Strong metal–support interaction (SMSI) denotes the formation, under reducing or oxidizing conditions, of ultrathin overlayers of substoichiometric oxide species that migrate from an oxide support onto the supported nanoparticles, producing a coating of one to several atomic layers that partially or extensively covers the metal. This phenomena has important implications in catalysis, as SMSI can significantly alter the catalytic properties and stability of the supported nanoparticle catalysts [62, 63, 64, 65]. Modeling SMSI at the atomic level can provide key insights into the mechanisms and characteristics of this process, providing guidance into preventing or controlling this behavior to enable more durable and active catalysts.

To assess whether our approach can reliably model these heterogeneous morphologies—slabs, nanoclusters, and their interfaces, we performed a simulation study of the SMSI process of a Pt nanocluster on a TiO2 support under reducing conditions. We use a training dataset generated with DFT calculations containing smaller Pt nanoclusters of several diameters on TiOx_x (110) slabs and running ab-initio MD between 300 K and 2000 K, collecting 16 200 configurations in total. We then fine-tuned a MatterSimV1-1M-MP×\times1 model on these data. For the simulation, we constructed a system comprising a 4 213-atom Pt cluster atop a 38 400-atom TiO2 slab, representing Pt nanoparticles at experimentally observed diameters. To emulate the reducing environment characteristic of catalytic operations, we uniformly removed 6.25% of the oxygen atoms from the TiO2 support, thereby introducing oxygen vacancies. We carried out an NVT simulation at an elevated temperature of 1500 K with a 2 fs time step to then observe the SMSI process. In Figure 4 we visualized some snapshots of the trajectory. After 2.5 ns, we can observe the Pt cluster to be fully encapsulated by Ti\mathrm{Ti}–O\mathrm{O} suboxide species, which is consistent with experimental observations of SMSI-induced encapsulation. These initial results demonstrate the promise large-scale simulations with AFMs can provide towards exploring these complex phenomena for the first time, though additional work is needed to rigorously validate these simulation results with additional experimental studies.

Image not extracted

Figure 4. Snapshots from an MD trajectory showing the progressive encapsulation of a Pt nanocluster by Ti\mathrm{Ti}–O\mathrm{O} suboxide species on an oxygen-deficient TiO2\mathrm{TiO}_{2} support. The simulation was driven by the MatterSimV1-1M-MP×1 model using ASE Langevin dynamics. After 2.5 ns, the Pt cluster is fully covered.

Conclusion

In this work, we introduced a general acceleration strategy and framework for AFMs that combines our proposed message-passing depth pruning with graph-partitioned multi-GPU parallelism. This strategy directly addresses two key obstacles to deploying AFMs in realistic molecular dynamics simulations: slow inference speeds and high memory demands. By simplifying AFM backbones to the most informative layers and distributing inference efficiently across GPUs, our framework makes large-scale and long-time simulations practical while preserving the hallmark accuracy and data efficiency of AFMs. Our experiments on three representative systems highlight both the effectiveness and the robustness of this approach across a broad range of different applications. In Li3PO4\mathrm{Li}_{3}\mathrm{PO}_{4}, we showed that AFMs trained on extremely small datasets can achieve competitive accuracy, and pruning further reduced complexity without compromising fidelity. In silicon-silica system and Pt/TiO2\mathrm{TiO}_{2} systems, we successfully capture the segregation phenomena and the metal–support encapsulation in defective oxide supports, demonstrating simulation robustness for challenging dynamics and complex heterogeneous structures. These capabilities have been integrated within the open-source MatterTune package.

Beyond the practical benefits in accelerating simulations, our findings also provide some conceptual insights for AFM design. The embedding analyses experiment revealed clear evidence of over-smoothing in AFMs: deeper message-passing layers contribute marginally to representation quality. This likely suggests that more message-passing layers do not necessarily translate into more useful information, and tailoring model depth to the target task may yield more efficient and equally accurate architectures. Further work is warranted in exploring oversmoothing in AFMs in the future.

Methods

Message Passing Depth Pruning

Figure 1a schematically illustrates our strategy for constructing lightweight models by pruning message-passing layers. Consider a generic AFM backbone that employs MM message-passing layers to generate node and edge embeddings. Our procedure simply truncates the backbone after the first KK layers and feeds the resulting node and edge features directly into the property-prediction heads. This approach is applicable to the vast majority of AFM architectures. For certain models, such as JMP, which predict properties by concatenating all intermediate states before a global read-out rather than using only the final node and edge features, we preserve dimensional consistency without modifying the read-out module by padding the missing intermediate states from K+1K+1 to MM layer with zero tensors.

Image not extracted

Figure 1. Overview of the pruning-and-partition acceleration workflow for AFMs. (a) Message-passing depth pruning. A pre-trained AFM backbone with M message-passing layers is truncated to the first K layers (K<M). The node and edge features after layer K feed directly into output heads for energy, forces, and stress. (b) Graph partitioning. The atomic graph is split by red dashed lines; the nodes in each piece define the set of root nodes. To preserve the receptive field of the pruned model, each piece is expanded to include the K-hop neighbors of its root nodes, following the original adjacency, yielding extended subgraphs. (c) Multi-GPU inference. The extended subgraphs are assigned to different GPUs for independent forward passes, and their outputs are gathered to produce predictions for the full structure.

Graph Partitioning

Our partitioning strategy ensures that each root node preserves its KK-hop neighborhood while enabling parallel evaluation. The partitioning procedure proceeds in two stages.

Stage 1. The nodes in the input atomic graph are split into disjoint sets that together cover all atoms; the atoms in each fragment constitute the root nodes of a partition. For bulk systems or other nearly uniform densities, we implemented an extremely fast spatial grid-based method: the unit cube is divided into a uniform grid of PP bins (with PP the desired number of partitions), and each atom is assigned to a bin based on its scaled position within the unit cube.

However, for heterogeneous geometries such as nanoparticles on supports, spatial partitioning alone can produce highly unbalanced or fragmented subgraphs. In such cases we have more advanced graph-based algorithms. One such algorithm, spectral partitioning, leverages the eigenvectors of the graph Laplacian to guide cuts that minimize inter-partition edges while maintaining balanced subgraphs. This approach yields high-quality partitions, but as the graph size increases, the cost of computing eigenvectors becomes prohibitive, limiting its applicability to very large systems. Thus, we instead choose to rely on METIS [66], a multilevel partitioner that approximates spectral cuts while scaling nearly linearly in the number of nodes, yielding balanced and compact subgraphs. In section S1, we present a comparative analysis of different graph partitioning methods.

Image not extracted

Figure 5. Example of partitioning and neighborhood expansion. Nodes are first grouped into P=3P = 3 disjoint partitions (colors). Boundary nodes are then expanded by a BFS of depth K=1K = 1, yielding overlapping regions (rings) that preserve local neighborhoods across partition boundaries

Stage 2. Each subgraph is extended by a breadth first search (BFS) whose depth is equal to the pruning parameter KK. This expansion ensures that, after partitioning, every root node retains exactly the same KK-hop effective neighborhood it would have in the full, unpartitioned structure. To avoid any performance penalty, the BFS expansion is implemented in Cython [67]. In our simulation of Li3PO4\mathrm{Li}_3\mathrm{PO}_4 containing 421 824 atoms, both partitioning stages together account for less than 2% of the wall-time per step.

Global Output Integration

After the model completes inference on each partitioned subgraph, the predicted quantities must be consistently assembled into global outputs for the full input structure, including energy, forces, and stress. For the total energy, most current AFMs decompose the system energy EE into a sum of per-atom contributions, i.e., E=∑iEiE = \sum_i E_i. Accordingly, energy aggregation reduces to simply summing up the per-atom energy contributions of the root nodes across all subgraphs.

For the forces, AFMs adopt two major strategies: direct prediction or energy-gradient prediction. In direct prediction, node or edge features are fed into a dedicated output head to predict forces without backpropagation. This yields computational advantages in terms of speed but does not guarantee a conservative potential energy surface. For this case, global forces can be obtained by concatenating the predicted forces of root nodes from all subgraphs. In energy-gradient prediction, forces are computed as derivatives of the energy with respect to atomic positions, i.e., F=−∇EF = -\nabla E. Since the total energy is additive over per-atom contributions, the chain rule allows us to express ∇E=∑i∇Ei\nabla E = \sum_i \nabla E_i. Thus, it suffices to compute the gradients of root-node energies with respect to atomic positions within each subgraph and then sum these contributions across subgraphs to obtain the global force field.

For stress predictions, a similar decomposition applies when forces are obtained via the energy-gradient approach. However, in the case of directly predicted forces, stress cannot be consistently reconstructed within our current framework.

A Unified ASE-Compatible Multi-GPU Calculator

We employ PyTorch Lightning’s Distributed Data Parallel (DDP) strategy [68, 69] to implement multi-GPU parallel inference, and based on this, we developed a unified ASE-compatible calculator [58] that supports accelerated prediction with multiple GPUs across a wide range of AFM models. This calculator, along with the model pruning and graph partitioning methods proposed in this work, has been fully integrated into the MatterTune platform [40] for public use.

Mean Feature Shift and Linear CKA

The mean feature shift is an average magnitude of the change in node features when moving from layer ii to layer i+1i+1 [70, 71]. To eliminate the effect of global translation, we normalized the feature matrices by subtracting the mean. Let H(i)∈RN×dH^{(i)} \in\mathbb{R}^{N \times d} and H(i+1)∈RN×dH^{(i+1)} \in\mathbb{R}^{N \times d} be the normalized node feature matrices of all NN nodes after successive message-passing layers. We compute mean feature shift as:

MFS⁡(i→i+1)=1N∥H(i+1)−H(i)∥F\operatorname{MFS}^{(i \to i+1)} = \frac{1}{N}\left\lVert H^{(i+1)} - H^{(i)} \right\rVert_F

where ∥⋅∥F\lVert\cdot\rVert_F is the Frobenius norm.

Linear CKA measures the similarity of two whole feature spaces, insensitive to isotropic scaling or orthogonal rotation [72]. Values that are close to 1 indicate that two feature spaces are nearly identical up to rotation/scale. To get linear CKA, we firstly center each feature matrix by H~=H−1N11TH\widetilde{H} = H - \frac{1}{N}\mathbf{1}\mathbf{1}^{T}H. Then the CKA is computed as:

CKA⁡(i→i+1)=∥H~(i)TH~(i+1)∥F2∥H~(i)TH~(i)∥F∥H~(i+1)TH~(i+1)∥F\operatorname{CKA}^{(i \to i+1)} = \frac{\left\lVert\widetilde{H}^{(i)T}\widetilde{H}^{(i+1)} \right\rVert_F^2}{\left\lVert\widetilde{H}^{(i)T}\widetilde{H}^{(i)} \right\rVert_F \left\lVert\widetilde{H}^{(i+1)T}\widetilde{H}^{(i+1)} \right\rVert_F}

Acknowledgements

The support of this work by the National Science Foundation grant CBET-2442223 is gratefully acknowledged. This research used resources of the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231 using NERSC award BES-ERCAP0032102.

Code Availability

The MatterTune platform, used to perform the experiments in this paper, are available on Github at https://github.com/Fung-Lab/MatterTune.

