OpenLAM: Open-Source Atom Models

• OpenLAM is a community-driven initiative that builds Large Atom Models to generalize across the full periodic table using scalable pretraining and vast datasets.
• It integrates benchmark infrastructure, open databases, and workflow tools to enable transparent model comparison across diverse chemical and structural domains.
• The framework leverages multi-task learning with dataset-specific encodings to address heterogeneous DFT data and bridge the gap to a universal potential energy surface.

Searching arXiv for OpenLAM and closely related Large Atomic Model benchmark papers to ground the article in recent literature. OpenLAM is a community-driven initiative to build, evaluate, and deploy Large Atom Models (LAMs) for materials and molecular science. It is explicitly inspired by the trajectory of LLMs: broad architectures that benefit from massive, diverse data and advanced pretraining to achieve generalizable performance. The initiative aims to develop an open-source foundation model “spanning the entire periodic table,” together with comprehensive, consensus benchmarks and large-scale, high-quality training and evaluation data (Peng et al., 20 Jan 2025). Within this framework, OpenLAM encompasses benchmark infrastructure, open databases of validated crystal structures, multi-domain pretraining collections, and workflow tooling for transparent model comparison. Subsequent work positions LAMBench as the benchmark and workflow engine for universal LAM evaluation, and describes OpenLAM-v1 as a multi-task collection of 31 independently constructed datasets used to train zero-shot generalizable LAMs (Peng et al., 28 Apr 2025, Zhang et al., 2 Jun 2025).

1. Origins, motivation, and scope

OpenLAM emerged from the view that the progress of LLMs was enabled by three coupled ingredients: transformer-like architectural advances, vast data coverage, and scalable pretraining. OpenLAM transfers this logic to atomistic modeling by arguing that LAMs require an open foundation model, transparent practices, and a community standard for benchmarking across extensive configurational and chemical space (Peng et al., 20 Jan 2025). Its target is not a narrow task-specific interatomic potential, but a general LAM intended to generalize across elements, structures, temperatures, and pressures.

The initiative is associated with the Deep Potential team, which had “been actively pretraining Large Atom Models (LAMs) since 2022” and launched OpenLAM to consolidate open-source models, training data, and benchmarks (Peng et al., 20 Jan 2025). The same line of work situates LAMs as foundation-scale neural potentials trained to approximate the universal potential energy surface from first-principles data across chemistry and materials (Peng et al., 28 Apr 2025). In this formulation, the universal PES is defined by

$E = E(R, Z),$

with forces derived as

$F_i = -\nabla_{R_i} E(R, Z),$

and, under periodic boundary conditions, stress related to the virial tensor by

$\sigma = -W/V$

(Peng et al., 28 Apr 2025).

OpenLAM’s stated scope is broad chemical and structural coverage. It targets the full periodic table and a wide range of crystal structures and materials classes, with the explicit aim of establishing benchmarks that reflect the “full capacity” of LAMs rather than performance on isolated tasks (Peng et al., 20 Jan 2025). This broad remit also appears in later descriptions of OpenLAM as a platform for building, sharing, and universally evaluating LAMs across domains, tasks, and simulation regimes (Peng et al., 28 Apr 2025).

2. Conceptual foundations: LAMs, universality, and physical constraints

In the OpenLAM literature, LAMs are large, pretrained atomistic models designed to approximate the ground-state Born–Oppenheimer PES across heterogeneous domains (Peng et al., 20 Jan 2025, Peng et al., 28 Apr 2025). Their intended universality is technical rather than metaphorical: the same Schrödinger equation, under the Born–Oppenheimer approximation, defines a universal PES for atomic systems, so a sufficiently general model could in principle serve “out-of-the-box” across distinct scientific settings (Peng et al., 28 Apr 2025).

This universality is constrained by physical and numerical requirements. The benchmark definition used within OpenLAM emphasizes that energy predictions should be invariant to global translation, rotation, and permutation of identical atoms; forces should be $E(3)$-equivariant; and models should be conservative and differentiable so that second derivatives exist for property calculations such as phonons (Peng et al., 28 Apr 2025). These requirements are not merely formal. LAMBench reports that non-conservative models can appear accurate on static test sets but fail in molecular dynamics or in workflows that demand smooth conservative force fields (Peng et al., 28 Apr 2025).

A recurrent theme in the OpenLAM papers is that heterogeneous DFT data complicates universal modeling. Exchange–correlation functionals, pseudopotentials, basis sets, and software implementations differ substantially across domains, impeding naive dataset merging and making cross-domain evaluation essential (Peng et al., 28 Apr 2025). OpenLAM-v1 addresses this by using multi-task learning with dataset encoding and dataset-specific energy biases, thereby separating shared structural representation from dataset-specific fidelity conventions (Zhang et al., 2 Jun 2025). This suggests that OpenLAM treats universality not as homogenization of all data into a single fidelity, but as a representation-learning problem under controlled multi-fidelity heterogeneity.

3. Data resources: crystal collections and OpenLAM-v1

A central OpenLAM objective is to build large, open databases of validated structures for both pretraining and benchmarking (Peng et al., 20 Jan 2025). The first major data effort was the LAM Crystal Philately competition, whose purpose was to collect stable crystal structures that expand the configurational and chemical space available to LAMs. In its first round, the competition collected “over 19.8 million valid structures,” including “1 million on the OpenLAM convex hull” (Peng et al., 20 Jan 2025). Of these, “12.8 million are uploaded by the participants, while the rest were sourced from open repositories,” including the Materials Project, the Crystallography Open Database, and the American Mineralogist Crystal Structure Database (Peng et al., 20 Jan 2025). The paper further states that “approximate 1.3 million participants-uploaded structures potentially expand the frontier of stable materials beyond what is currently documented in open-source structure repositories” (Peng et al., 20 Jan 2025).

The OpenLAM convex hull is defined through MLIP-predicted energetics. The initiative computes energy above the hull, $E_{\mathrm{hull}}$, using an MLIP-based framework, and structures on the hull are interpreted as thermodynamically stable with respect to decomposition into competing phases, whereas structures above the hull are interpreted as metastable or unstable (Peng et al., 20 Jan 2025). Reliability is supported by validation against the Materials Project on known structures, with “MAE: 0.0096 eV/atom” and “R2: 0.9553” for the correlation between OpenLAM $E_{\mathrm{hull}}$ and MP $E_{\mathrm{hull}}$ (Peng et al., 20 Jan 2025).

A later and distinct resource is OpenLAM-v1, described as a multi-task collection of 31 independently constructed datasets used to train LAMs for zero-shot cross-domain generalization (Zhang et al., 2 Jun 2025). Its scope explicitly includes organic materials, organic and bio-molecules such as SPICE2, catalysis such as OC20, chemical reactions, and inorganic materials such as OMat24 and Materials Project trajectories (Zhang et al., 2 Jun 2025). The paper does not report a single aggregate size for OpenLAM-v1, but it provides detailed statistics for MPtrj, one of its widely used components: about 1.58 million atomic configurations, energies and forces (49.3 million), and stress tensors (14.22 million), spanning approximately 146,000 inorganic materials across 89 elements (Zhang et al., 2 Jun 2025).

The distinction between the crystal-collection program and the multi-domain pretraining collection is important. The former emphasizes validated stable crystal discovery and hull-based ranking; the latter emphasizes multi-task pretraining across heterogeneous DFT conventions. A plausible implication is that OpenLAM combines discovery-oriented structure aggregation with representation-oriented pretraining corpora rather than relying on a single monolithic dataset.

4. Evaluation infrastructure: LAM Crystal Philately and LAMBench

The Crystal Philately competition provides one of OpenLAM’s earliest operational workflows. Its pipeline begins with CIF files encoding chemical species, atomic positions, lattice and space-group information, and periodic boundary conditions. In the pre-submission phase, CIFs are parsed and subjected to full geometry relaxation using an MLIP. In the sampling phase, a randomly sampled subset is checked, and submissions must achieve “valid rate $M/N > 0.3$” to proceed. In the evaluation phase, all structures are checked, each valid structure is scored by energy above the hull, and accepted structures are added to the open-source database (Peng et al., 20 Jan 2025). Validity checks shown in the schematic include “max(abs(force)) < 0.05 eV/Å,” formation-energy screening with “$E_{\mathrm{form}} < 0$,” deduplication and structure matching, and additional filters such as energy change bounds after relaxation (Peng et al., 20 Jan 2025).

The competition also functions as a generative-model testbed. It explicitly allows any generative algorithm or database, and participants applied methods such as ConCDVAE and InvDesFlow. The paper reports that the leaderboard evolved as scoring criteria shifted between formation energy and energy above hull, revealing “an intriguing interplay between participant strategies and the competition rules” (Peng et al., 20 Jan 2025). This indicates that OpenLAM uses benchmark design not only to rank models, but also to probe how generative strategies respond to altered physical objectives.

LAMBench generalizes this benchmarking logic from crystal generation to LAM evaluation across domains. It is described as the benchmark and workflow engine that operationalizes universal evaluation of LAMs across domains, tasks, and simulation regimes, with an interactive leaderboard aggregating results in real time (Peng et al., 28 Apr 2025). LAMBench organizes evaluation around three attributes: generalizability, adaptability, and applicability (Peng et al., 28 Apr 2025).

Generalizability is defined as out-of-distribution performance on downstream datasets. It includes force-field tasks across 17 datasets spanning five domains and property-calculation tasks such as phonons and torsion profiles (Peng et al., 28 Apr 2025). Adaptability measures how well a pretrained LAM can be fine-tuned for direct property prediction. Applicability is assessed through efficiency and stability in realistic simulation settings (Peng et al., 28 Apr 2025). The toolkit is modular, with separate “Models,” “Tasks,” and “Workflow” modules, and it accepts ASE-compatible calculators, physics workflows such as phonon or MD tasks, and Dflow-based job orchestration on cloud or HPC resources (Peng et al., 28 Apr 2025).

5. Metrics, benchmark results, and the universality gap

LAMBench introduces a dimensionless generalizability metric that normalizes model errors against a dummy baseline and aggregates them across domains and prediction types (Peng et al., 28 Apr 2025). For model $m$, domain $F_i = -\nabla_{R_i} E(R, Z),$0, prediction type $F_i = -\nabla_{R_i} E(R, Z),$1, and dataset $F_i = -\nabla_{R_i} E(R, Z),$2,

$F_i = -\nabla_{R_i} E(R, Z),$3

with log-averaging and weighted aggregation producing an overall generalizability score $F_i = -\nabla_{R_i} E(R, Z),$4, where a perfect model scores 0 and the dummy baseline scores 1 (Peng et al., 28 Apr 2025). Applicability is quantified by an efficiency metric

$F_i = -\nabla_{R_i} E(R, Z),$5

with higher values better, and an instability metric based on NVE energy-drift slopes, with lower values better (Peng et al., 28 Apr 2025).

The first LAMBench study benchmarked eight state-of-the-art LAMs released prior to April 1, 2025. For force-field tasks, the reported overall dimensionless errors $F_i = -\nabla_{R_i} E(R, Z),$6 were: DPA-2.4-7M, 0.265; GRACE-2L-OAM, 0.340; SevenNet-l3i5, 0.355; MACE-MPA-0, 0.356; Orb-v2, 0.356; SevenNet-MF-ompa, 0.358; MatterSim-v1-5M, 0.389; and MACE-MP-0, 0.405 (Peng et al., 28 Apr 2025). For property-calculation tasks, the overall dimensionless errors $F_i = -\nabla_{R_i} E(R, Z),$7 were: DPA-2.4-7M, 0.208; SevenNet-l3i5, 0.240; GRACE-2L-OAM, 0.262; MatterSim-v1-5M, 0.280; MACE-MPA-0, 0.291; SevenNet-MF-ompa, 0.300; MACE-MP-0, 0.341; and Orb-v2, 0.560 (Peng et al., 28 Apr 2025).

These results are accompanied by qualitative findings. Most models perform relatively well in Inorganic Materials, Reactions, and Small Molecules, but struggle in Catalysis and Biomolecules; DPA-2.4-7M has the lowest errors in Catalysis and Biomolecules as well as strong performance elsewhere (Peng et al., 28 Apr 2025). Conservative models perform substantially better than Orb-v2 on phonon tasks because phonons require accurate second derivatives, emphasizing conservativeness and smoothness (Peng et al., 28 Apr 2025). For torsional energy profiles in TorsionNet500, multitask DPA-2.4-7M was the best among the LAMs evaluated, but domain-specific models such as MACE-OFF23 remained much better, which the authors interpret as evidence of a gap to the ideal universal PES (Peng et al., 28 Apr 2025).

This conclusion is sharpened in later work using DPA3. When trained on OpenLAM-v1, DPA-3.1-3M is reported to achieve the lowest overall zero-shot generalization error across 17 downstream tasks in LAMBench, with further improvement in a “bestXC” setting that selects the dataset encoding most compatible with the downstream XC functional and, in one case, applies a D3(BJ) dispersion correction (Zhang et al., 2 Jun 2025). Orb-v3 is described as competitive and best in some domains, but DPA-3.1-3M attains the lowest aggregate error across all 17 tasks (Zhang et al., 2 Jun 2025). The same paper does not provide per-task numerical error values, but it frames the result as state-of-the-art aggregate zero-shot performance within the benchmark (Zhang et al., 2 Jun 2025).

6. Architectures, pretraining strategies, and scaling

OpenLAM is architecturally plural. The initial challenge paper cites Deep Potential LAM efforts such as an “attention-based deep learning potential model for molecular simulation” and “DPA-2: a large atomic model as a multi-task learner,” while also noting compatibility with graph neural networks and related atomistic ML designs (Peng et al., 20 Jan 2025). LAMBench benchmarks a heterogeneous set of conservative and non-conservative models, including DeePMD-family descriptors, $F_i = -\nabla_{R_i} E(R, Z),$8-equivariant GNNs, message-passing GNNs, MatterSim, GRACE, and direct-force architectures such as Orb-v2 (Peng et al., 28 Apr 2025).

DPA3 provides the most detailed architectural account in the OpenLAM literature. It is a multi-layer graph neural network founded on line graph series (LiGS), where the first graph $F_i = -\nabla_{R_i} E(R, Z),$9 represents atoms and neighbor pairs, the line graph transform produces $\sigma = -W/V$0 with bonds as vertices and angles as edges, and higher orders can encode dihedrals (Zhang et al., 2 Jun 2025). The paper states that $\sigma = -W/V$1 provides the best accuracy, whereas $\sigma = -W/V$2 degraded force accuracy (Zhang et al., 2 Jun 2025). DPA3 uses residual message-passing updates on each LiGS graph, smooth distance-based switch weights, and a symmetrization block that preserves rotational and permutational invariance while injecting directional information (Zhang et al., 2 Jun 2025). Because forces and virials are obtained by differentiating the predicted scalar energy, the model is conservative by construction (Zhang et al., 2 Jun 2025).

A defining training mechanism is dataset encoding. For training dataset $\sigma = -W/V$3, atomic energy is written as

$\sigma = -W/V$4

where $\sigma = -W/V$5 is a one-hot dataset encoding and $\sigma = -W/V$6 is a dataset-specific chemical-formula energy bias fitted by least squares (Zhang et al., 2 Jun 2025). Total energy, forces, and virials are then

$\sigma = -W/V$7

(Zhang et al., 2 Jun 2025). This mechanism is explicitly intended to accommodate inconsistent XC functionals, basis sets, and software implementations across constituent datasets.

The DPA3 paper also presents a scaling-law analysis. On a down-sampled WBM set with training on MPtrj, test MAE is fit by

$\sigma = -W/V$8

with coefficients $\sigma = -W/V$9, $E(3)$0; $E(3)$1, $E(3)$2; $E(3)$3, $E(3)$4; and $E(3)$5 (Zhang et al., 2 Jun 2025). The authors interpret this as evidence that increasing model size, effective data scale, and compute systematically reduces error in this setting (Zhang et al., 2 Jun 2025). This directly links OpenLAM to the scaling-law paradigm that motivated the initiative from the outset.

7. Access, practical use, limitations, and future directions

OpenLAM includes both user-facing and programmatic access mechanisms. For structure data, the Crystal Craft APP allows queries by Wyckoff positions, chemical formulas, sources, or elements, and the repository provides APIs together with the OpenLAM homepage at AIS Square and competition hosting at Bohrium (Peng et al., 20 Jan 2025). For benchmarking, LAMBench code is open-sourced at GitHub under an MIT License, and an interactive leaderboard is available through the OpenLAM benchmark tab (Peng et al., 28 Apr 2025). OpenLAM-v1 is available on AIS Square through the datasets portal, although licensing terms are not specified in the DPA3 paper (Zhang et al., 2 Jun 2025).

The initiative presents several concrete practical implications. For materials discovery, conservative and stable LAMs with strong cross-domain generalizability can accelerate structure relaxation, phonon calculations, elastic-property workflows, and screening for stability (Peng et al., 28 Apr 2025). For catalysis, improved cross-domain models support adsorption, reaction energetics, and surface dynamics with reduced DFT cost. For molecular and biomolecular modeling, the literature emphasizes torsion profiles, conformational energetics, and MD stability as critical stress tests where current LAMs are improving but still trail domain-specialized models (Peng et al., 28 Apr 2025). The OpenLAM structure database also reduces the need to “create structures from scratch” by enabling researchers to filter candidates by crystalline attributes and stability metrics for downstream screening (Peng et al., 20 Jan 2025).

The published limitations are explicit. Existing benchmarks are described as fragmented, task-specific, or restricted in chemical space, motivating OpenLAM’s consensus-benchmark objective (Peng et al., 20 Jan 2025, Peng et al., 28 Apr 2025). LAMBench identifies a “significant gap between the current LAMs and the ideal universal potential energy surface,” especially in Catalysis and Biomolecules and in comparison with domain-specialized small-molecule models (Peng et al., 28 Apr 2025). Multi-fidelity inconsistency remains a core challenge: matching downstream XC fidelity can materially reduce errors, as shown for DPA-2.4-7M, where choosing better-matched task heads reduced domain errors in Small Molecules from 0.19 to 0.09 and in Biomolecules from 0.25 to 0.17 (Peng et al., 28 Apr 2025). DPA3 further notes that invariant-feature LiGS models can be outperformed by equivariant models such as MACE on some small-molecule AIMD datasets, that larger models can overfit in low-diversity settings, and that memory limits constrained the maximum depth explored in some runs (Zhang et al., 2 Jun 2025).

Future directions are framed as open invitations. The challenge paper explicitly invites collaboration on comprehensive benchmark construction, condition-generation based on experimental properties, and large-scale LAM pretraining (Peng et al., 20 Jan 2025). LAMBench proposes to broaden out-of-distribution datasets, add more property workflows in reactions, catalysis, and biomolecular properties, integrate multi-fidelity modeling more systematically, and develop a unified fine-tuning framework across LAM backends (Peng et al., 28 Apr 2025). The first challenge paper noted that the second competition round was scheduled to begin in February 2025 (Peng et al., 20 Jan 2025). Taken together, these plans define OpenLAM not as a single model release, but as a continuing infrastructure project for open, multi-domain atomistic foundation modeling.