Pith. sign in

REVIEW 2 major objections 8 minor 36 references

LLM Agent Bridges Natural Language to XANES Simulation

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-06 06:07 UTC pith:FST23ET5

load-bearing objection Useful workflow paper with a real inter-component compatibility gap between RAG-retrieved parameters and downstream normalization the 2 major comments →

arxiv 2604.16205 v2 pith:FST23ET5 submitted 2026-04-17 cond-mat.mtrl-sci cs.AIphysics.chem-ph

ChemGraph-XANES: An Agentic Framework for XANES Simulation and Curation

Vitor F. Grizzi , Thang Duc Pham , Luke N. Pretzie , Jiayi Xu , Murat Keceli , Cong Liu This is my paper
classification cond-mat.mtrl-sci cs.AIphysics.chem-ph
keywords executionxaneschemgraph-xanesframeworkgenerationsimulationagenticcomputational
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper presents ChemGraph-XANES, a software framework that uses a large language model (LLM) agent to orchestrate X-ray absorption near-edge structure (XANES) simulations. The core claim is that a documentation-grounded, schema-constrained orchestration layer can connect natural-language requests—such as 'compute the XANES for Ti in TiO2'—to deterministic execution of the FDMNES simulation code, while preserving full provenance linking each output spectrum to its input structure and parameters. The framework retrieves simulation parameters from the FDMNES manual via retrieval-augmented generation (RAG), validates them against typed schemas, and routes them through a common Python backend that handles structure acquisition, input file generation, execution, spectral parsing, and normalization. Three proof-of-capability demonstrations show: (1) a RAG agent retrieving an energy-range parameter from the manual and propagating it into a validated tool call, (2) file-to-spectrum execution from a local POSCAR structure file, and (3) chemistry-level natural-language specification resolving a composition query into a database lookup, absorber identification, and full simulation. The central design principle is controlled autonomy: the LLM interprets requests and populates constrained parameters, but never authors arbitrary simulation inputs, ensuring that parameter provenance remains traceable to either the user, the documentation, or an explicit default.

Core claim

The paper establishes that an LLM-based agent can serve as a trustworthy orchestration layer for computational spectroscopy by operating through schema-validated tool interfaces rather than free-form generation. The key mechanism is a three-layer architecture: a deterministic backend for scientific computation, a schema-constrained orchestration layer that translates natural-language requests into validated tool calls, and a documentation-grounded retrieval layer that consults external manuals to inform parameter selection. The recorded trace showing a FDMNES Range parameter retrieved from the manual and propagated into an executable tool call demonstrates that simulation parameters can be在外

What carries the argument

The load-bearing mechanism is the coupling of a RAG-enabled expert agent to schema-validated execution tools. The agent queries a knowledge base built from the FDMNES manual, retrieves parameter guidance, and passes numerical values through typed function arguments that are validated before reaching the deterministic backend. This design ensures that every parameter entering a simulation is attributable to an identifiable source—user input, retrieved documentation, or implementation default—rather than to opaque model generation.

Load-bearing premise

The paper assumes that the LLM-based agent can reliably interpret natural-language requests and retrieve correct parameters from documentation in non-trivial cases, but only demonstrates three curated examples without measuring retrieval accuracy, answer correctness, or robustness to ambiguous manual passages at scale.

What would settle it

Run a large, diverse set of natural-language XANES requests through the agent and measure the end-to-end success rate: how often does the agent correctly interpret the request, retrieve accurate parameters from the FDMNES manual, generate valid inputs, and produce scientifically meaningful spectra? If the success rate is low outside curated examples, the framework's practical utility is limited.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Computational spectroscopy workflows that currently require manual scripting and expert knowledge of code-specific keywords could become accessible through natural-language interfaces, lowering the barrier to large-scale XANES campaigns.
  • The provenance-aware design—preserving structure identifiers, input files, execution logs, and both raw and normalized spectra—provides a foundation for building machine-learning-ready structure-spectrum datasets with auditable data quality.
  • The task-parallel structure of the framework, combined with Parsl-based distributed execution, positions the system for future high-throughput XANES generation on HPC systems, enabling screening of structural databases and ensemble analysis.
  • The architectural separation of orchestration from simulation backend suggests that similar adapters could be built for other spectroscopy codes, enabling comparative-method studies within a unified data-curation framework.
  • The conservative parameter-exposure strategy—surfacing only a validated subset of FDMNES controls—provides a template for how other scientific automation systems can balance agent flexibility with input safety.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the RAG-based parameter retrieval proves reliable at scale, the same architecture could be extended to other scientific software packages with complex, documentation-heavy parameter spaces, potentially transforming how non-expert users interact with specialized simulation codes.
  • The framework's emphasis on provenance-aware data curation could become a standard requirement for computational spectroscopy datasets used in machine learning, where hidden heterogeneity in simulation settings can silently degrade model training.
  • The schema-constrained approach to agent autonomy suggests a broader design pattern for scientific AI agents: rather than granting LLMs unrestricted control over scientific software, trustworthy automation may require deliberately limiting the action space to validated, documented operations.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 8 minor

Summary. This manuscript presents ChemGraph-XANES, an LLM-based agentic framework that orchestrates XANES simulations using FDMNES. The framework combines retrieval-augmented generation (RAG) for parameter selection from the FDMNES manual, schema-constrained tool execution, deterministic input file generation, task-parallel execution via Parsl, spectral parsing, normalization helpers, and provenance-aware structure–spectrum linkage. Three proof-of-capability demonstrations are presented: (1) RAG-assisted retrieval and propagation of the FDMNES Range parameter into an executable tool call, (2) structure-file-based execution (Cu K-edge on a MnO₂ slab), and (3) chemistry-level natural-language specification (Ti K-edge in TiO₂). The paper is explicitly positioned as a workflow-construction paper rather than a benchmark study.

Significance. The paper addresses a genuine workflow bottleneck in computational XANES — the fragmented pipeline from structure acquisition through spectral curation — and proposes a principled, schema-constrained orchestration layer. The design choices (typed tool interfaces, deterministic backend, conservative parameter exposure, retention of raw spectra alongside normalized ones) are sound and well-motivated. The authors provide a public code repository and recorded agent traces, which supports reproducibility. The provenance-aware approach to linking spectra to their structural and parametric origins is a practical contribution for ML-oriented dataset construction. The framework's extensibility to other spectroscopy backends is a reasonable claim.

major comments (2)
  1. §3.1, Table 2 and §2.5: The flagship RAG demonstration retrieves the FDMNES default Range of −5 to 60 eV (0.5 eV step) and propagates it into run_xanes_single as energy_range=[-5.0, 0.5, 60.0]. However, the normalization procedure in §2.5 requires pre-edge fitting points at E ≤ E₀ − ΔE_pre = −20 eV. With the retrieved range starting at −5 eV, there are zero points in the pre-edge window, so the normalization falls back to max-intensity normalization — losing the edge-step normalization that the paper presents as a key post-processing feature. The paper acknowledges this ('routine ChemGraph executions can instead use the built-in ChemGraph mesh when a normalization-ready energy grid is desired'), but this acknowledgment does not resolve the tension with the central claim that the framework provides a coherent orchestration layer that 'connects' documentation-grounded parameter selection,F
  2. §3.1, Table 2 and §2.5: The flagship RAG demonstration retrieves the FDMNES default Range of −5 to 60 eV (0.5 eV step) and propagates it into run_xanes_single as energy_range=[-5.0, 0.5, 60.0]. However, the normalization procedure in §2.5 requires pre-edge fitting points at E ≤ E₀ − ΔE_pre = −20 eV. With the retrieved range starting at −5 eV, there are zero points in the pre-edge window, so the normalization falls back to max-intensity normalization — losing the edge-step normalization that the paper presents as a key post-processing feature. The paper acknowledges this ('routine ChemGraph executions can instead use the built-in ChemGraph mesh when a normalization-ready energy grid is desired'), but this acknowledgment does not resolve the tension with the central claim that the framework provides a coherent orchestration layer that 'connects' documentation-grounded parameter selection,
minor comments (8)
  1. §2.1: The distinction between 'single-agent' and 'multi-agent' execution modes is mentioned but never illustrated. Since all three demonstrations appear to use single-agent or RAG-augmented modes, clarifying which mode was used in each example would help readers understand the capabilities.
  2. §2.3: The list of exposed FDMNES controls (energy_range, radius, edge, magnetism, Green, Density_all, Quadrupole, Spherical, SCF) is provided, but the implementation-level defaults for each are not specified. Providing a table of defaults would improve reproducibility.
  3. Figures 2 and 3: The energy axes are not labeled with units (eV). The y-axes are labeled 'Normalized absorption' but the normalization method used (edge-step vs. max-intensity fallback) is not stated in the figure captions.
  4. §2.4: The statement 'A scientifically successful run requires at least one convolution output file' is important but buried. Consider highlighting this as a validation criterion, especially given the intended use for ML dataset generation where silent failures could corrupt training data.
  5. §3.1: The qualitative retrieval examples (Table 1) were generated with GPT-4o while the executable trace (Table 2) used Gemini 2.5 Flash. The rationale for using different models across examples is not explained.
  6. Reference [15] (XANE(3)) and [14] are co-authored by the present authors and are closely related. This is disclosed but could be stated more explicitly in the Introduction to contextualize the relationship between ChemGraph-XANES and the XANE(3) ML predictor.
  7. §2.2: The energy-above-hull threshold for Materials Project queries is mentioned but the default value is not specified.
  8. The abstract states 'provenance-aware data curation' but the provenance tracking described in §2.4 is limited to filename conventions and serialized structure files. A brief discussion of what metadata is captured (beyond MP IDs and chemical formulas) would strengthen this claim.