Supplementary Information

Comparison of Graph Partitioning Methods

To assess both the runtime characteristics of different graph partition methods and their behavior across structures with distinct geometries, we performed a head-to-head evaluation. For each system size, we partitioned the atomic graph into 8 subgraphs. Under this common setting, we recorded the wall-clock time required by three algorithms—spatial grid partitioning, METIS, and spectral partitioning (Figure S1). Spectral partitioning becomes prohibitively expensive once the system size exceeds several tens of thousands of atoms, whereas METIS and the spatial grid exhibit substantially flatter scaling; owing to its simplicity, the spatial grid is consistently the fastest.

Image not extracted

Figure S1. Runtime comparison of different graph partitioning algorithms on Li3_3PO4_4 structures of varying sizes.

We further compared the quality of partitions produced by METIS and the spatial grid using the halo growth factor HH on Li3_3PO4_4 and Pt-TiO2_2 systems of various sizes (Figure S2). We define the halo growth factor HH as the ratio of the atoms introduced by halo expansion to the number of atoms in the original subgraph. Smaller HH indicates less overhead from halo expansion and thus higher-quality partition boundaries. On the homogeneous bulk Li3_3PO4_4 structures, METIS holds only a slight advantage over the spatial grid. In contrast, for the heterogeneous Pt-TiO2_2, which are spherical particles placed on planar slabs, METIS achieves a noticeably lower halo growth factor, indicating a clearer advantage.

Image not extracted

Figure S2. Comparison of halo growth factor across graph partitioning algorithms for (a) Li3PO4\mathrm{Li}_3\mathrm{PO}_4 structures and (b) Pt-TiO2\mathrm{TiO}_2 structures, evaluated over varying system sizes.

