PaperlayerView exact revision on arXiv

Source-linked AI summary

A strategic roadmap for an atomistic machine-learning ecosystem

Jörg Behler, Michele Ceriotti, Cecilia Clementi, Gábor Csányi, Alin-Marin Elena, Aditi Krishnapriyan, Joseph W. Abbott, Fabio Affinito, Albert P. Bartók, Ilyes Batatia, Filippo Bigi, Florian N. Brünig, Yannick Calvino Alonso, Giuseppe Carleo, Aurélie Champagne, Stefan Chmiela, Marc L. Descoteaux, Ralf Drautz, Alexandra Farcas, Meng Gao, Rohit Goswami, Michael F. Herbst, Christian Holm, James R. Kermode, Alexander L. M. Knoll, Tobias Kreiman, Hoang-Thien Luu, Yury Lysogorskiy, Mihai-Cosmin Marinica, Rocco Meli, Klaus-Robert Müller, Frank Noé, Mohamadhosein Nosratjoo, Simon Olsson, Christoph Ortner, Aldo S. Pasos-Trejo, Anyang Peng, Eric Qu, Andrea Rizzi, Mariana Rossi, Bassem Sboui, Gregor N. C. Simm, Alexandre Tkatchenko, Jacopo Venturin, O. Anatole von Lilienfeld, William C. Witt, Brandon M. Wood, Tigany Zarrouk, Fabian Zills

arXiv:2609.39090v1physics.chem-phcond-mat.mtrl-scics.LG

TL;DR

Atomistic models can perform well on fixed-test-set errors yet fail to reproduce the thermodynamic or dynamical properties a study needs, and established practice for validated uncertainty estimates is missing. This roadmap surveys model and simulation approaches alongside ecosystem recommendations. Broadly transferable potentials can already provide reasonable first structural models with little or no in-house training, while transfer across conditions and to non-protein systems remains an active objective.

Abstract

from arXiv · show

Data-driven machine learning (ML) techniques have become an essential tool in many domains of science. Their application to atomistic simulations of matter is particularly widespread and impactful. This success is due largely to the existence of a well-developed and established physics-based modeling framework, ranging from first-principles electronic-structure calculations to molecular dynamics and statistical sampling, into which ML was integrated naturally to reshape long-standing trade-offs between accuracy, efficiency, and scale. Nevertheless, this integration raises both conceptual and practical challenges, from choosing between data-centric and physics-based modeling approaches to adapting established software stacks to modern hardware accelerators and ML libraries. As the field evolves rapidly, fueled in part by widespread enthusiasm but also by tangible impact, it seems appropriate to take a moment to consider the current state of the art and open challenges, and reflect on what can be done to better coordinate efforts across the community. With this goal in mind, several members of this community met in Lausanne in January 2026 at CECAM to discuss algorithms, models, software and hardware infrastructure, and the most promising scientific applications that have become possible thanks to the use of artificial intelligence in atomic-scale simulations. This strategic roadmap paper summarizes the outcomes of these discussions, suggesting some long-term goals, and some concrete actions, to establish a healthy, sustainable and impactful atomistic ML ecosystem.

I. WHAT IS ATOMISTIC MACHINE LEARNING

Atomistic ML builds on mature physics-based simulation methods to address the cost-limited size and timescales of accurate first-principles simulations. The roadmap defines its scope and maps a rapidly evolving field’s open problems and coordination needs.

I. WHAT IS ATOMISTIC MACHINE LEARNING

  • Accurate first-principles costs limit simulations to smaller systems and shorter timescales than many consequential questions require, while ML has emerged as a successful way to ease that constraint.The roadmap focuses on particle-based simulations predicting matter’s structure and properties from fundamental physical laws, without experimental input.Source: §I · p. 2
  • Atomistic ML combines a well-understood physical source of truth with routinely generated electronic-structure data, but dataset sharing is hindered by incompatible reference settings and extrapolation remains central.Differences in pseudopotentials, functionals, basis sets, or convergence settings can make datasets numerically incompatible.Source: §I · p. 2
  • System-specific reference datasets train ML interatomic potentials (MLIPs) that scale simulations in size and time, but usefulness depends on reliable predictions beyond training geometries and chemistries.The roadmap uses MLIP rather than the interchangeable term machine-learning force field (MLFF).Source: §I · pp. 2–3
  • A fast-growing, uneven landscape of methods, datasets, and software has made coordination timely; the roadmap surveys use cases, models, assessment, infrastructure, and community actions.The roadmap follows discussions among developers and users at a January 2026 CECAM workshop.Source: §I · pp. 2–3

II. USE CASES FOR ATOMISTIC MACHINE LEARNING

  • Interatomic-potential parameterization is atomistic ML’s most mature application and the roadmap’s primary focus, although adjacent problems are also developing rapidly.The broader landscape includes atomic properties and electronic-structure surrogates.Source: §II · p. 3

A. Machine learning interatomic potentials and force fields

The section traces ML interatomic potentials from early local representations to specialized and broadly transferable models, then describes their simulation uses and extensions to other properties. It emphasizes that broader transferability does not eliminate the role of focused models, and that electronic-structure surrogates bring distinct opportunities and challenges.

A. Machine learning interatomic potentials and force fields

  • Atom-centered local geometry representations enabled potentials for very large systems, including early high-dimensional neural-network potentials and Gaussian Approximation Potentials.Earlier transferable approaches were constrained to a few degrees of freedom; subsequent generations expanded applications and model complexity.Source: A. Machine learning interatomic… · p. 3
  • The field now spans focused models trained on thousands of structures for high fidelity and general-purpose potentials trained on diverse data to cover much of the periodic table.General-purpose models are often fine-tuned, so “universal” remains aspirational; broadly transferable models can nevertheless provide reasonable first structural models with little or no in-house training.Source: A. Machine learning interatomic… · pp. 3–4
  • Specialized and general-purpose models complement one another: broad models can seed fine-tuning, while focused models remain accuracy standards and economical inference targets; further progress also depends on training-data quality.In some sophisticated simulation tasks, a general model’s zero-shot error against its reference electronic-structure model is lower than that reference model’s error against experiments.Source: A. Machine learning interatomic… · p. 4

B. What are MLIPs good for

  • Up to 90%: pairing MLIP-based optimization with ab initio refinement can reduce force evaluations for local-minimum and saddle-point searches while retaining nominal first-principles accuracy.For transition-state searches, the text advises using MLIPs to accelerate rather than wholly replace the search procedure.Source: B. What are MLIPs good for · p. 4

C. Other atomic and electronic properties

  • ML models can predict response properties such as multipoles, polarizabilities, NMR shifts, XPS binding energies, and vibrational intensities, supporting structural validation and inverse inference from experiments.These properties are mature enough to be auxiliary MLIP outputs with little overhead.Source: C. Other atomic and electronic… · p. 4
  • Electronic-structure surrogates predict outputs such as electron densities or Hamiltonians, potentially reducing calculation bottlenecks while preserving access to derived properties, but fixed-basis targets constrain their representation.Their data, software, and validation challenges are greater than those for MLIPs.Source: C. Other atomic and electronic… · p. 4

III. LONGER TIME AND LENGTH SCALES

Coarse-grained models reduce particle counts and permit longer time steps, but matching equilibrium structure alone does not ensure correct dynamics. Recent work shows progress toward transferable biomolecular models while highlighting needs for compatible data, standardized mappings, and benchmarks.

A. Coarse-graining

  • Coarse-graining reduces particle counts and averages out fast motions, enabling longer time steps; the formal target is the free-energy surface matching the mapped atomistic equilibrium distribution.Coarse-grained coordinates can represent retained particles such as backbone and side-chain beads.Source: A. Coarse-graining · p. 5
  • Bottom-up force matching or relative-entropy models can represent many-body coarse-grained free energies, while top-down models tune interactions to experimental or macroscopic observables.Bottom-up models inherit reference-data accuracy limits; top-down models can match targets while misrepresenting the underlying ensemble.Source: A. Coarse-graining · p. 5
  • Hybrid objectives could correct or fine-tune bottom-up models against experimental targets, anchoring them to experiment rather than potentially erroneous reference methods.The roadmap presents this as a principled route beyond the reference-method “quantum ceiling.”Source: A. Coarse-graining · p. 5
  • Protein-scale coarse-grained models preserve important atomistic thermodynamic features while accelerating sampling by several orders of magnitude.Transferable models trained on diverse atomistic data reproduce folded structures, intermediates, disordered ensembles, folding-upon-binding, and mutation-induced stability changes.Source: A. Coarse-graining · p. 5
  • Equilibrium-consistent coarse-grained models generally do not reproduce diffusion constants or transition rates because eliminating fast degrees of freedom removes implicit friction and memory.Learned memory kernels or transition propagators are possible approaches, but a learned propagator gives up an explicit energy function.Source: A. Coarse-graining · p. 6
  • A unified learnable pipeline could connect mapping, potential, sampler, and backmapping, while current efforts only partially integrate these stages.Backmapping reconstructs atomistic detail from coarse-grained coordinates and can be treated as a conditional generative problem.Source: A. Coarse-graining · p. 6

B. Enhanced sampling with ML potentials

Enhanced-sampling methods combine ML potentials with established estimators to address metastable trapping and costly direct trajectories. Generative samplers offer direct ensemble proposals, but Boltzmann generators and emulators face limits in scale and target-density fidelity.

B. Enhanced sampling with ML potentials

  • Direct ML-potential trajectories can remain trapped in metastable basins, making ensemble and path observables costly; established enhanced-sampling estimators provide complementary tools.Relevant methods include umbrella sampling, replica exchange, metadynamics, thermodynamic integration, free-energy perturbation, and nonequilibrium switching.Source: B. Enhanced sampling with ML potentials · p. 6
  • ML potentials combined with enhanced-sampling methods have supported free-energy calculations across solution and catalytic chemistry, condensed-phase chemistry, and materials science.Examples include reactive potential construction, solvation free energies, and formation free energies in metastable basins.Source: B. Enhanced sampling with ML potentials · p. 6
  • Future benchmarks should assess biased-simulation stability, uncertainty under extrapolative sampling, and final thermodynamic or kinetic observables, not only force and energy errors.At coarse resolution, differentiable coarse-grained potentials can also be combined with free-energy perturbation for mutation-based folding free-energy changes.Source: B. Enhanced sampling with ML potentials · p. 6

C. Generative samplers: Boltzmann generators and emulators

  • Boltzmann generators use generative models with reweighting or Monte Carlo acceptance to sample equilibrium distributions, with applications extended across systems and sampling tasks.They have been used to accelerate free-energy estimator convergence and to propose moves in Markov chain Monte Carlo.Source: C. Generative samplers: Boltzmann… · pp. 6–7
  • Boltzmann generators are practically limited to a few thousand degrees of freedom, while emulators generally cannot be reweighted to a prescribed microscopic target density.These constraints leave learning an MLIP energy and efficiently sampling from it as separate tasks, motivating approaches that make energy functions and samplers mutually consistent.Source: C. Generative samplers: Boltzmann… · p. 7

D. Force-free integrators

Learned evolution maps absorb many integration steps into finite-time updates, potentially bypassing small-step restrictions but placing long-time accuracy and stability on the model. At very long steps, generative propagation must also preserve kinetic properties.

D. Force-free integrators

  • Learned evolution maps predict finite-time updates to amortize many integration steps and target longer simulations despite high-frequency modes, shifting integration-error control from solver to model.Long-time accuracy, stability, conservation laws, and transferability remain central open questions.Source: D. Force-free integrators · p. 7
  • Beyond tens of femtoseconds, propagation becomes stochastic because the conditional transition density must be sampled to represent molecular thermodynamics and kinetics.Generative propagation aims to preserve kinetic properties such as transition rates, time-correlation functions, and kinetic experimental observables while enhancing sampling beyond direct MD.Source: D. Force-free integrators · p. 7

IV. STATE OF THE ART

  • The roadmap organizes model design around incorporated physics, mathematical architecture, and training data, whose coupled choices shape performance.Symmetry-constrained architectures enforce physical symmetry, while data augmentation can also produce highly symmetric predictions.Source: §IV · p. 7

A. Physics

  • Atomistic models must respect constraints such as translation and rotation symmetry, smoothness, and conservative forces, while their physical parameterization can reflect processes such as Coulomb interactions, magnetism, and charge transfer.The physical form can be represented in different ways, including explicit or indirect charge learning and descriptors capturing Coulomb asymptotics.Source: A. Physics · p. 8

B. Model architectures

Atomistic model architectures trade physical inductive bias against capacity, locality, and inference cost. Data growth has expanded model scope, but inconsistent reference settings, fragmented datasets, and mismatched theory levels constrain reliable combination and transfer.

B. Model architectures

  • Atom-centered descriptors and message passing exploit locality, while reciprocal-space methods, all-to-all transformers, linear attention, and virtual nodes provide routes to longer-range information transfer.Large-chemical-space models are often graph neural networks; shallow linear, nonlinear, and kernel approaches remain useful for small data, fast inference, and distillation.Source: B. Model architectures · p. 8
  • Equivariant and high-body-order architectures offer inductive bias and data efficiency, while invariant constructions can also be universal approximators; unconstrained models can recover approximate symmetry and locality from data.Whether emergent constraints behave like hard-coded ones out of distribution remains an open validation question.Source: B. Model architectures · p. 8
  • Direct-force models infer forces independently of energies and enable two- to three-fold faster inference than conservative models, at the cost of energy conservation.Direct forces can also reduce pretraining cost or serve inner steps in multiple-time-stepping schemes, while conservative forces handle outer steps.Source: B. Model architectures · pp. 8–9

C. Data

  • First-principles datasets have grown to hundreds of millions of structures, but differences in electronic-structure settings can make datasets numerically incompatible and even yield qualitatively different forces.Magnetism, spin polarization, and Hubbard +U choices are particularly problematic; DFT accuracy also varies across elements and bonding regimes.Source: C. Data · p. 9
  • A dataset can validate against itself while being systematically wrong, and foundation potentials inherit their training corpus’s level of theory; transfer to another reference requires new-level fine-tuning or multi-fidelity strategies.Multi-fidelity methods require internally consistent sources and a ground-truth anchor, making curated higher-theory reference subsets essential.Source: C. Data · p. 9
  • Neural exchange-correlation functionals have shown benchmark improvements, and Skala improved molecular thermochemistry while retaining Kohn–Sham DFT’s O(N^3) scaling; robust methods could generate higher-accuracy MLIP data at lower cost than CCSD(T).Neural VMC offers high-end reference data for strongly correlated chemistry but remains very expensive, limiting routine applications to a few heavy atoms and requiring effective core potentials for metals.Source: C. Data · pp. 9–10
  • Materials and molecule datasets developed largely independently, and smaller-than-1M datasets can be competitive when representative, highly converged, and internally consistent.Coordinated curation could produce datasets that are diverse, accurate, and larger than individual groups can generate.Source: C. Data · p. 10

D. What is now possible

Current MLIPs support applications from materials discovery to molecular and extreme-condition simulations, while important physical frontiers and end-to-end evaluation challenges remain.

D. What is now possible

  • MLIPs support large-scale materials simulations, high-throughput candidate screening, quantitatively confident water studies, extreme-condition modeling, and advances in single-molecule chemistry.For solid-state electrolytes, MLIP-driven molecular dynamics has reduced ionic-conductivity computation costs by orders of magnitude; molecular and biomolecular systems remain harder to simulate at relevant timescales.Source: D. What is now possible · p. 10
  • Long-range electrostatics and dispersion lack a unified treatment, and the field has no consensus on when explicit long-range modeling matters.Charge equilibration, latent long-range messages, and Ewald-like augmentations coexist, with suitability often depending on the system's physics.Source: D. What is now possible · p. 10
  • Excited states, multiple spin states, non-collinear magnetism, and strong static correlation challenge both reference data and model architectures.Sampling algorithms for quantum nuclear effects have been combined with ML but are not broadly available in end-to-end modeling tools.Source: D. What is now possible · pp. 10–11

V. ASSESSING THE QUALITY OF MODELS AND DATA

  • A model's fixed-test-set energy or force RMSE may not predict physical trajectories or target properties, while mediocre RMSE can still suit a particular question.Derived properties such as free energies and transport coefficients lack affordable electronic-structure reference values, and disagreements with experiment can also reflect dynamics or insufficient sampling.Source: §V · p. 11

A. Beyond test-set error

Evaluating MLIPs requires downstream physical tests and scrutiny of reference labels, while benchmarks should be designed to remain informative rather than become targets.

A. Beyond test-set error

  • Test-set RMSE misses transferability, potential smoothness, and long-run stability; downstream tests should include physical observables and stability checks.Localized features such as reactive sites, defects, transition states, or interfaces can dominate behavior while contributing little to global RMSE.Source: A. Beyond test-set error · p. 11, Figure 3 · p. 11
  • DFT-referenced benchmark errors inherit reference-method errors, so benchmark reports should include variation across reasonable DFT settings and uncertainty estimates.Where higher-level references are available, the authors recommend reporting model error against them as well.Source: A. Beyond test-set error · p. 11

B. Benchmarks: useful, but not as goals in themselves

  • Benchmarks can improve access and prediction accuracy, but leaderboard incentives may favor in-distribution scores over practical out-of-distribution performance and test leakage can inflate accuracy.The authors recommend an explicit benchmark lifecycle, including periodic refresh or retirement.Source: B. Benchmarks: useful, but not as… · p. 12
  • Benchmarks should pair raw-label errors with physically meaningful diagnostics and disclose reference settings, label uncertainty, and reference-method limitations.These design principles make the basis and scope of benchmark targets clearer.Source: B. Benchmarks: useful, but not as… · p. 12
  • Benchmark comparisons should account for simulation protocols and costs across systems, remain open to stakeholders, and use complementary tests instead of one ranking number.Benchmark maintainers should be independent of model developers, computational costs should be assessed consistently, and benchmarks should have an explicit retirement path.Source: B. Benchmarks: useful, but not as… · p. 12

C. Community one-time challenges as a complement to benchmarks

One-time blind challenges complement fixed benchmarks, while system-specific uncertainty estimates and reference-method uncertainty remain central to judging model validity.

C. Community one-time challenges as a complement to benchmarks

  • Blind community challenges complement fixed benchmarks by evaluating predictions on new problems and can use independent experiments for properties without affordable computational references.They are resource-intensive to organize, and new data can still resemble existing training data.Source: C. Community one-time challenges as a… · p. 12

D. Uncertainty and validity

  • End users need validated uncertainty estimates that indicate a model's expected accuracy for their system and are reported in standard forms usable by downstream tools.The paper calls for uncertainty reporting alongside consideration of model misspecification and internal consistency, rather than relying only on hold-out calibration.Source: D. Uncertainty and validity · p. 13
  • Probabilistic DFT models and differentiable codes can propagate parameter distributions to quantities such as optimized structures and forces.The paper also notes methods for estimating plane-wave basis-set errors and argues that such estimates may inform future MLIP training.Source: D. Uncertainty and validity · p. 13
  • An MLIP trained on electronic-structure calculations is bounded in experimental accuracy by its reference method, and higher-level training data are costly to generate routinely.The paper identifies multi-fidelity, delta-learning, and hybrid training against experimental targets as promising routes to bridge this gap.Source: D. Uncertainty and validity · p. 13

VI. A CHALLENGE OF SOFTWARE AND HARDWARE

Atomistic ML depends on a layered, accelerator-oriented software stack, but integrating it with established simulation codes and coordinating reusable components remain difficult.

VI. A CHALLENGE OF SOFTWARE AND HARDWARE

  • Industry-led ML frameworks and accelerators have boosted atomistic ML, but their rapid evolution and some closed low-level components make the ecosystem fragile.Scientific code bases require continuing investment to keep pace with these lower layers.Source: §VI · p. 13
  • Integrating Python-based ML architectures with classical C++ or Fortran molecular-dynamics engines requires accelerator-native representations and neighbor lists within domain-decomposed codes.This is particularly demanding for equivariant message-passing networks that evaluate tensor products on irreducible representations on the fly.Source: §VI · p. 13

A. Anatomy of the stack

  • The stack progresses from hardware and low-level libraries through ML frameworks and atomistic-ML libraries to applications and simulation drivers.Its layers provide progressively higher-level components, from tensor operations and automatic differentiation to simulation and data-labeling workflows.Source: A. Anatomy of the stack · p. 14, Figure 4 · p. 14
  • Fragmented training code, dataset formats, neighbor-list implementations, and simulation bindings repeat effort and impede standardized reuse of low-level operations.Existing reusable libraries have not always led to standardized adoption.Source: A. Anatomy of the stack · p. 14
  • Integrating ML with electronic-structure software is further complicated by large memory demands, varied formalisms, and high parallelism requirements.Autodifferentiable quantum-chemistry codes using similar software frameworks offer promising examples for such integration.Source: A. Anatomy of the stack · p. 14

B. Reasons for monolithic codes, and reasons against

The roadmap weighs the optimization and governance advantages of monolithic codes against the coordination and maintenance costs of modular ecosystems. It recommends shared, language-agnostic interfaces and modular design that accounts for end-to-end performance.

B. Reasons for monolithic codes, and reasons against

  • Monolithic codes offer tight integration, clear governance, and a self-contained artifact, while modular ecosystems bring dependency, coordination, responsibility, and credit-tracking costs.The roadmap argues that fragmentation costs now outweigh the benefits of unrestricted independent development, and that shared interfaces can yield disproportionate gains.Source: B. Reasons for monolithic codes, and… · pp. 14–15
  • Separate key algorithmic primitives, readable reference implementations, and interchangeable performant implementations behind common interfaces.The proposed primitives include sparse operations, equivariant tensor products, long-range routines, neighbor lists, and flexible loss functions.Source: B. Reasons for monolithic codes, and… · p. 15
  • Modern frameworks can blur reference and performance roles, while Julia supports high-level custom algorithms with minimal performance impact but remains weakly integrated with standard simulation ecosystems.The roadmap describes integration as a prerequisite for wider adoption.Source: B. Reasons for monolithic codes, and… · p. 15

C. Interoperability and isomorphic interfaces

  • Formalize a language-agnostic calculator interface and a shared dataset-field ontology rather than mandating one binary format.ASE’s positions–species–cell to energy–forces–stress contract is a de facto interface; proposed dataset options include extended XYZ, HDF5, LMDB, Zarr, and Parquet.Source: C. Interoperability and isomorphic… · p. 15
  • Use optional plugins and stable cross-language bindings, but choose interfaces with end-to-end performance and rigorous testing in view.The roadmap identifies a stable C ABI and model formats such as ONNX or StableHLO as possible routes to interoperability.Source: C. Interoperability and isomorphic… · p. 15

D. Hardware: leverage and lock-in

The roadmap proposes community practices to improve the long-term health of atomistic ML, whose impact and costs of inaction are rising. Its recommendations emphasize coordination without consolidation, shared standards, and sustained stewardship.

D. Hardware: leverage and lock-in

  • Coordinate with vendors and HPC centers around a small set of shared operations, and benchmark end-to-end MD throughput with specified hardware and software settings.The roadmap identifies whole-MD throughput as the most application-relevant metric, while noting that it is especially prone to confounding factors.Source: D. Hardware: leverage and lock-in · p. 16

VII. STRATEGY FOR THE COMMUNITY

  • Preserve diverse architectures and software stacks while formalizing language-agnostic interfaces, dataset ontologies, model-card metadata, and shared low-level primitives.The roadmap calls for collaborative API documentation, governance, and maintenance of existing community tools.Source: §VII · p. 16
  • Improve reference-data fidelity and consistency through metadata standards, multi-fidelity strategies, and a public set of structures recomputed across datasets.The roadmap proposes using resulting fingerprints to compare electronic-structure settings and expose incompatibilities before model training.Source: §VII · p. 16
  • Supplement existing benchmarks with application-driven tests, cost reporting, and recurring blind challenges whose committee maintains versioned community standards.The roadmap proposes a regular challenge series, for example biennially, paired with a CECAM conference and an explicit decision procedure.Source: §VII · pp. 16–17
  • Use controlled model problems with known ground truth to probe theoretical questions about model capabilities and limitations.Examples include a known long-range tail, an isolated defect, and an analytically tractable energy landscape.Source: §VII · p. 17
  • Make calibrated, held-out-data-validated uncertainty estimates a default for production models, exposed through standard interfaces and reported alongside reference-method accuracy.Source: §VII · p. 17
  • Support sustainable shared infrastructure with research-software career paths and visible academic credit for software contributions.The roadmap also calls for established citation conventions for low-level libraries.Source: §VII · p. 17
  • Build partnerships between traditional simulation codes, vendors, and HPC centers, keep vendor implementations behind community-defined interfaces, and maintain the roadmap through versioned updates.The roadmap proposes in-memory interoperability, joint workshops and infrastructure, and tracking which recommendations are delivered or remain open.Source: §VII · p. 17

AUTHOR CONTRIBUTIONS STATEMENT

The statement identifies the organizers and manuscript coordinators, describes the involvement of other authors, and notes the use of AI in drafting.

AUTHOR CONTRIBUTIONS STATEMENT

Related papers

Loading 2609.39090v1…