Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

xChemAgents: Agentic AI for Explainable Quantum Chemistry

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Two frozen language-model agents select and validate three to five weighted descriptors per molecule, cutting QM9 mean absolute error by up to 22% while producing natural-language rationales.

desk verdict The agentic feature-selection wrapper is coherent and the tables are new, but missing ablations leave the core attribution claim—that the Selector–Validator dialogue, not the added text channel, causes the gains—unsupported. read the letter →

arxiv 2505.20574 v2 pith:WLYV5FSM submitted 2025-05-26 cs.MA physics.chem-phphysics.comp-ph

Can Polat , Mehmet Tuncel , Mustafa Kurban , Erchin Serpedin , Hasan Kurban This is my paper · ORCID
classification cs.MAphysics.chem-phphysics.comp-ph
keywords xChemAgentsmulti-agentLLMsystemsmultimodalmolecularrepresentationlearningQM9propertypredictionequivariantgraphneuralnetworksdescriptorselectionphysics-awarefeaturevalidationinterpretableAIforchemistry
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that molecular property prediction improves when descriptor selection is an explicit, physics-checked dialogue between two frozen large language model agents: a Selector that picks three to five weighted textual descriptors and explains its choice, and a Validator that rejects proposals violating unit consistency, scaling laws, or sparsity. On the QM9 benchmark, this cooperative loop lowers mean absolute error by up to 22% relative to the multimodal baselines, with the clearest gains for electronic quantities such as HOMO and LUMO energies, while adding about 0.2 seconds per molecule. The point of the work is that feature selection for surrogate models can be both adaptive and auditable: the accepted descriptors and their rationales are logged, and the text channel shrinks from nine descriptors to at most five. The paper also shows where the approach stops helping—thermodynamic energies such as internal energies often stagnate or degrade, which the authors tie to the limited physical content of high-level textual descriptors.

What carries the argument

The load-bearing mechanism is the Selector–Validator dialogue: $A_{\mathrm{sel}} : (x,y) \mapsto (S,w,r_{\mathrm{sel}})$ outputs a three-to-five element subset, weights on the simplex $\Delta^{p-1}$ summing to one, and a chain-of-thought rationale, while $A_{\mathrm{val}} : (S,w,y) \mapsto (v,r_{\mathrm{val}})$ returns a verdict and critique guided by unit consistency, scaling laws, and sparsity or complementarity. Rejected proposals are fed back to the Selector for up to $L=3$ rounds. The accepted, weighted descriptors are combined into a physics-aware text embedding $t_{\mathrm{phys}} = \sum_j \hat{w}_j \, \phi_{k_j}$ from frozen CLIP embeddings of each descriptor, and this embedding is fused with the graph embedding from an equivariant GNN by a learned gate $f = z \odot \tilde{g} + (1-z) \odot \tilde{t}$. The loop decides which textual information reaches the predictor, reducing the channel from nine descriptors to at most five and leaving an audit trail of rationales.

What would settle it

Replace the Selector with a fixed rule that always picks the same three to five descriptors—say, the most frequently chosen ones reported in the paper—and disable the Validator, keeping the same GNN backbone, fusion module, text embeddings, and training schedule; if the mean absolute errors on HOMO energy, LUMO energy, and HOMO–LUMO gap are statistically indistinguishable from the full system's, then the agent dialogue is not responsible for the improvement.

Watch

Extended reading notes

Core claim

The central claim is that a cooperative Selector–Validator agent loop, with both agents frozen, turns descriptor augmentation from a static preprocessing step into an adaptive, physics-vetted process. For each molecule and target, the Selector returns a subset of three to five descriptors from a bank of nine, a normalized weight vector summing to one, and a natural-language rationale; the Validator checks the proposal against unit consistency, scaling laws such as Koopmans' theorem for HOMO/LUMO, and sparsity and complementarity, and up to three rounds of critique can revise it. The accepted descriptors are embedded with frozen CLIP vectors and combined as $t_{\mathrm{phys}} = \sum_j \hat{w}_j \phi_{k_j}$, which is then fused with an equivariant graph neural network embedding through a learned gate. Across SchNet, DimeNet++, Equiformer, and FAENet, the agentic selection cuts mean absolute error for dipole moment and frontier-orbital targets, with the largest reductions for HOMO and LUMO energies, and the selection statistics show the agents rely on molecular weight as a frequent but low-weight default while assigning high weights to XLogP for LUMO energy and gross formula for the HOMO–LUMO gap. Thermodynamic targets show limited benefit or degradation, which the authors attribute to the weak physical signal in string descriptors for potential-energy-surface properties.

Load-bearing premise

The load-bearing assumption is that the accuracy gains come from the Selector–Validator agent process itself, rather than from merely adding text descriptors or from the particular language model used; the reported experiments compare base models with the full agentic system but do not include a control using all nine descriptors, a random selector, or a fixed-weight selector.

Editorial extensions

If this is right

  • If the central claim holds, any geometry-based surrogate can be augmented with a small, target-specific text channel that improves electronic-property predictions without retraining the language models.
  • Fusion placement matters: late fusion at the global readout gives net gains on five of nine targets, while early fusion into every attention layer of one backbone degrades geometry-dominated properties.
  • Cutting the descriptor channel from nine to at most five, at about 0.2 seconds overhead per molecule, makes the interpretability gain cheap enough to keep in screening workflows.
  • The logged rationales and accepted descriptor sets create an audit trail, addressing the common interpretability deficit of graph neural networks in materials science.
  • The selection pattern—molecular weight chosen often but weighted low, XLogP and gross formula weighted high for specific orbital targets—indicates the agent balances a generic size proxy with task-specific chemistry.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The experiments do not include a control that uses all nine descriptors, a random three-to-five descriptor subset, or fixed weights; if any of those matched the reported errors, the agent dialogue would not be the cause of the gains.
  • Because the agents are frozen and the descriptor bank is replaceable, the same dialogue could be applied to crystals, alloys, or polymers by swapping in a different set of textual descriptors.
  • The authors' suggestion that thermodynamic targets need physics-derived descriptors such as force constants or vibrational modes is directly testable by adding those descriptors to the bank and rerunning the same protocol.
  • An ablation that removes the Validator while keeping the Selector's choices would isolate whether the physics-checking dialogue, not the sparse selection alone, drives accuracy.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper introduces xChemAgents, a multi-agent framework in which a frozen LLM-based Selector proposes a sparse, weighted subset of textual descriptors for a target quantum-chemical property, a Validator critiques the proposal against physical constraints, and the accepted descriptors are fused via CLIP embeddings into a GNN backbone. Experiments on QM9 with four backbones (SchNet, DimeNet++, Equiformer, FAENet) report MAE reductions of up to 22% for frontier-orbital properties, alongside qualitative analyses of the agents' selection behavior. The authors argue that the agentic dialogue improves both accuracy and interpretability relative to geometry-only baselines.

Significance. If the claimed gains were shown to derive from the agentic Selector-Validator process, the paper would offer a novel and potentially useful way to inject domain knowledge into multimodal molecular property prediction, with the added value of explicit, auditable rationales. The release of code and data is a strength that supports follow-up work. However, the current experimental design cannot separate the effect of the agent dialogue from the simple addition of text features, and the baselines are not state-of-the-art, so the significance of the headline result is presently unclear and needs substantial additional evidence.

major comments (5)
  1. [§4.1, Table 1] The central claim that cooperative Selector-Validator agents improve accuracy is not supported by the experimental design. The comparison between Base and xChemAgents changes three factors simultaneously: the addition of CLIP text embeddings, the LLM-driven selection of a 3-5 descriptor subset with weights, and the iterative Validator dialogue. There is no control using all nine descriptors, a random descriptor subset, fixed equal weights, or a Selector-only variant without the Validator. Since the Introduction argues that naively appending large descriptor sets can degrade performance, it is plausible that any sparse subset—or even a simple statistical feature-selection routine—would reproduce the observed MAE changes, making the agent dialogue an irrelevant wrapper. Please add these ablations and report whether the agentic selection is necessary for the improvements.
  2. [§4 (Implementations) and Table 1] The abstract and Section 4 claim 'up to a 22% reduction in mean absolute error over the state-of-the-art baselines,' but the baselines are not state-of-the-art: SchNet uses six interaction blocks, DimeNet++ is 'scaled down by half,' Equiformer uses 50% fewer channels, and FAENet uses two interaction blocks. These are reduced-capacity reimplementations, and the paper provides no comparison with published state-of-the-art errors on QM9. Furthermore, metrics are reported as means over three randomized folds with no standard deviations, error bars, or per-fold values, so the statistical reliability of the differences (e.g., SchNet εH 0.1790 vs 0.1358 eV) cannot be assessed. Please either compare with published SOTA numbers or explicitly characterize the baselines as reduced-capacity, and provide variance estimates.
  3. [§4.1, Table 1, Figure 2, Abstract] The paper's summary claims are selective relative to its own Table 1. While improvements are reported for HOMO, LUMO, and gap for most backbones, Table 1 also shows large degradations: Equiformer U0K increases from 12.3913 to 19.5572 eV, FAENet U298K from 11.0204 to 19.8619 eV, and SchNet α from 0.3154 to 0.4194 ų. Moreover, the text in §4.1 states 'consistent reductions in MAE across all backbones for HOMO energy εH, LUMO energy εL, and the HOMO–LUMO gap Δε,' but Table 1 shows FAENet εL degrades from 0.0980 to 0.1014 eV. This overstates the consistency and the abstract's 'up to 22%' framing omits the severe degradations. Please report all targets comprehensively and temper the claims accordingly.
  4. [§3.1, Algorithm 1, §4 (Implementations)] The agentic system is not reproducible as described: the only reported implementation detail is 'Ollama serving as the backbone model' (Section 4), with no specification of the underlying LLM checkpoint, version, temperature, or prompt templates. Because the Selector and Validator outputs are nondeterministic and model-specific, the reported results cannot be independently reproduced, and the claimed benefit could be an artifact of one particular unstated model rather than of the framework. Please disclose the exact model, inference settings, and prompts, and ideally include a sensitivity analysis over at least two different LLMs.
  5. [§3.1, Eq. (1)] The descriptor weights w in Eq. (1) are generated by a frozen LLM and are not learned or optimized against the prediction target. The paper provides no evidence that these LLM-assigned weights outperform simple alternatives such as equal weights, weights derived from a statistical filter, or a learned attention mechanism over the full descriptor set. Without such a comparison, the claim that the 'physics-aware' weighting contributes to accuracy is unsupported; this is closely tied to the missing ablations in Table 1 and should be addressed in the same set of experiments.
minor comments (6)
  1. [§3.1] The heading 'Physics-aware texual embedding' contains a typo ('texual' should be 'textual').
  2. [Algorithm 1, line 12] Line 12 writes tphys ← Σ_{φ_k∈Ŝ} ŵ_k CLIP(φ_k), but the descriptor embeddings are already computed with a frozen CLIP encoder (Section 3.1); clarify whether CLIP(φ_k) denotes the precomputed embedding and align notation with Eq. (1).
  3. [Figure 2] The y-axis label 'Percentile change in MAE' is unclear: it should specify whether the values are relative percentage changes or percentile ranks, and relative to which baseline.
  4. [§4 (Implementations)] The description of the gated-fusion layer appears both in the main text and in a separate 'Gated-fusion layer' paragraph, with slightly different notation (e.g., W_g vs W); unify the notation to avoid confusion.
  5. [References] Several references are incomplete, e.g., 'arXiv preprint arXiv:2412.11693' for CrystalFlow and 'arXiv preprint arXiv:2503.16874' for MARS; please supply full entries, including titles and author lists where available.
  6. [Table 2] The caption states 'Best values in each row are shown in bold, and second-best are underlined,' but for 'Normalized Importance' rows, higher weights are not necessarily 'better' in a predictive sense unless tied to the accuracy analysis; clarify the criterion.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the descriptor weights are produced by a frozen LLM and fed forward, not fitted to the targets, so no prediction reduces to its input by construction.

full rationale

The claimed derivation chain is empirical rather than analytic. The Selector maps (x, y) to a subset and weights using a frozen LLM; the weighted text embedding tphys = sum of w_j * CLIP(phi_j) is a fixed feedforward combination of frozen CLIP embeddings; the Validator applies LLM-based critique; only the GNN, gating, and MLP parameters are trained on QM9 labels. The paper states 'During training and inference, both agents remain frozen,' so the descriptor weights are not optimized against the target values and no fitted parameter is relabeled as a prediction. No equation in Section 3 defines a reported MAE in terms of the model's own inputs by construction. The self-citations (Polat et al., 2024; 2025a; 2025b) support background premises about multimodal descriptor behavior, but they are not used as derivation steps that force the numerical results; the QM9 labels provide external evidence. The absence of an all-descriptor or random-selector control in Table 1 weakens the causal attribution of the gains to the agent dialogue, and the 'state-of-the-art baselines' wording is inflated, but those are experimental-design and reporting concerns, not circular reductions. Thus no circular step is identified; the minor score reflects the self-citations and the unisolated comparison, not circular reasoning.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The paper is an empirical benchmark, not a derivation, so the ledger captures the design choices and domain assumptions that the accuracy claim depends on. The most important items are the assumed causal role of the agent dialogue, the assumption that CLIP and LLM outputs carry physics that can be validated, and the unquantified exclusion of molecules missing PubChem fields.

free parameters (6)
  • Descriptor subset size p = 3 to 5
    The Selector is prompted to output 3-5 features; this range is a hand-set design choice, not optimized or derived.
  • Dialogue rounds L = 3
    The Selector-Validator loop is capped at three iterations by design (Algorithm 1); no analysis of the effect of this cap is provided.
  • Text embedding projection dimension = 32
    The CLIP text vector is projected to 32 dimensions before fusion; this latent size is chosen by hand.
  • Graph hidden dimensions and interaction blocks = 128 or 256 hidden, 2-6 blocks
    Backbones are configured with reduced capacity relative to published models, including DimeNet++ at half scale and Equiformer with 50% fewer channels, which changes the baseline and the comparison.
  • Training schedule = 35 epochs, batch 64, lr 1e-3
    A shared training schedule is chosen for all models; no learning-rate search or convergence analysis is reported.
  • Randomized folds k = 3
    Performance is averaged over three randomized folds with no seeds and no variance reporting.
assumptions (6)
  • domain assumption QM9 target values from DFT are valid ground truth for the benchmark.
    All MAE comparisons assume the QM9 labels are the reference; Section 4, Dataset.
  • domain assumption Frozen CLIP embeddings of textual descriptors encode chemically relevant information for the target properties.
    The method replaces the text vector with a weighted average of CLIP embeddings (Eq. 1) and assumes these vectors carry physical signal.
  • ad hoc to paper An LLM can judge unit consistency and scaling laws from descriptor names and property descriptions alone.
    The Validator prompt encodes unit consistency and Koopmans' theorem (Section 3.1); no mechanism shows how these constraints can be evaluated on text embeddings.
  • ad hoc to paper The Selector's outputs are causally responsible for prediction improvements, not merely the presence of text metadata.
    Table 1 lacks a full-descriptor or random-selection control, so this attribution is assumed.
  • domain assumption Excluding molecules missing PubChem fields does not bias the benchmark.
    Section 4 states exclusion without quantifying how many molecules or which chemical classes are removed.
  • ad hoc to paper Three randomized folds with no variance reporting are sufficient to support the accuracy claims.
    Section 4 reports mean over k=3 folds only; no error bars or significance tests are given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of xChemAgents: Agentic AI for Explainable Quantum Chemistry." pith.science (2026). https://pith.science/paper/WLYV5FSM

@misc{pith2026250520574,
  author       = {Pith},
  title        = {Pith review of: xChemAgents: Agentic AI for Explainable Quantum Chemistry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WLYV5FSM}},
  note         = {Machine review of arXiv:2505.20574}
}
read the original abstract

Recent progress in multimodal graph neural networks has demonstrated that augmenting atomic XYZ geometries with textual chemical descriptors can enhance predictive accuracy across a range of electronic and thermodynamic properties. However, naively appending large sets of heterogeneous descriptors often degrades performance on tasks sensitive to molecular shape or symmetry, and undermines interpretability. xChemAgents proposes a cooperative agent framework that injects physics-aware reasoning into multimodal property prediction. xChemAgents comprises two language-model-based agents: a Selector, which adaptively identifies a sparse, weighted subset of descriptors relevant to each target, and provides a natural language rationale; and a Validator, which enforces physical constraints such as unit consistency and scaling laws through iterative dialogue. On standard benchmark datasets, xChemAgents achieves up to a 22% reduction in mean absolute error over the state-of-the-art baselines, while producing faithful, human-interpretable explanations. Experiment results highlight the potential of cooperative, self-verifying agents to enhance both accuracy and transparency in foundation-model-driven materials science. The implementation and accompanying dataset are available at https://github.com/KurbanIntelligenceLab/xChemAgents.

Figures

Figures reproduced from arXiv: 2505.20574 by the authors.

Figure 1
Figure 1. Flowchart of the Selector–Validator pipeline for molecu￾lar feature selection. The process begins with an input molecule and its property description; the Selector agent identifies 3–5 can￾didate features, assigns normalized weights, and documents its reasoning. The Validator agent then evaluates feature relevance, weight accuracy, and overall completeness. If validation succeeds, the finalized, validated selections… view at source ↗
Figure 2
Figure 2. Percentile change in MAE across models for each target property. Positive values denote an increase in error relative to the baseline, while negative values indicate an improvement in the model’s performance. FAENet. FAENet is configured with a cutoff radius of 6 A˚ , two interaction blocks, a hidden dimension of 128, and 20 Gaussian distance channels. The textual CLIP embedding (768-dimensional) is reduced to 32 di… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Stress-Testing Multimodal Foundation Models for Crystallographic Reasoning

    cs.CV 2025-06 reject novelty 6.0 of 10

    Across nine vision-language models, performance collapses when chemical composition is held out, but the reported magnitude and internal consistency of this collapse are not supported by the paper's own tables.

  2. EMPATHIA: Multi-Faceted Human-AI Collaboration for Refugee Integration

    cs.AI 2025-08 reject novelty 4.0 of 10

    A multi-agent LLM system produces weighted refugee placement recommendations; its reported 87.4% 'validation convergence' only measures self-consistency, not real-world validity.

Reference graph

Works this paper leans on

41 extracted references · 26 canonical work pages · cited by 2 Pith papers

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    and Xia, T

    Aykent, S. and Xia, T. GotenNet: Rethinking Efficient 3D Equivariant Graph Neural Networks . In The Thirteenth International Conference on LearningRepresentations, 2025. URL https://openreview.net/forum?id=5wxCQDtbMo

    work page 2025

  3. [3]

    P., Kornbluth, M., Molinari, N., Smidt, T

    Batzner, S., Musaelian, A., Sun, L., Geiger, M., Mailoa, J. P., Kornbluth, M., Molinari, N., Smidt, T. E., and Kozinsky, B. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications, 13 0 (1): 0 2453, 2022

    2022

  4. [4]

    A geometric-information-enhanced crystal graph network for predicting properties of materials

    Cheng, J., Zhang, C., and Dong, L. A geometric-information-enhanced crystal graph network for predicting properties of materials. Communications Materials, 2 0 (1): 0 92, 2021

    work page 2021

  5. [5]

    J., Mori-S \'a nchez, P., and Yang, W

    Cohen, A. J., Mori-S \'a nchez, P., and Yang, W. Challenges for density functional theory. Chemical Reviews, 112 0 (1): 0 289--320, 2012

    work page 2012

  6. [6]

    P., and Geiger, A

    Coors, B., Condurache, A. P., and Geiger, A. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pp.\ 518--533, 2018

    work page 2018

  7. [7]

    Crysmmnet: multimodal representation for crystal property prediction

    Das, K., Goyal, P., Lee, S.-C., Bhattacharjee, S., and Ganguly, N. Crysmmnet: multimodal representation for crystal property prediction. In Uncertainty in Artificial Intelligence, pp.\ 507--517. PMLR, 2023

    work page 2023

  8. [8]

    S., and Jurdak, R

    Dorri, A., Kanhere, S. S., and Jurdak, R. Multi-agent systems: A survey. Ieee Access, 6: 0 28573--28593, 2018

    work page 2018

Show all 41 references
  1. [9]

    and Lenssen, J

    Fey, M. and Lenssen, J. E. Fast graph representation learning with PyTorch Geometric . In ICLR Workshop on Representation Learning on Graphs and Manifolds, 2019

    2019

  2. [10]

    Se (3)-transformers: 3d roto-translation equivariant attention networks

    Fuchs, F., Worrall, D., Fischer, V., and Welling, M. Se (3)-transformers: 3d roto-translation equivariant attention networks. Advances in neural information processing systems, 33: 0 1970--1981, 2020

    1970

  3. [11]

    Dftb3: Extension of the self-consistent-charge density-functional tight-binding method (scc-dftb)

    Gaus, M., Cui, Q., and Elstner, M. Dftb3: Extension of the self-consistent-charge density-functional tight-binding method (scc-dftb). Journal of chemical theory and computation, 7 0 (4): 0 931--948, 2011

    2011

  4. [12]

    Molecular simulation of water and hydration effects in different environments: Challenges and developments for dftb based models

    Goyal, P., Qian, H.-J., Irle, S., Lu, X., Roston, D., Mori, T., Elstner, M., and Cui, Q. Molecular simulation of water and hydration effects in different environments: Challenges and developments for dftb based models. The Journal of Physical Chemistry B, 118 0 (38): 0 11007--...

    2014

  5. [13]

    Linear-scaling density-functional-theory technique: The density-matrix approach

    Hern \'a ndez, E., Gillan, M., and Goringe, C. Linear-scaling density-functional-theory technique: The density-matrix approach. Physical Review B, 53 0 (11): 0 7147, 1996

    1996

  6. [14]

    Dftb+, a software package for efficient approximate density functional theory based atomistic simulations

    Hourahine, B., Aradi, B., Blum, V., Bonafe, F., Buccheri, A., Camacho, C., Cevallos, C., Deshaye, M., Dumitric a , T., Dominguez, A., et al. Dftb+, a software package for efficient approximate density functional theory based atomistic simulations. The Journal of Chemical Physi...

    2020

  7. [15]

    Closed-loop framework for discovering stable and low-cost bifunctional metal oxide catalysts for efficient electrocatalytic water splitting in acid

    Jia, X., Zhou, Z., Liu, F., Wang, T., Wang, Y., Zhang, D., Liu, H., Wang, Y., Ye, S., Amezawa, K., et al. Closed-loop framework for discovering stable and low-cost bifunctional metal oxide catalysts for efficient electrocatalytic water splitting in acid. Journal of the America...

    2025

  8. [16]

    A., Thiessen, P

    Kim, S., Chen, J., Cheng, T., Gindulyte, A., He, J., He, S., Li, Q., Shoemaker, B. A., Thiessen, P. A., Yu, B., et al. Pubchem 2025 update. Nucleic Acids Research, 53 0 (D1): 0 D1516--D1525, 2025

    2025

  9. [17]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  10. [18]

    and Sham, L

    Kohn, W. and Sham, L. J. Self-consistent equations including exchange and correlation effects. Physical review, 140 0 (4A): 0 A1133, 1965

    1965

  11. [19]

    A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges

    Li, X., Wang, S., Zeng, S., Wu, Y., and Yang, Y. A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges. Vicinagearth, 1 0 (1): 0 9, 2024

    2024

  12. [20]

    Numerical methods for kohn--sham density functional theory

    Lin, L., Lu, J., and Ying, L. Numerical methods for kohn--sham density functional theory. Acta Numerica, 28: 0 405--539, 2019

    2019

  13. [21]

    Crystalflow: A flow-based generative model for crystalline materials

    Luo, X., Wang, Z., Wang, Q., Lv, J., Wang, L., Wang, Y., and Ma, Y. Crystalflow: A flow-based generative model for crystalline materials. arXiv preprint arXiv:2412.11693, 2024

    2024

  14. [22]

    Multi-agent systems and simulation: A survey from the agent commu-nity’s perspective

    Michel, F., Ferber, J., and Drogoul, A. Multi-agent systems and simulation: A survey from the agent commu-nity’s perspective. In Multi-Agent Systems, pp.\ 17--66. CRC Press, 2018

    2018

  15. [23]

    A., and Neese, F

    Orio, M., Pantazis, D. A., and Neese, F. Density functional theory. Photosynthesis Research, 102: 0 443--453, 2009

    2009

  16. [24]

    L., Buonassisi, T., and Butler, K

    Oviedo, F., Ferres, J. L., Buonassisi, T., and Butler, K. T. Interpretable and explainable machine learning for materials science and chemistry. Accounts of Materials Research, 3 0 (6): 0 597--607, 2022

    2022

  17. [25]

    Automatic differentiation in pytorch

    Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., and Lerer, A. Automatic differentiation in pytorch. In Advances in Neural Information Processing Systems Workshop on Autodiff, 2017

    2017

  18. [26]

    Multimodal neural network-based predictive modeling of nanoparticle properties from pure compounds

    Polat, C., Kurban, M., and Kurban, H. Multimodal neural network-based predictive modeling of nanoparticle properties from pure compounds. Machine Learning: Science and Technology, 5 0 (4): 0 045062, 2024

    2024

  19. [27]

    Quantumshellnet: ground-state eigenvalue prediction of materials using electronic shell structures and fermionic properties via convolutions

    Polat, C., Kurban, H., and Kurban, M. Quantumshellnet: ground-state eigenvalue prediction of materials using electronic shell structures and fermionic properties via convolutions. Computational Materials Science, 246: 0 113366, 2025 a

    2025

  20. [28]

    Understanding the capabilities of molecular graph neural networks in materials science through multimodal learning and physical context encoding, 2025 b

    Polat, C., Kurban, H., Serpedin, E., and Kurban, M. Understanding the capabilities of molecular graph neural networks in materials science through multimodal learning and physical context encoding, 2025 b . URL https://arxiv.org/abs/2505.12137

    2025 arXiv

  21. [29]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pp.\ 8748--8763. PmLR, 2021

    2021

  22. [30]

    O., Rupp, M., and von Lilienfeld, O

    Ramakrishnan, R., Dral, P. O., Rupp, M., and von Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data, 1, 2014

    2014

  23. [31]

    Graph neural networks for materials science and chemistry

    Reiser, P., Neubert, M., Eberhard, A., Torresi, L., Zhou, C., Shao, C., Metni, H., van Hoesel, C., Schopmans, H., Sommer, T., et al. Graph neural networks for materials science and chemistry. Communications Materials, 3 0 (1): 0 93, 2022

    2022

  24. [32]

    C., and Reymond, J.-L

    Ruddigkeit, L., Van Deursen, R., Blum, L. C., and Reymond, J.-L. Enumeration of 166 billion organic small molecules in the chemical universe database gdb-17. Journal of chemical information and modeling, 52 0 (11): 0 2864--2875, 2012

    2012

  25. [33]

    G., Hoogeboom, E., and Welling, M

    Satorras, V. G., Hoogeboom, E., and Welling, M. E (n) equivariant graph neural networks. In International conference on machine learning, pp.\ 9323--9332. PMLR, 2021

    2021

  26. [34]

    Equivariant message passing for the prediction of tensorial properties and molecular spectra

    Sch \"u tt, K., Unke, O., and Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning, pp.\ 9377--9388. PMLR, 2021

    2021

  27. [35]

    A multiagent-driven robotic ai chemist enabling autonomous chemical research on demand

    Song, T., Luo, M., Zhang, X., Chen, L., Huang, Y., Cao, J., Zhu, Q., Liu, D., Zhang, B., Zou, G., et al. A multiagent-driven robotic ai chemist enabling autonomous chemical research on demand. Journal of the American Chemical Society, 147 0 (15): 0 12534--12545, 2025

    2025

  28. [36]

    and De Fabritiis, G

    Th \"o lke, P. and De Fabritiis, G. Torchmd-net: equivariant transformers for neural network based molecular potentials. arXiv preprint arXiv:2202.02541, 2022

    2022 arXiv

  29. [37]

    and Wooldridge, M

    Van der Hoek, W. and Wooldridge, M. Multi-agent systems. Foundations of Artificial Intelligence, 3: 0 887--928, 2008

    2008

  30. [38]

    Huggingface's transformers: State-of-the-art natural language processing

    Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. Huggingface's transformers: State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019

    1910 arXiv

  31. [39]

    and Liu, C.-C

    Xie, J. and Liu, C.-C. Multi-agent systems and their applications. Journal of International Council on Electrical Engineering, 7 0 (1): 0 188--197, 2017

    2017

  32. [40]

    Crystal diffusion variational autoencoder for periodic material generation

    Xie, T., Fu, X., Ganea, O.-E., Barzilay, R., and Jaakkola, T. Crystal diffusion variational autoencoder for periodic material generation. arXiv preprint arXiv:2110.06197, 2021

    2021 arXiv

  33. [41]

    Mars: A multi-agent framework incorporating socratic guidance for automated prompt optimization

    Zhang, J., Wang, Z., Zhu, H., Liu, J., Lin, Q., and Cambria, E. Mars: A multi-agent framework incorporating socratic guidance for automated prompt optimization. arXiv preprint arXiv:2503.16874, 2025

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.