Simulated Author's Rebuttal

1 responses · 0 unresolved

The referee raises a single substantive point (duplicated in the report) concerning the tension between the RAG-retrieved FDMNES default energy range (−5 to 60 eV) and the normalization procedure's pre-edge fitting window (E ≤ E₀ − 20 eV). We agree this is a real limitation in the current demonstration and will revise the manuscript to address it more directly.

read point-by-point responses

  1. Referee: §3.1, Table 2 and §2.5: The flagship RAG demonstration retrieves the FDMNES default Range of −5 to 60 eV (0.5 eV step) and propagates it into run_xanes_single as energy_range=[-5.0, 0.5, 60.0]. However, the normalization procedure in §2.5 requires pre-edge fitting points at E ≤ E₀ − ΔE_pre = −20 eV. With the retrieved range starting at −5 eV, there are zero points in the pre-edge window, so the normalization falls back to max-intensity normalization — losing the edge-step normalization that the paper presents as a key post-processing feature. The paper acknowledges this ('routine ChemGraph executions can instead use the built-in ChemGraph mesh when a normalization-ready energy grid is desired'), but this acknowledgment does not resolve the tension with the central claim that the framework provides a coherent orchestration layer that 'connects' documentation-grounded parameter selection,

    Authors: We thank the referee for this careful observation, which identifies a genuine gap in the current demonstration. The referee is correct: the RAG-retrieved FDMNES default Range of −5 to 60 eV does not provide any points in the pre-edge fitting window (E ≤ −20 eV), so the normalization routine falls back to max-intensity normalization rather than performing edge-step normalization. We agree that the current manuscript's acknowledgment of this issue is insufficient and does not fully resolve the tension with our claim that the framework provides a coherent orchestration layer connecting documentation-grounded parameter selection to downstream post-processing. We will make the following revisions to address this point: (1) We will add an explicit discussion in §3.1 stating that the RAG-propagated default range is incompatible with the default normalization windows, and that this incompatibility illustrates an important design consideration: documentation-grounded retrieval of a software default does not automatically guarantee compatibility with downstream post-processing conventions. This is itself a meaningful finding for agent-mediated workflows, and we will frame it as such rather than burying it in a parenthetical. (2) We will add a second RAG demonstration trace in which the agent retrieves a normalization-compatible energy range (e.g., −30 to 80 eV with 0.5 eV step) from the FDMNES manual and propagates it into the execution call, showing that the same retrieval-and-propagation pathway can produce a mesh that supports edge-step normalization. This will demonstrate that the tension is not architectural but rather a consequence of which default value was retrieved in the specific trace. (3) We will revise the language in §2.5, §3.1, and the Discussion to clarify that RG revision: no

Circularity Check

0 steps flagged

No significant circularity: the framework's claims are grounded in external software and data sources, with only minor self-citation to the parent ChemGraph framework

full rationale

The paper presents ChemGraph-XANES as a workflow orchestration layer over established external tools (FDMNES, ASE, Materials Project, Parsl). The central claim—that the framework connects natural-language requests to deterministic XANES execution and data curation—is supported by three proof-of-capability demonstrations that route through these external dependencies. The RAG agent retrieves parameters from the FDMNES manual (an external document), not from the authors' own prior work. The only self-citation is to the parent ChemGraph framework [28], authored by some of the same team, which provides the underlying agent architecture. However, this citation is not load-bearing for the XANES-specific claims: the paper describes the XANES workflow components (FDMNES input generation, spectral parsing, normalization) in full detail within the manuscript itself, and the demonstrations use standard external inputs (TiO2, Fe2O3, Cu on MnO2). No 'prediction' or 'first-principles result' is claimed that would reduce to its inputs by construction. The paper explicitly frames its contributions as workflow construction and proof-of-capability demonstrations rather than benchmarking or spectral fidelity validation. The normalization procedure (§2.5) is defined by explicit equations with stated assumptions, not derived from self-cited prior work. The RAG-retrieved Range parameter (-5 to 60 eV) comes from the FDMNES manual, not from the authors' own results. While the skeptic correctly notes an incompatibility between the retrieved range and the normalization defaults, this is a correctness/coordination concern, not circularity—the paper does not claim the retrieved parameter produces normalized spectra, and explicitly notes the fallback behavior. No step in the derivation chain reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The framework introduces no new physical entities or forces. The free parameters are standard defaults for XANES normalization and energy meshes. The axioms are standard domain assumptions for computational chemistry workflows.

free parameters (3)
  • Pre-edge fitting window = E_0 - 20 eV
    Default parameter for spectral normalization.
  • Post-edge fitting window = E_0 + 50 eV
    Default parameter for spectral normalization.
  • Default energy mesh = [-5, 0.5, 60.0] eV
    Default energy range and step size used when not overridden by RAG or user.
axioms (3)
  • domain assumption FDMNES is a valid and reliable tool for XANES simulation.
    The framework uses FDMNES as its deterministic backend without questioning its scientific validity.
  • domain assumption LLMs can reliably interpret natural-language scientific requests and map them to structured tool calls.
    The agentic orchestration layer depends on this assumption, which is not formally benchmarked in the paper.
  • domain assumption The FDMNES manual is a sufficient knowledge base for parameter selection.
    The RAG agent uses the manual as its sole source of documentation for parameter retrieval.

pith-pipeline@v1.1.0-glm · 17578 in / 2099 out tokens · 1434490 ms · 2026-07-06T06:07:20.499873+00:00 · methodology

0 comments
read the original abstract

Computational X-ray absorption near-edge structure (XANES) is widely used to interpret local coordination environments, oxidation states, and electronic structure in chemically complex systems. In practice, routine computational XANES at scale is often constrained by workflow complexity rather than by the simulation method. We present ChemGraph-XANES, a large-language-model (LLM)-based agentic framework for XANES simulation and analysis that combines retrieval-augmented generation (RAG)-assisted parameter selection from documentation, schema-constrained tool execution, deterministic FDMNES input generation, and provenance-aware data curation. The framework supports both direct scripted execution and natural-language orchestration, with both modes routed through a deterministic backend for structure handling, absorber and edge specification, input generation, execution, spectral extraction, and post-processing. We demonstrate three proof-of-capability use cases: RAG-assisted selection and propagation of FDMNES input parameters, structure-file-based execution, and chemistry-level natural-language specification of absorber and composition requests. In a recorded trace, a simulation parameter is retrieved from the FDMNES manual by the RAG-enabled agent and propagated into a schema-validated tool call, illustrating traceable parameter selection. We further show that the same execution pathway supports both explicit local structures and chemistry-level user inputs. Because XANES calculations are independent once inputs are defined, ChemGraph-XANES is designed to support task-parallel execution and the creation of structure-linked XANES collections. ChemGraph-XANES therefore serves as a practical agentic framework for computational spectroscopy and data generation, emphasizing constrained orchestration, reproducibility, and traceable outputs.

Figures

Figures reproduced from arXiv: 2604.16205 by Cong Liu, Jiayi Xu, Luke N. Pretzie, Murat Keceli, Thang Duc Pham, Vitor F. Grizzi.

Figure 1
Figure 1. Figure 1: Schematic overview of the ChemGraph-XANES workflow. User requests can be provided either as chemistry-level natural-language queries or as explicit local structure files. These inputs are interpreted by an agentic orchestration layer that supports both single-agent and multi-agent execution, with an optional documentation-grounded expert agent that consults the FDMNES manual to inform parameter selection. … view at source ↗
Figure 2
Figure 2. Figure 2: Normalized XANES spectrum computed for a Cu absorber on a MnO [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Normalized XANES spectrum computed for a Ti absorber in bulk TiO [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 4 internal anchors

  1. [1]

    Jiayi Xu, Prajay Patel, A Jeremy Kropf, David Kaphan, Massimiliano Delferro, and Cong Liu. Theoretical investigation of the hydrogenation of cyclohexene catalyzed by supported single- atom sites on redox noninnocent limn2o4 and li2mn2o4 surfaces.The Journal of Physical Chemistry C, 128(12):4946–4957, 2024

    work page 2024

  2. [2]

    K-edge xanes investigation of fe-based oxides by density functional theory calculations.The Journal of Physical Chemistry C, 125(47):26229– 26239, 2021

    Jing Zhu, Zhenhua Zeng, and Wei-Xue Li. K-edge xanes investigation of fe-based oxides by density functional theory calculations.The Journal of Physical Chemistry C, 125(47):26229– 26239, 2021

    work page 2021

  3. [3]

    Fe k-edge x-ray absorption spectroscopy study of nanosized nominal magnetite.The Journal of Physical Chemistry C, 118(2):1332–1346, 2014

    Cristina Piquer, MA Laguna-Marco, Alejandro G Roca, R Boada, C Guglieri, and Jes´ us Chaboy. Fe k-edge x-ray absorption spectroscopy study of nanosized nominal magnetite.The Journal of Physical Chemistry C, 118(2):1332–1346, 2014

    work page 2014

  4. [4]

    Electrosynthesis of urea by using fe2o3 nanoparticles encapsulated in a conductive metal–organic framework.Nature Synthesis, 3(11):1404–1413, 2024

    Da-Shuai Huang, Xiao-Feng Qiu, Jia-Run Huang, Min Mao, Lingmei Liu, Yu Han, Zhen-Hua Zhao, Pei-Qin Liao, and Xiao-Ming Chen. Electrosynthesis of urea by using fe2o3 nanoparticles encapsulated in a conductive metal–organic framework.Nature Synthesis, 3(11):1404–1413, 2024

    work page 2024

  5. [5]

    Jiayi Xu, Colton Lund, Prajay Patel, Yu Lim Kim, and Cong Liu. Recent advances on compu- tational modeling of supported single-atom and cluster catalysts: characterization, catalyst– support interaction, and active site heterogeneity.Catalysts, 14(4):224, 2024

    work page 2024

  6. [6]

    Recent advances in x-ray absorption near edge structure (xanes) simulations for catalysis: Theories and applications.Annual Reports in Computational Chemistry, 20:157–187, 2024

    Jiayi Xu, Yu Lim Kim, Rishu Khurana, Shana Havenridge, Prajay Patel, and Cong Liu. Recent advances in x-ray absorption near edge structure (xanes) simulations for catalysis: Theories and applications.Annual Reports in Computational Chemistry, 20:157–187, 2024

    work page 2024

  7. [7]

    In silico chemical experiments in the age of ai: From quantum chemistry to machine learning and back.Advanced Materials, 36(30):2402369, 2024

    Abdulrahman Aldossary, Jorge Arturo Campos-Gonzalez-Angulo, Sergio Pablo-Garc´ ıa, Shi Xuan Leong, Ella Miray Rajaonson, Luca Thiede, Gary Tom, Andrew Wang, Davide Avagliano, and Al´ an Aspuru-Guzik. In silico chemical experiments in the age of ai: From quantum chemistry to machine learning and back.Advanced Materials, 36(30):2402369, 2024

    work page 2024

  8. [8]

    Machine-learned potentials for next-generation matter simulations.Nature Materials, 20(6):750–761, 2021

    Pascal Friederich, Florian H¨ ase, Jonny Proppe, and Al´ an Aspuru-Guzik. Machine-learned potentials for next-generation matter simulations.Nature Materials, 20(6):750–761, 2021

    work page 2021

  9. [9]

    Active-learned potentials reveal the interplay between the structure and dynamics in thorium molten salts.Physical Review Materials, 10(6):065606, 2026

    Vitor Ferreira Grizzi, Benjamin Nebgen, and Y Z. Active-learned potentials reveal the interplay between the structure and dynamics in thorium molten salts.Physical Review Materials, 10(6):065606, 2026

    work page 2026

  10. [10]

    The joint automated repository for various integrated simulations (jarvis) for data-driven materials design.npj computational materials, 6(1):173, 2020

    Kamal Choudhary, Kevin F Garrity, Andrew CE Reid, Brian DeCost, Adam J Biacchi, An- gela R Hight Walker, Zachary Trautt, Jason Hattrick-Simpers, A Gilad Kusne, Andrea Cen- trone, et al. The joint automated repository for various integrated simulations (jarvis) for data-driven materials design.npj computational materials, 6(1):173, 2020

    work page 2020

  11. [11]

    Enhancing materials property prediction by leveraging 16 computational and experimental data using deep transfer learning.Nature communications, 10(1):5316, 2019

    Dipendra Jha, Kamal Choudhary, Francesca Tavazza, Wei-keng Liao, Alok Choudhary, Care- lyn Campbell, and Ankit Agrawal. Enhancing materials property prediction by leveraging 16 computational and experimental data using deep transfer learning.Nature communications, 10(1):5316, 2019

    work page 2019

  12. [12]

    Zeeshan Ahmad, Tian Xie, Chinmay Maheshwari, Jeffrey C Grossman, and Venkatasubra- manian Viswanathan. Machine learning enabled computational screening of inorganic solid electrolytes for suppression of dendrite formation in lithium metal anodes.ACS central sci- ence, 4(8):996–1006, 2018

    work page 2018

  13. [13]

    Monte carlo thought search: Large language model querying for complex scientific reasoning in catalyst design

    Henry Sprueill, Carl Edwards, Mariefel Olarte, Udishnu Sanyal, Heng Ji, and Sutanay Choud- hury. Monte carlo thought search: Large language model querying for complex scientific reasoning in catalyst design. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 8348–8365, 2023

    work page 2023

  14. [14]

    Rational design of heterogeneous single-site catalysts via surface organometallic chemistry.Coordination Chemistry Reviews, 566:218186, 2026

    Jiayi Xu, Rishu Khurana, Taylor Harville, Vitor Ferreira Grizzi, Luke Pretzie, Anik Biswas, Prajay Patel, and Cong Liu. Rational design of heterogeneous single-site catalysts via surface organometallic chemistry.Coordination Chemistry Reviews, 566:218186, 2026

    work page 2026

  15. [15]

    XANE(3): An E(3)-Equivariant Graph Neural Network for Accurate Prediction of XANES Spectra from Atomic Structures

    Vitor F. Grizzi, Luke N. Pretzie, Jiayi Xu, and Cong Liu. XANE(3): An E(3)-equivariant graph neural network for accurate prediction of XANES spectra from atomic structures.arXiv preprint arXiv:2604.12140, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    Soft x-ray absorption spectroscopy of liquids and solutions.Chemical reviews, 117(23):13909–13934, 2017

    Jacob W Smith and Richard J Saykally. Soft x-ray absorption spectroscopy of liquids and solutions.Chemical reviews, 117(23):13909–13934, 2017

    work page 2017

  17. [17]

    React: Synergizing reasoning and acting in language models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations, 2022

    work page 2022

  18. [18]

    Large Language Model Agent: A Survey on Methodology, Applications and Challenges

    Junyu Luo, Weizhi Zhang, Ye Yuan, Yusheng Zhao, Junwei Yang, Yiyang Gu, Bohan Wu, Binqi Chen, Ziyue Qiao, Qingqing Long, et al. Large language model agent: A survey on methodology, applications and challenges.arXiv preprint arXiv:2503.21460, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Mixture-of-Agents Enhances Large Language Model Capabilities

    Junlin Wang, Jue Wang, Ben Athiwaratkun, Ce Zhang, and James Zou. Mixture-of-agents enhances large language model capabilities.arXiv preprint arXiv:2406.04692, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    From what to why: A multi-agent system for evidence-based chemical reaction condition reasoning.arXiv preprint arXiv:2509.23768, 2025

    Cheng Yang, Jiaxuan Lu, Haiyuan Wan, Junchi Yu, and Feiwei Qin. From what to why: A multi-agent system for evidence-based chemical reaction condition reasoning.arXiv preprint arXiv:2509.23768, 2025

    work page arXiv 2025

  21. [21]

    El agente: An autonomous agent for quantum chemistry.Matter, 8(7), 2025

    Yunheng Zou, Austin H Cheng, Abdulrahman Aldossary, Jiaru Bai, Shi Xuan Leong, Jorge Ar- turo Campos-Gonzalez-Angulo, Changhyeok Choi, Cher Tian Ser, Gary Tom, Andrew Wang, et al. El agente: An autonomous agent for quantum chemistry.Matter, 8(7), 2025

    work page 2025

  22. [22]

    Foundation models for materials discovery– current state and future directions.Npj Computational Materials, 11(1):61, 2025

    Edward O Pyzer-Knapp, Matteo Manica, Peter Staar, Lucas Morin, Patrick Ruch, Teodoro Laino, John R Smith, and Alessandro Curioni. Foundation models for materials discovery– current state and future directions.Npj Computational Materials, 11(1):61, 2025

    work page 2025

  23. [23]

    Unlocking azobenzene isomerization mechanisms via an llm agent-driven workflow integrating simulation, experiment, and machine learning

    Yixi Shen, Ledu Wang, Yan Huang, Xiaolong Zhang, Meng Huang, Huirong Li, Jing He, Aoran Cai, Yang Wang, Pieter ES Smith, et al. Unlocking azobenzene isomerization mechanisms via an llm agent-driven workflow integrating simulation, experiment, and machine learning. Chemical Science, 17(19):9663–9673, 2026. 17

    work page 2026

  24. [24]

    Atomgpt: Atomistic generative pretrained transformer for forward and inverse materials design.The Journal of Physical Chemistry Letters, 15(27):6909–6917, 2024

    Kamal Choudhary. Atomgpt: Atomistic generative pretrained transformer for forward and inverse materials design.The Journal of Physical Chemistry Letters, 15(27):6909–6917, 2024

    work page 2024

  25. [25]

    Steering towards safe self-driving laboratories.Nature Reviews Chemistry, 9(10):707–722, 2025

    Shi Xuan Leong, Caleb E Griesbach, Rui Zhang, Kourosh Darvish, Yuchi Zhao, Abhijoy Mandal, Yunheng Zou, Han Hao, Varinia Bernales, and Al´ an Aspuru-Guzik. Steering towards safe self-driving laboratories.Nature Reviews Chemistry, 9(10):707–722, 2025

    work page 2025

  26. [26]

    Ac- celerating catalytic advancements through the precision of high-throughput experiments & calculations.Digital Discovery, 5(2):497–509, 2026

    Jenny G Vitillo, Al´ an Aspuru-Guzik, Eric Doskocil, Omar K Farha, Timur Islamoglu, Heather J Kulik, Peter M Margl, Stuart Miller, Jordan Reddel, Aayush R Singh, et al. Ac- celerating catalytic advancements through the precision of high-throughput experiments & calculations.Digital Discovery, 5(2):497–509, 2026

    work page 2026

  27. [27]

    Accelerated inorganic materials design with generative ai agents.Cell Reports Physical Science, 6(12), 2025

    Izumi Takahara, Teruyasu Mizoguchi, and Bang Liu. Accelerated inorganic materials design with generative ai agents.Cell Reports Physical Science, 6(12), 2025

    work page 2025

  28. [28]

    Pham, Aditya Tanikanti, and Murat Ke¸ celi

    Thang D. Pham, Aditya Tanikanti, and Murat Ke¸ celi. ChemGraph as an agentic framework for computational chemistry workflows.Communications Chemistry, 9(1):33, 2026

    work page 2026

  29. [29]

    DREAMS: Density Functional Theory Based Research Engine for Agentic Materials Simulation

    Ziqi Wang, Hongshuo Huang, Hancheng Zhao, Changwen Xu, Shang Zhu, Jan Janssen, and Venkatasubramanian Viswanathan. Dreams: Density functional theory based research engine for agentic materials simulation.arXiv preprint arXiv:2507.14267, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    Cactus: Chemistry agent connecting tool usage to science.ACS omega, 9(46):46563–46573, 2024

    Andrew D McNaughton, Gautham Krishna Sankar Ramalaxmi, Agustin Kruel, Carter R Knut- son, Rohith A Varikoti, and Neeraj Kumar. Cactus: Chemistry agent connecting tool usage to science.ACS omega, 9(46):46563–46573, 2024

    work page 2024

  31. [31]

    Towards end-to-end automation of ai research.Nature, 651(8107):914–919, 2026

    Chris Lu, Cong Lu, Robert Tjarko Lange, Yutaro Yamada, Shengran Hu, Jakob Foer- ster, David Ha, and Jeff Clune. Towards end-to-end automation of ai research.Nature, 651(8107):914–919, 2026

    work page 2026

  32. [32]

    The atomic simulation environ- ment—a python library for working with atoms.Journal of Physics: Condensed Matter, 29(27):273002, 2017

    Ask Hjorth Larsen, Jens Jørgen Mortensen, Jakob Blomqvist, Ivano E Castelli, Rune Chris- tensen, Marcin Du lak, Jesper Friis, Michael N Groves, Bjørk Hammer, Cory Hargus, Eric D Hermes, Paul C Jennings, Peter Bjerre Jensen, James Kermode, John R Kitchin, Esben Leon- hard Kolsbjerg, Joseph Kubal, Kristen Kaasbjerg, Steen Lysgaard, J´ on Bergmann Maronsson,...

    work page 2017

  33. [33]

    Self-consistent aspects of x-ray absorption calculations.Journal of Physics: Condensed Matter, 21(34):345501, 2009

    O Bun˘ au and Yves Joly. Self-consistent aspects of x-ray absorption calculations.Journal of Physics: Condensed Matter, 21(34):345501, 2009

    work page 2009

  34. [34]

    Parsl: Pervasive parallel program- ming in python

    Yadu Babuji, Anna Woodard, Zhuozhao Li, Daniel S Katz, Ben Clifford, Rohan Kumar, Lukasz Lacinski, Ryan Chard, Justin M Wozniak, Ian Foster, et al. Parsl: Pervasive parallel program- ming in python. InProceedings of the 28th International Symposium on High-Performance Parallel and Distributed Computing, pages 25–36, 2019

    work page 2019

  35. [35]

    Com- mentary: The materials project: A materials genome approach to accelerating materials inno- vation.APL materials, 1(1), 2013

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, et al. Com- mentary: The materials project: A materials genome approach to accelerating materials inno- vation.APL materials, 1(1), 2013. 18

    work page 2013

  36. [36]

    Shyue Ping Ong, Shreyas Cholia, Anubhav Jain, Miriam Brafman, Dan Gunter, Gerbrand Ceder, and Kristin A Persson. The materials application programming interface (api): A simple, flexible and efficient api for materials data based on representational state transfer (rest) principles.Computational Materials Science, 97:209–215, 2015. 19

    work page 2015