References

  1. [1]Tao Cheng, Hai Xiao, and William A Goddard III. Full atomistic reaction mechanism with kinetics for co reduction on cu (100) from ab initio molecular dynamics free-energy calculations at 298 k. Proceedings of the National Academy of Sciences, 114(8):1795–1800, 2017.DOI
  2. [2]Jia-Bo Le, Ao Chen, Lang Li, Jing-Fang Xiong, Jinggang Lan, Yun-Pei Liu, Marcella Iannuzzi, and Jun Cheng. Modeling electrified pt(111)-had/water interfaces from ab initio molecular dynamics. JACS Au, 1(5):569–577, 2021. PMID: 34467320.DOI
  3. [3]Bingqing Cheng, Edgar A Engel, Jörg Behler, Christoph Dellago, and Michele Ceriotti. Ab initio thermodynamics of liquid and solid water. Proceedings of the National Academy of Sciences, 116(4):1110–1115, 2019.
  4. [4]Yifei Mo, Shyue Ping Ong, and Gerbrand Ceder. First principles study of the li10gep2s12 lithium super ionic conductor material. Chemistry of Materials, 24(1):15–17, 2012.
  5. [5]Meirbek Islamov, Hasan Babaei, Ryther Anderson, Kutay B Sezginel, Jeffrey R Long, Alan JH McGaughey, Diego A Gomez-Gualdron, and Christopher E Wilmer. High-throughput screening of hypothetical metal-organic frameworks for thermal conductivity. npj Computational Materials, 9(1):11, 2023.
  6. [6]William D Richards, Tomoyuki Tsujimura, Lincoln J Miara, Yan Wang, Jae Chul Kim, Shyue Ping Ong, Ichiro Uechi, Naoki Suzuki, and Gerbrand Ceder. Design and synthesis of the superionic conductor na10snp2s12. Nature communications, 7(1):11009, 2016.DOI
  7. [7]Yifei Mo, Shyue Ping Ong, and Gerbrand Ceder. Insights into diffusion mechanisms in p2 layered oxide materials by first-principles calculations. Chemistry of Materials, 26(18):5208–5214, 2014.DOI
  8. [8]Pierre Hohenberg and Walter Kohn. Inhomogeneous electron gas. Physical review, 136(3B):B864, 1964.
  9. [9]Walter Kohn and Lu Jeu Sham. Self-consistent equations including exchange and correlation effects. Physical review, 140(4A):A1133, 1965.DOI
  10. [10]Krishnan Raghavachari, Gary W Trucks, John A Pople, and Martin Head-Gordon. A fifth-order perturbation comparison of electron correlation theories. Chemical Physics Letters, 157(6):479–483, 1989.DOI
  11. [11]Thomas L Beck. Real-space mesh techniques in density-functional theory. Reviews of Modern Physics, 72(4):1041, 2000.DOI
  12. [12]Mike C Payne, Michael P Teter, Douglas C Allan, TA Arias, and ad JD Joannopoulos. Iterative minimization techniques for ab initio total-energy calculations: molecular dynamics and conjugate gradients. Reviews of modern physics, 64(4):1045, 1992.DOI
  13. [13]Georg Kresse and Jürgen Furthmüller. Efficiency of ab-initio total energy calculations for metals and semiconductors using a plane-wave basis set. Computational materials science, 6(1):15–50, 1996.DOI
  14. [14]Dominik Marx and Jurg Hutter. Ab initio molecular dynamics: Theory and implementation. Modern methods and algorithms of quantum chemistry, 1(301-449):141, 2000.
  15. [15]Mark E Tuckerman. Ab initio molecular dynamics: basic concepts, current trends and novel applications. Journal of Physics: Condensed Matter, 14(50):R1297, 2002.
  16. [16]Albert P Bartók, Mike C Payne, Risi Kondor, and Gábor Csányi. Gaussian approximation potentials: The accuracy of quantum mechanics, without the electrons. Physical review letters, 104(13):136403, 2010.pubmed.ncbi.nlm.nih.gov/20481899
  17. [17]Kristof Schutt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert Muller. SchNet: A continuous-filter convolutional neural network for modeling quantum interactions. Advances in neural information processing systems, 30, 2017.
  18. [18]Kristof Schutt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International conference on machine learning, pages 9377–9388. PMLR, 2021.arxiv.org/abs/2102.03150
  19. [19]Han Wang, Linfeng Zhang, Jiequn Han, et al. Deepmd-kit: A deep learning package for many-body potential energy representation and molecular dynamics. Computer Physics Communications, 228:178–184, 2018.DOI
  20. [20]Ralf Drautz. Atomic cluster expansion for accurate and transferable interatomic potentials. Physical Review B, 99(1):014104, 2019.DOI
  21. [21]Chi Chen, Weike Ye, Yunxing Zuo, Chen Zheng, and Shyue Ping Ong. Graph networks as a universal machine learning framework for molecules and crystals. Chemistry of Materials, 31(9):3564–3572, 2019.DOI
  22. [22]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
  23. [23]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
  24. [24]Ilyes Batatia, David P Kovacs, Gregor Simm, Christoph Ortner, and Gábor Csányi. Mace: Higher order equivariant message passing neural networks for fast and accurate force fields. Advances in neural information processing systems, 35:11423–11436, 2022.
  25. [25]Albert Musaelian, 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
  26. [26]Yi-Lun Liao and Tess Smidt. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. arXiv preprint arXiv:2206.11990, 2022.arxiv.org/abs/2206.11990
  27. [27]Aidan P Thompson, H Metin Aktulga, Richard Berger, Dan S Bolintineanu, W Michael Brown, Paul S Crozier, Pieter J In’t Veld, Axel Kohlmeyer, Stan G Moore, Trung Dac Nguyen, et al. Lammps-a flexible simulation tool for particle-based materials modeling at the atomic, meso, and continuum scales. Computer physics communications, 271:108171, 2022.DOI
  28. [28]KT Schutt, Pan Kessel, Michael Gastegger, Kim A Nicoli, Alexandre Tkatсhenko, and K-R Muuller. Schnetpack: A deep learning toolbox for atomistic systems. Journal of chemical theory and computation, 15(1):448–455, 2018.DOI
  29. [29]Ivan S Novikov, Konstantin Gubaev, Evgeny V Podryabinkin, and Alexander V Shapeev. The mlip package: moment tensor potentials with mpi and active learning. Machine Learning: Science and Technology, 2(2):025002, 2020.arxiv.org/abs/2007.08555
  30. [30]Venkat Kapil, Mariana Rossi, Ondrej Marsalek, Riccardo Petraglia, Yair Litman, Thomas Spura, Bingqing Cheng, Alice Cuzzocrea, Robert H Meißner, David M Wilkins, et al. i-pi 2.0: A universal force engine for advanced molecular simulations. Computer Physics Communications, 236:214–223, 2019.
  31. [31]Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M Elena, Dávid P Kovács, Janosh Riebesell, Xavier R Advincula, Mark Asta, Matthew A Vaylon, William J Baldwin, et al. A foundation model for atomistic materials chemistry. arXiv preprint arXiv:2401.00096, 2023.arxiv.org/abs/2401.00096
  32. [32]Amil Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon, and Ekin Dogus Cubuk. Scaling deep learning for materials discovery. Nature, 624(7990):80–85, 2023.DOI
  33. [33]Duo Zhang, Xinzijian Liu, Xiangyu Zhang, Chengqian Zhang, Chun Cai, Hangrui Bi, Yiming Du, Xuejian Qin, Anyang Peng, Jiameng Huang, et al. Dpa-2: a large atomic model as a multi-task learner. npj Computational Materials, 10(1):293, 2024.DOI
  34. [34]Nima Shoghi, Adeesh Kolluru, John R Kitchin, Zachary W Ulissi, C Lawrence Zitnick, and Brandon M Wood. From molecules to materials: Pre-training large generalizable models for atomic property prediction. arXiv preprint arXiv:2310.16802, 2023.arxiv.org/abs/2310.16802
  35. [35]Han Yang, Chenxi Hu, Yichi Zhou, Xixian Liu, Yu Shi, Jielan Li, Guanzhi Li, Zekun Chen, Shuizhou Chen, Claudio Zeni, et al. Mattersim: A deep learning atomistic model across elements, temperatures and pressures. arXiv preprint arXiv:2405.04967, 2024.arxiv.org/abs/2405.04967
  36. [36]Mark Neumann, James Gin, Benjamin Rhodes, Steven Bennett, Zhiyi Li, Hitarth Choubisa, Arthur Hussey, and Jonathan Godwin. Orb: A fast, scalable neural network potential, 2024.arxiv.org/abs/2410.22570
  37. [37]Benjamin Rhodes, Sander Vandenhaute, Vaidotas Šimkus, James Gin, Jonathan Godwin, Tim Duignan, and Mark Neumann. Orb-v3: atomistic simulation at scale, 2025.arxiv.org/abs/2504.06231
  38. [38]Yi-Lun Liao, Brandon Wood, Abhishek Das, and Tess Smidt. EquiformerV2: Improved equivariant transformer for scaling to higher-degree representations. arXiv preprint arXiv:2306.12059, 2023.arxiv.org/abs/2306.12059
  39. [39]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. arXiv preprint arXiv:2502.12147, 2025.arxiv.org/abs/2502.12147
  40. [40]Lingyu Kong, Nima Shoghi, Guoxiang Hu, Pan Li, and Victor Fung. Mattertune: An integrated, user-friendly platform for fine-tuning atomistic foundation models to accelerate materials simulation and discovery. arXiv preprint arXiv:2504.10655, 2025.arxiv.org/abs/2504.10655
  41. [41]Mariia Radova, Wojciech G Stark, Connor S Allen, Reinhard J Maurer, and Albert P Bartók. Fine-tuning foundation models of materials interatomic potentials with frozen transfer learning. npj Computational Materials, 11(1):237, 2025.
  42. [42]Harveen Kaur, Flaviano Della Pia, Ilyes Batatia, Xavier R Adyincula, Benjamin X Shi, Jing-gang Lan, Gábor Csányi, Angelos Michaelides, and Venkat Kapil. Data-efficient fine-tuning of foundational models for first-principles quality sublimation enthalpies. Faraday Discussions, 256:120–138, 2025.arxiv.org/abs/2405.20217
  43. [43]Xiaochen Du, James K Damewood, Jaclyn R Lunger, Reisel Millan, Bilge Yildiz, Lin Li, and Rafael Gómez-Bombarelli. Machine-learning-accelerated simulations to enable automatic surface reconstruction. Nature Computational Science, 3(12):1034–1044, 2023.DOI
  44. [44]John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al. Highly accurate protein structure prediction with alphafold. nature, 596(7873):583–589, 2021.DOI
  45. [45]Tan Shi, Qingsong Tu, Yaosen Tian, Yihan Xiao, Lincoln J Miara, Olga Kononova, and Gerbrand Ceder. High active material loading in all-solid-state battery electrode via particle size optimization. Advanced Energy Materials, 10(1):1902881, 2020.DOI
  46. [46]Tucker Holstun, Tara P Mishra, Liliang Huang, Han-Ming Hau, Shashwat Anand, Xiaochen Yang, Colin Ophus, Karen Bustillo, Lu Ma, Steven Ehrlich, et al. Accelerating the electrochemical formation of the δ phase in manganese-rich rocksalt cathodes. Advanced Materials, 37(6):2412871, 2025.
  47. [47]Qimai Li, Zhichao Han, and Xiao-Ming Wu. Deeper insights into graph convolutional networks for semi-supervised learning. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018.arxiv.org/abs/1801.07606
  48. [48]Kenta Oono and Taiji Suzuki. Graph neural networks exponentially lose expressive power for node classification. arXiv preprint arXiv:1905.10947, 2019.arxiv.org/abs/1905.10947
  49. [49]Chen Cai and Yusu Wang. A note on over-smoothing for graph neural networks. arXiv preprint arXiv:2006.13318, 2020.arxiv.org/abs/2006.13318
  50. [50]Deli Chen, Yankai Lin, Wei Li, Peng Li, Jie Zhou, and Xu Sun. Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3438–3445, 2020.arxiv.org/abs/1909.03211
  51. [51]Paul Fuchs, Weilong Chen, Stephan Thaler, and Julija Zavdavljav. chemtrain-deploy: a parallel and scalable framework for machine learning potentials in million-atom md simulations. Journal of Chemical Theory and Computation, 21(15):7550–7560, 2025.DOI
  52. [52]James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, et al. Jax: composable transformations of python+ numpy programs. 2018.
  53. [53]Kevin Han, Bowen Deng, Amir Barati Farimani, and Gerbrand Ceder. Distmlip: A distributed inference platform for machine learning interatomic potentials. arXiv preprint arXiv:2506.02023, 2025.arxiv.org/abs/2506.02023
  54. [54]Xiaohua Yu, JB Bates, GE Jellison, and FX Hart. A stable thin-film lithium electrolyte: lithium phosphorus oxynitride. Journal of the electrochemical society, 144(2):524, 1997.DOI
  55. [55]JB Bates, NJ Dudney, B Neudecker, A Ueda, and CD Evans. Thin-film lithium and lithium-ion batteries. Solid state ionics, 135(1-4):33–45, 2000.DOI
  56. [56]Andrew S Westover, Andrew K Kercher, Mordechai Kornbluth, Michael Naguib, Max J Palmer, David A Cullen, and Nancy J Dudney. Plasma synthesis of spherical crystalline and amorphous electrolyte nanopowders for solid-state batteries. ACS Applied Materials & Interfaces, 12(10):11570–11578, 2020.DOI
  57. [57]Wenwen Li, Yasunobu Ando, Emi Minamitani, and Satoshi Watanabe. Study of li atom diffusion in amorphous li3po4 with neural network potential. The Journal of chemical physics, 147(21), 2017.DOI
  58. [58]Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E Castelli, Rune Christensen, Marcin Dulak, Jesper Friis, Michael N Groves, Björk Hammer, Cory Hargus, et al. The atomic simulation environment—a python library for working with atoms. Journal of Physics: Condensed Matter, 29(27):273002, 2017.DOI
  59. [59]Khalid AlKaabi, Dasari LVK Prasad, Peter Kroll, NW Ashcroft, and Roald Hoffmann. Silicon monoxide at 1 atm and elevated pressures: crystalline or amorphous? Journal of the American Chemical Society, 136(9):3410–3423, 2014.DOI
  60. [60]Linus C Erhard, Jochen Rohrer, Karsten Albe, and Volker L Deringer. Modelling atomic and nanoscale structure in the silicon–oxygen system through active machine learning. Nature Communications, 15(1):1927, 2024.
  61. [61]Akihiko Hirata, Shinji Kohara, Toshihiro Asada, Masazumi Arao, Chihiro Yogi, Hideto Imai, Yongwen Tan, Takeshi Fujita, and Mingwei Chen. Atomic-scale disproportionation in amorphous silicon monoxide. Nature communications, 7(1):11591, 2016.DOI
  62. [62]SJ Tauster, SC Fung, RTK Baker, and JA Horsley. Strong interactions in supported-metal catalysts. Science, 211(4487):1121–1125, 1981.pubmed.ncbi.nlm.nih.gov/17755135
  63. [63]Shaofeng Liu, Wei Xu, Yiming Niu, Bingsen Zhang, Lirong Zheng, Wei Liu, Lin Li, and Junhu Wang. Ultrastable au nanoparticles on titania through an encapsulation strategy under oxidative atmosphere. Nature communications, 10(1):5790, 2019.DOI
  64. [64]Chuanbo Gao, Fenglei Lyu, and Yadong Yin. Encapsulated metal nanoparticles for catalysis. Chemical Reviews, 121(2):834–881, 2020.DOI
  65. [65]Felipe Polo-Garzon, Thomas F Blum, Zhenghong Bao, Kristen Wang, Victor Fung, Zhennan Huang, Elizabeth E Bickel, De-en Jiang, Miaofang Chi, and Zili Wu. In situ strong metal–support interaction (smsi) affects catalytic alcohol conversion. ACS Catalysis, 11(4):1938–1945, 2021.
  66. [66]George Karypis and Vipin Kumar. A fast and high quality multilevel scheme for partitioning irregular graphs. SIAM Journal on scientific Computing, 20(1):359–392, 1998.DOI
  67. [67]Stefan Behnel, Robert Bradshaw, Craig Citro, Lisandro Dalcin, Dag Sverre Seljebotn, and Kurt Smith. Cython: The best of both worlds. Computing in Science & Engineering, 13(2):31–39, 2010.DOI
  68. [68]William A Falcon. Pytorch lightning. GitHub, 3, 2019.DOI
  69. [69]Shen Li, Yanli Zhao, Rohan Varma, Omkar Salpekar, Pieter Noordhuis, Teng Li, Adam Paszke, Jeff Smith, Brian Vaughan, Pritam Damania, et al. Pytorch distributed: Experiences on accelerating data parallel training. arXiv preprint arXiv:2006.15704, 2020.arxiv.org/abs/2006.15704
  70. [70]Vijay Prakash Dwivedi, Chaitanya K Joshi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Benchmarking graph neural networks. Journal of Machine Learning Research, 24(43):1–48, 2023.
  71. [71]T Konstantin Rusch, Michael M Bronstein, and Siddhartha Mishra. A survey on oversmoothing in graph neural networks. arXiv preprint arXiv:2303.10993, 2023.arxiv.org/abs/2303.10993
  72. [72]Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. In International conference on machine learning, pages 3519–3529. PMLR, 2019.arxiv.org/abs/1905.00414

Paper details

vlnc:W_9vbadtgx
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents