betaFeedback

WorksComputer Science

Toward Exascale AI for Science: A Scalable AI Skill for Autonomous Microkinetics Discovery

Abstract

We present a scalable AI-driven framework that advances autonomous scientific discovery by combining agentic workflow automation, high-performance computing, and scientific surrogate models. Using microkinetics discovery as a testbed, the work demonstrates how AI can reduce expert intervention, recover from failed simulations, and systematically evaluate surrogate model reliability. This study shows how AI skills can transform complex domain workflows into robust, scalable capabilities for next-generation materials research.

CCS Concepts

• Applied computing → Engineering; • Computing methodologies → Multi-agent systems; Parallel algorithms.

ACM Reference Format: Ken-ichi Nomura, William Dawson, Kai Ito, Nabankur Dasgupta, Taufeq Mohammed Razakh, Thomas Linker, and Aiichiro Nakano. 2026. Toward Exascale AI for Science: A Scalable AI Skill for Autonomous Microkinetics Discovery. In Proceedings of ACM AI Leadership Summit 2026 (AI Summit ’26). ACM, New York, NY, USA, 6 pages. https://doi.org/XXXXXXX.XXXXXXX

Introduction

A grand challenge in science is to understand complex materials from first principles – a challenge best addressed by multiscale simulation [1]. Today, the state of the art (SOTA) is combined deployment of multi-exaflop quantum mechanical (QM) simulations with surrogate foundation models [2, 3], a formidable modeling task at the intersection of high-performance computing (HPC), AI, and applied science. A prototypical use case is reaction kinetics of complex material structures [4, 5], a regime where even small scale QM simulations are laborious and the establishment of robust foundation models remains unrealized.

Agentic AI – characterized by autonomy, planning, and task execution – promises to vault materials research over these technical barriers through the automation of workflows that once only humans could manage. Agents deployed for scientific tasks integrate domain-specific knowledge and tools for long-term memory or information retrieval [6, 7], as well as interaction with external software [8, 9] or resources [10, 11]. Composable augmentation is typically achieved today through two means: the Model Context Protocol (MCP) [12] and agent skills [13, 14]. Skills in particular have gained traction as a lightweight abstraction layer to teach Large Language Models (LLMs) domain-specific and procedural knowledge. Embedding these capabilities into agent harnesses transforms general-purpose frameworks (e.g., OpenClaw [15], Nemo-Claw [16]) into "AI scientists." With the rapid adoption of automated AI workflows across the scientific community [17, 10], ensuring their consistency, reproducibility, and reliability has become an urgent priority [18].

Agent tools and skills to leverage HPC resources would accelerate materials research through three complementary pillars: agentic automation, materials-specific surrogate models, and large-scale simulations on exascale computers. Agents backed by open models running on national supercomputing infrastructure would in turn manage simulations on the same or complementary machines. Unfortunately, domain-specific frameworks – particularly incorporating HPC [19] – able to robustly perform simulations despite the sycophantic nature of LLMs [20], remain in their infancy. In this work, we realize this vision as an HPC-centric agent able to perform reaction kinetics simulations at scale, and evaluate its performance for comparing different universal models for material properties.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected].

Design of NEB skill

Figure 1. Design of NEB skill. (a) Directory and file organization of the five key steps. (b) Skill flowchart. A user initiates the skill by providing reactant and product coordinates. Locally executed tasks are indicated by blue boxes, while yellow boxes indicate HPC inference by LLM. During the Monitor and Retry step, an agent is given the system task (c), and decides strategies (d) to improve the success rate of NEB calculation.

Methodology

Skill Design

To explore the potential of agent skills, we have developed “nebskill” [21] for Nudged Elastic Band (NEB) [22, 23] calculations. NEB searches for the minimum energy pathway (MEP) between a reactant and product to determine the transition state and activation energy of a chemical reaction. Although the methodology is well established, NEB calculations are highly sensitive to the choice of parameters and the system of interest. This dependency often turns conceptually simple work into labor-intensive jobs that take days, if not weeks. Hence, agentic automation offers the promise of increasing scientific productivity for these studies. Though fully autonomous AI scientists have garnered significant attention [18], their widespread adoption remains hindered by the lack of one-size-fits-all solutions; we therefore adopt a co-scientist approach for the ”nebskill”, emphasizing immediate, practical value for a community increasingly exploring agentic workflows.

Figure 1 (a) shows the organization of directories and files. The top directory contains a SKILL.md markdown file and each subdirectory contains another markdown file INSTRUCTIONS.md. This allows the agent to process the skill as it progresses, as well as to divide the entire task into smaller subtasks to reduce the chance of context rot [24]. The five key steps of the skill are shown below:

  • Step1: Initialization: setup environment (e.g., download MLIP models, verify reactant and product coordinates).

  • Step2: Relax reactant and product configurations.

  • Step3: Interpolate between the reactant and product, then run a two-phase calculation – standard NEB followed by Climbing-Image (CI) NEB.

  • Step4: Monitoring and failure recovery. Upon a calculation failure, LLM diagnoses the cause and proposes a new set of parameters (e.g., number of NEB images, spring constant).

  • Step5: Analysis and report. Methods from domain experts are applied to converged calculations to verify the chemistry.

When a job fails, the agent diagnoses the cause and proposes a new job to try (Step 4) following the ReAct (reasoning and acting) [25] approach. Based on observations from the failed job, the agent makes a decision on whether to exploit accumulated knowledge of the ongoing campaign (e.g., slightly update the value of spring constant) or to explore a new strategy (e.g., change the interpolation scheme from linear to spline). In this work, we use the ALCF (Argonne Leadership Computing Facility) Inference Endpoint, through which we access high-throughput AI services at scale. The API is called with a system prompt (Fig. 1c) along with the current NEB parameters and the diagnosis to propose a next strategy (Fig. 1d). Building on our previous integration of agents with supercomputer Fugaku [19], we performed scalability tests on the Perlmutter cluster at the National Energy Research Scientific Computing Center (NERSC). Under the US-Japan Genesis partnership, we plan to use these globally distributed supercomputers as a single metacomputer to perform grand challenge AI-simulation workflows otherwise impossible, leveraging our previous US-Japan metacomputing paradigm [26].

Surrogate Models and Dataset

Equivariant networks for interatomic potentials encode atomic environments by expanding interatomic displacements into spherical harmonics, yielding internal features that transform as proper geometric tensors under rotation [27, 28]. By enforcing equivariance throughout network architecture, the model does not need to relearn the same physics in arbitrarily rotated orientations, yielding substantially better data efficiency, smoother and more physically consistent force/stress predictions, and improved extrapolation to geometries and compositions outside the training distribution.

While many network architectures have been proposed to date [29], NequIP [30] and MACE [31] are widely used equivariant MLIP models based on message-passing scheme to build multi-body correlations over successive graph convolutions. The Allegro model [32] employs a strictly local feature approach where pairwise equivariant feature tensors are iteratively refined through tensor products with neighboring pairs sharing a central atom, preserving algorithmic scalability without sacrificing expressivity. Beyond these models supported by the AI skill, we expanded the evaluation pool to include a total of 12 MLIPs.

Activation energy along the reaction coordinate predicted by MLIP models

Figure 3. Activation energy along the reaction coordinate predicted by the default pretrained MLIP models compared against ground truth QM calculations (black). The horizontal axis is normalized by the total number of NEB images.

In this work, the agent examines a number of SOTA pretrained universal MLIP models as a surrogate of graphite surface microkinetics. We evaluate over ten SOTA MLIP models to find the model that best matches with a given ground truth result. Figure 2 presents the atomic configurations during the sublimation of CO2_2 from a graphite surface obtained by ab-initio quantum mechanical calculations and used to evaluate the developed AI skill.

Atomic configurations of reactant, transition state, and product

Figure 2. (a) Atomic configurations of reactant, (b) transition state, and (c) product obtained with ab-initio calculation. Red spheres for oxygen and gray ones for carbon atoms, respectively. The first step of NEB calculation is to smoothly interpolate the reactant and product configurations and generate an initial guess of intermediate states (i.e. NEB images).

QM calculation: We perform adiabatic quantum molecular dynamics simulations using the QXMD software [33]. A QMD simulation [34, 35] follows the trajectory of all atoms, while calculating interatomic forces quantum mechanically from first principles within the framework of density functional theory (DFT) [36]. The electronic states are calculated using the projector-augmented wave (PAW) method [37], where projector functions are generated for 2s22p22s^{2}2p^{2} for C, 2s22p42s^{2}2p^{4} for O atoms. The PBE generalized gradient approximation (GGA) is used for the exchange-correlation functional [38] and DFT-D2 for the dispersion correction [39]. The plane-wave cutoff energies are set to 30 Ry for the wavefunction and 300 Ry for the electron density in all calculations. We used 4×4×14 \times4 \times1 kk-points for Brillouin zone sampling.

Graphite model: The graphite model consists of two graphene layers. The graphene sheet consists of 4 unit cells with a lattice constant of 1.43 Å. To prevent interaction between periodic images along the zz direction due to the periodic boundary condition, a vacuum layer of thickness 14 Å is introduced along the zz axis. The supercell is orthorhombic with dimensions of 8.38×9.91×17.78.38 \times9.91 \times17.7 Å3^{3}.

Results

Reaction Coordinate and Activation Energy

Fig. 3 presents the energy profiles obtained by SOTA MLIP models together with DFT calculation as the ground truth. We evaluated pretrained universal MLIPs spanning the MACE, NequIP, Allegro, DPA-2/DPA-3, EquiformerV3, PET-MAD and MatRIS models [40, 30, 32, 41, 42, 43, 44, 45]. While quantitative comparisons may not be straightforward without fine-tuning, some of the MLIP models agree reasonably well with the ground truth calculation. For example, among all models, mace-mp-1 shows the closest activation energy, while the energy curve of nequip-oam and allegro-oam closely follow the DFT result. In terms of the reaction mechanism, all models predict the lactone-mediated CO2_2 sublimation mechanism that was also verified by other QM calculations [46, 47].

Scalability Test

To maximize parallel efficiency and improve time-to-solution, our agent skill employs two-dimensional parallelism over the MLIP model and NEB images. While the concurrency of model parallelism differs case by case, typically 5 to 10 MLIP models are evaluated simultaneously to improve the statistics. Since the model parallelism is loosely coupled, each MLIP evaluation is distributed over different nodes. In contrast, parallelization over NEB images is tightly coupled due to frequent atomic force exchanges. To minimize communication latency, the NEB images (typically 10 to 30) are distributed across the GPUs of a single node in a round-robin fashion to accelerate the overall calculation.

We have performed scalability tests of the AI skill on the Department of Energy (DOE) leadership computing facilities – specifically the Perlmutter cluster at NERSC and the Inference Endpoint at ALCF. The Perlmutter cluster is an HPE Cray system consisting of 3,072 CPU and 1,792 GPU nodes. Each GPU node hosts four NVIDIA A100 GPUs. The ALCF Inference Endpoint is powered by the Sophia (NVIDIA DGX A100) and Metis (SambaNova) clusters, and offers DOE scientists high-throughput, low-latency AI inference capabilities. Currently dozens of open foundation models are supported, including Google Gemma [48], Meta LLaMA [49], and OpenAI GPT-OSS [50].

Table 1 shows the improvement in runtime from baseline. All measurements were performed on Perlmutter at NERSC. While the scalability performance is impacted by many external factors such as file system access and network congestion, we have achieved a nearly perfect parallel efficiency, i.e. 3.75× speedup on four GPU nodes.

NodeGPUswallclock (min)speed up
14116.91.0 (baseline)
2884.31.39
31261.01.92
41632.13.75

Table 1. Scalability of distributed NEB computing

Failure and Retry Analysis

Because NEB calculations explore the energy landscape of target system, its success is highly dependent on researchers’ knowledge and familiarity of the target system. Researchers often figure out an optimal configuration after time-consuming trial-and-error efforts. Therefore, investigating the cause of NEB calculation failures, retry strategies, and its cost would provide useful insights of the system and actionable strategies for improvement. Here, we present a case study on failed NEB jobs and the robustness of MLIP model predictions.

We have identified 54 jobs in this study that failed in the first attempt but successfully converged at the end. We call them "resilient" NEB job. Of all, 51 jobs exhibited “image collapse” where the distance between each NEB image became too close, most likely attributed to the value of spring constant being too low. The remaining three jobs are “kinking” in the energy profile, exhibiting discontinuities or being very noisy.

Table 2 presents five representative cases out of the resilient job pool. The default strategy taken by the agent turned out to be adjusting the spring constant. While many jobs were successfully converged by one or a few attempts of tightening the spring constant, several resilient jobs went through multi-step parameter adjustments. An example is mace-mp shown in Table 2 where the agent chose to increase the spring constant twice consecutively, i.e. 0.3 → 0.5 → 1.2, then increased the NEB image size by 2 from 13. Finally, the agent decreased the spring constant from 1.2 → 0.5 and successfully achieved the energy convergence. Token usage averaged approximately 1,600 per retry, with communication times ranging from 6 to 14 seconds, signifying sound throughput performance across the DOE leadership computing facilities.

MLIPretrytokenscomm. (s)
mace-mpadjust spring664223.17
adjust spring
set N images
adjust spring
allegro-oam-ladjust spring159514.48
pet-omatadjust spring468619.43
adjust spring
set N images
pet-omatadjust spring330511.03
adjust spring
allegro-oam-ladjust spring159314.41

Table 2. Five representative NEB calculations that successfully converged after several retries by the agent. Used MLIP model, attempted retries (from top to bottom), the number of tokens consumed, and communication time with ALCF Inference Endpoint are shown respectively.

AI-driven workflow automation can also validate the robustness of scientific discoveries. Table 3 shows the sensitivity of MLIP prediction by adjusting a convergence criterion; the maximum force between 0.1 - 0.3 eV/Å. While the mean is an important indicator to

MLIPmeanminmaxstdev
allegro-mp-l3.34253.30643.51440.0615
allegro-oam-l3.04532.8123.13030.1226
dpa22.28422.17332.57630.1145
dpa32.55552.51282.59820.0438
equiformer-v32.1532.15232.15430.0008
mace-mp2.09361.94542.1460.0529
mace-omat1.79491.76762.07260.0667
matris-oam2.98472.87983.37520.1337
nequip-mp-l1.4761.36042.08540.1602
nequip-oam-l2.29942.21582.61370.0895
nequip-oam-m2.42652.42442.4310.0018
pet-omat2.07132.0472.20210.0535

Table 3. Sensitivity of activation energy with NEB parameters

select a suitable surrogate from a dozen of very powerful pretrained MLIP models currently available, highly accurate model prediction does not guarantee its robustness [51]. The distribution of the model predictions helps researchers to systematically select a set of MLIP models suitable for microkinetics studies.

Conclusion

We have developed a portable and scalable AI skill that automates NEB calculations by combining agentic reasoning, HPC, and universal MLIPs. Beyond reducing manual intervention, our study demonstrates how domain expertise can be encapsulated into reusable AI skills that autonomously recover from failures, benchmark competing surrogate models, and quantify the robustness of scientific predictions. AI automation also facilitates uncertainty quantification; a critical step for reliable and reproducible science. As scientific foundation models and HPC resources continue to grow, collections of interoperable AI skills executed across distributed computing facilities could form the basis of autonomous scientific campaigns capable of accelerating materials discovery in the era of exascale computing and beyond.

Acknowledgments

The authors acknowledge support from the ARO cooperative agreement W911NF-25-2-0183. This research used computing resources under Innovative and Novel Computational Impact on Theory and Experiment (INCITE) at the Argonne Leadership Computing Facility. This research also used resources of the National Energy Research Scientific Computing Center (NERSC), a Department of Energy User Facility using NERSC award BES-ERCAP 0035639. This work used computational resources of Hokusai provided by RIKEN through the HPCI System Research Project (Project ID: hp260089).

References

  1. [1]NobelPrize.org. 2026. The Nobel Prize in Chemistry 2013. Retrieved Jun 25, 2026 from https://www.nobelprize.org/prizes/chemistry/2013/summary.
  2. [2]Taufiq Mohammed Razakh et al. 2025. Multiscale light-matter dynamics in quantum materials: from electrons to topological superlattices. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis (SC ’25). Association for Computing Machinery, 36–47. doi:10.1145/3712285.3771785.DOI
  3. [3]Shinmosuke Hattori, Kohei Shimamura, Aiichiro Nakano, Rajiv K. Kalia, Priya Vashishta, and Ken-ichi Nomura. 2025. Beyond scaling: chemical intuition as emergent ability of universal machine learning interatomic potentials. (June 2025). arXiv: 2506.07579 [cond-mat.mtrl-sci]. https://arxiv.org/abs/2506.07579.
  4. [4]Jian Zhao et al. 2026. High-temperature memristors enabled by interfacial engineering. Science, 392, 6799, 771–779. eprint: https://www.science.org/doi/pdf/10.1126/science.aeb9934. doi:10.1126/science.aeb9934.DOI
  5. [5]Colonel Vijay Kumar and Balasubramanian Kandasubramanian. 2019. Ad-carbon: new insights into carbon based materials: a review. Industrial & Engineering Chemistry Research, 58, 51, 22663–22701. eprint: https://doi.org/10.1021/acs.iecr.9b04625. doi:10.1021/acs.iecr.9b04625.
  6. [6]Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schütze. 2008. Introduction to Information Retrieval. Cambridge University Press, New York. ISBN: 0521865719.
  7. [7]Yikang Han, Ai Liu, and Haihao Wang. 2023. A survey of vector databases in the era of large language models. arXiv preprint arXiv:2310.11703.
  8. [8]Ziqi Wang, Hongshuo Huang, Hancheng Zhao, Changwen Xu, Shang Zhu, Jian Janssen, and Venkatasubramaniam Viswanathan. 2025. Dreams: density functional theory based research engine for agentic materials simulation. arXiv:2507.14267.arxiv.org/abs/2507.14267
  9. [9]Shanghai Gao et al. 2025. Democratizing ai scientists using tooluniverse. arXiv:2509.23426.
  10. [10]Daniil A. Boiko, Robert MacKnight, Ben Kline, and Gabe Gomes. 2023. Autonomous chemical research with large language models. Nature, 624, 7992, (Dec. 2023), 570–578. doi:10.1038/s41586-023-06792-0.DOI
  11. [11]Andres M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari, Andrew D. White, and Philippe Schwaller. 2024. Augmenting large language models with chemistry tools. Nat. Mach. Intell., 6, 5 (May 2024), 525–535. doi:10.1038/s42256-024-00832-8.DOI
  12. [12]Anthropic and Agentic AI Foundation. 2024. Model context protocol (mcp): an open standard for integrating ai models with external tools and data. Retrieved June 18, 2026 from https://modelcontextprotocol.io.
  13. [13]Agent Skills Working Group. 2026. Agent skills specification: portable ai agent expertise via skill.md. Retrieved June 18, 2026 from https://agentskills.io.
  14. [14][SW] Vercel Labs and The Agent Skills Directory Contributors, The Agent Skills Directory and CLI (skills.sh) 2026. URL: https://www.skills.sh.
  15. [15][SW] Peter Steinberger and The OpenClaw Foundation, OpenClaw: An Open-Source Autonomous AI Agent Framework 2026. URL: https://github.com/OpenClaw/openclaw.
  16. [16][SW] NVIDIA Corporation, NVIDIA NemoClaw: Enterprise-Grade Agent Orchestration and Security Stack 2026. URL: https://blogs.nvidia.com/blog/industrial-software-leaders-secure-autonomous-ai-engineers-nemoclaw/.
  17. [17]Edward O. Pyzer-Knapp, Jed W. Pitera, Peter W. J. Staar, Seiji Takeda, Teodoro Laino, Daniel P. Sanders, James Sexton, John R. Smith, and Alessandro Curioni. 2022. Accelerating materials discovery using artificial intelligence, high performance computing and robotics. npj Computational Materials, 8, 1, (Apr. 2022), 84. doi:10.1038/s41524-022-00765-z.DOI
  18. [18]Chris Lu, Cong Lu, Robert Tjarko Lange, Yutaro Yamada, Shengran Hu, Jakob Foerster, David Ha, and Jeff Clune. 2026. Towards end-to-end automation of ai research. Nature, 651, 8107, (Mar. 2026), 914–919. doi:10.1038/s41586-026-10265-5.DOI
  19. [19]William Dawson, Louis Beal, Yoann Curé, Giuseppe Fisicaro, Dorian Rolland, and Luigi Genovese. 2026. Lara: validation-driven agentic supercomputer workflows for atomistic modeling. arXiv:2604.22571.arxiv.org/abs/2604.22571
  20. [20]Robert MacKnight, Daniil A. Boiko, Jose Emilio Regio, Liliana C. Gallegos, Théo A. Neukomm, and Gabe Gomes. 2025. Rethinking chemical research in the age of large language models. Nat. Comput. Sci., 5, 9, (Sept. 2025), 715–726. doi:10.1038/s43588-025-00811-y.
  21. [21][SW] NVIDIA Corporation, NebsKill: Nudged Elastic Band software utilities and analysis tools 2026. URL: https://github.com/KenichiNomura/nebskill.
  22. [22]Hannes Jönsson, Greg Mills, and Karsten W. Jacobsen. 1998. Nudged elastic band method for finding minimum energy paths of transitions. In Classical and Quantum Dynamics in Condensed Phase Simulations. Bruce J. Berne, Giovanni Ciccotti, and David F. Coker, (Eds.) World Scientific, 385–404.DOI
  23. [23]Graeme Henkelman, Blas P. Uberuaga, and Hannes Jönsson. 2000. A climbing image nudged elastic band method for finding saddle points and minimum energy paths. The journal of Chemical Physics, 113, 22, 9901–9904. doi:10.1063/1.1329672.DOI
  24. [24]Kelly Hong, Anton Troynikov, and Jeff Huber. 2025. Context Rot: How Increasing Input Tokens Impacts LLM Performance. Tech. rep. Chroma, (July 2025). https://research.trychroma.com/context-rot.
  25. [25]Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. ReAct: synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR). OpenReview.net. https://openreview.net/forum?id=WE_vluYUL-X.arxiv.org/abs/2210.03629
  26. [26]Hiroshi Takemiya, Yoshio Tanaka, Satoshi Sekiguchi, Shuji Ogata, Rajiv K. Kalia, Aiichiro Nakano, and Priya Vashishta. 2006. Sustainable adaptive grid supercomputing: multiscale simulation of semiconductor processing across the pacific. In SC ’06: Proceedings of the 2006 ACM/IEEE Conference on Supercomputing, 23–23. doi:10.1109/SC.2006.59.DOI
  27. [27]Ralf Drautz. 2019. Atomic cluster expansion for accurate and transferable interatomic potentials. Physical Review B, 99, 1, (Jan. 2019), 014104. doi:10.1103/PhysRevB.99.014104.DOI
  28. [28]Ilyes Batatia, Simon Batzer, Dávid P. Kovács, Albert Musaelian, Gregor N. C. Smith, Ralf Drautz, Christoph Ortner, Boris Kozinsky, and Gábor Csányi. 2025. The design space of E(3)-equivariant atom-centred interatomic potentials. Nature Machine Intelligence, 7, 1, 56–67. doi:10.1038/s42256-024-00956-x.DOI
  29. [29]Janosh Riebesell et al. 2023. Matbench Discovery – a framework to evaluate machine learning crystal stability predictions. arXiv preprint arXiv:2308.14920, (Aug. 2023). arXiv: 2308.14920 [cond-mat.mtrl-sci]. doi:10.48550/arXiv.2308.14920.
  30. [30]Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P. Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E. Smidt, and Boris Kozinsky. 2022. E(3)-Equivariant Graph Neural Networks for Data-Efficient and Accurate Interatomic Potentials. Nature Communications, 13, 2453. doi:10.1038/s41467-022-29939-5.DOI
  31. [31]Ilyes Batatia, Dávid Péter Kovács, Gregor N. C. Simm, Christoph Ortner, and Gábor Csányi. 2022. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields. In Advances in Neural Information Processing Systems. Vol. 35, 11423–11436.
  32. [32]Albert Musaelian, Simon Batzner, Anders Johansson, Lixin Sun, Cameron J. Owen, Mordechai Kornbluth, and Boris Kozinsky. 2023. Learning local equivariant representations for large-scale atomistic dynamics. Nature Communications, 14, 579. doi:10.1038/s41467-023-36329-y.DOI
  33. [33]Fuyuki Shimojo et al. 2019. Qxmd: an open-source program for nonadiabatic quantum molecular dynamics. SoftwareX, 10, 10307. doi:10.1016/j.softx.2019.10307.DOI
  34. [34]Roberto Car and Michele Parrinello. 1985. Unified approach for molecular dynamics and density-functional theory. Physical Review Letters, 55, (Nov. 1985), 2471–2474, 22, (Nov. 1985). doi:10.1103/PhysRevLett.55.2471.DOI
  35. [35]Michael C. Payne, Michael P. Teter, Douglas C. Allan, Tomás A. Arias, and Jón H. Joannopoulos. 1992. Iterative minimization techniques for ab initio total-energy calculations: molecular dynamics and conjugate gradients. Reviews of Modern Physics, 64, (Oct. 1992), 1045–1097, 4, (Oct. 1992). doi:10.1103/RevModPhys.64.1045.
  36. [36]Pierre Hohenberg and Walter Kohn. 1964. Inhomogeneous electron gas. Physical Review, 136, (Nov. 1964), B864–B871, 3B, (Nov. 1964). doi:10.1103/PhysRev.136.B864.DOI
  37. [37]Peter E. Blöchl. 1994. Projector augmented-wave method. Physical Review B, 50, (Dec. 1994), 17953–17979, 24, (Dec. 1994). doi:10.1103/PhysRevB.50.17953.DOI
  38. [38]John P. Perdew, Kieron Burke, and Matthias Ernzerhof. 1996. Generalized gradient approximation made simple. Physical Review Letters, 77, (Oct. 1996), 3865–3868, 18, (Oct. 1996). doi:10.1103/PhysRevLett.77.3865.DOI
  39. [39]Stefan Grimme. 2004. Accurate description of van der waals complexes by density functional theory including empirical corrections. Journal of Computational Chemistry, 25, 12, 1463–1473. doi:10.1002/jcc.20078.DOI
  40. [40]Ilyes Batatia, Phillip Benner, Yuan Chiang, Adrian M. Elena, Dávid Péter Kovács, et al. 2025. A foundation model for atomistic materials chemistry. The Journal of Chemical Physics, 163, 18, 184110.
  41. [41]Duo Zhang, Xinzijian Liu, Xiangyu Zhang, et al. 2024. DPA-2: a large atomic model as a multi-task learner. npj Computational Materials, 10, 293. arXiv:2312.15492. doi:10.1038/s41524-024-01493-2.DOI
  42. [42]Duo Zhang et al. 2025. A graph neural network for the era of large atomistic models. arXiv:2506.01686. doi:10.48550/arXiv.2506.01686.arxiv.org/abs/2506.01686
  43. [43]Yi-Lun Liao, Alexander J. Hoffman, Sabrina C. Shen, Alexandre Duval, Sam Walton Norwood, and Tess Smidt. 2026. EquiformerV3: scaling efficient, expressive, and general SE(3)-equivariant graph attention transformers. (2026). arXiv: 2604.09130. doi:10.48550/arXiv.2604.09130.DOI
  44. [44]Arslan Mazitov, Filippo Bigi, Matthias Kellner, Paolo Pegolo, Davide Tisi, Guillaume Fraux, Sergey Pozdnyakov, Philip Loche, and Michele Cerruti. 2025. PET-MAD, a lightweight universal interatomic potential for advanced materials modeling. (2025). arXiv: 2503.14118. doi:10.48550/arXiv.2503.14118.
  45. [45]Yuanchang Zhou, Siyu Hu, Xiangyu Zhang, Hongyu Wang, Guangming Tan, and Weile Jia. 2026. MatRS: toward reliable and efficient pretrained machine learning interatomic potentials. (2026). arXiv: 2603.02002. doi:10.48550/arXiv.2603.02002.DOI
  46. [46]Tao Sun, Stefano Fabris, and Stefano Baroni. 2011. Surface precursors and reaction mechanisms for the thermal reduction of graphene basal surfaces oxidized by atomic oxygen. The Journal of Physical Chemistry C, 115, 11, 4730–4737. eprint: https://doi.org/10.1021/jp111372k. doi:10.1021/jp111372k.
  47. [47]Rosanna Laricpirete, Stefano Fabris, Tao Sun, Paolo Lacovig, Alessandro Baraldi, and Silvano Lizzit. 2011. Dual path mechanism in the thermal reduction of graphene oxide. Journal of the American Chemical Society, 133, 43, 17315–17321. eprint: https://doi.org/10.1021/ja205168x. doi:10.1021/ja205168x.
  48. [48]Gemma Team et al. 2024. Gemma: open models based on gemini research. https://arxiv.org/abs/2403.08295 arXiv: 2403.08295.
  49. [49]Abhimanyu Dubey et al. 2024. The llama 3 herd of models. https://arxiv.org/abs/2407.21783 arXiv: 2407.21783 [cs.AI].
  50. [50]OpenAI. 2025. Gpt-oss: open-weight language models with advanced reasoning and tool use capabilities. https://github.com/openai/gpt-oss. (2025).
  51. [51]Xiang Fu, Zhenghao Wu, Wujie Wang, Tian Xie, Sinan Keten, Rafael Gomez-Bombarelli, and Tommi Jaakkola. 2023. Forces are not enough: benchmark and critical evaluation for machine learning force fields with molecular simulations. Transactions on Machine Learning Research. arXiv: 2210.07237 [physics.comp-ph]. doi:10.48550/arXiv.2210.07237.DOI

Paper details

vlnc:W_q8fkrtxs
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 8

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents