PaperlayerView exact revision on arXiv

Source-linked AI summary

Autonomous Chemical Mechanistic Discovery through Agentic Reasoning and Validation

Dong Li, Sixuan Mi, Zihao Ye, Huan Xiong, Tao XU, Tong Zhu, Aijia Zhang, Junqi Gao, Kaiyan Zhang, Shijie Wang, Bowen Zhou, Yuqiang Li, Biqing Qi

arXiv:2609.11147v1cs.AI

TL;DR

Reaction-mechanism discovery remains dependent on expert intervention because existing computational systems largely automate fixed workflows rather than open-ended mechanistic reasoning. ARCHE integrates generalist hypothesis generation, domain-specialized computational validation, and structured tool orchestration in a closed loop. Across three scenarios, it reconstructs stereocontrolling transition states, identifies a plausible radical pathway, and extracts an interpretable selectivity descriptor, while currently operating within a Gaussian-centric workflow.

  • Problem

    Existing systems automate predefined computational procedures but do not provide the closed-loop process needed for autonomous discovery of unresolved reaction mechanisms.

  • Method

    ARCHE couples a general reasoning model for mechanistic hypotheses with domain-specialized computational expertise and structured tool orchestration in an iterative reasoning–validation cycle.

  • Results

    ARCHE reproduces established stereocontrolling transition states, identifies a previously unreported radical pathway consistent with observations, and extracts a minimal torsional descriptor.

  • Takeaways & Limitations

    The three scenarios support ARCHE as a framework for autonomous, interpretable mechanistic exploration across reliability, discovery, and descriptor abstraction.

  • Takeaways & Limitations

    ARCHE currently operates within a Gaussian-centric workflow, although its modular tool registry provides a basis for broader extension.

Abstract

from arXiv · show

Unraveling reaction mechanisms is central to modern chemistry, yet automating these investigations remains challenging because computational workflows still rely heavily on expert intervention. Here we introduce ARCHE, an autonomous agentic system that integrates a general-purpose reasoning model, a domain-specialized computational chemistry model, and a structured tool registry to transform mechanistic inquiry into a scalable, self-validating process. ARCHE interprets scientific questions, generates and prioritizes mechanistic hypotheses, orchestrates computational workflows, and iteratively refines conclusions based on computed evidence within a closed loop. We validate its capabilities across three increasingly demanding scenarios: reconstructing stereocontrolling transition states and validating the corresponding reaction mechanism in a previously reported asymmetric catalytic reaction; proposing and validating a plausible radical pathway through iterative hypothesis refinement for a recently discovered but unpublished $α$-iodoboronate C-I cleavage reaction; and identifying a chemically interpretable descriptor that governs selectivity in nickel-catalysed migratory cross-coupling reactions. By coupling agentic reasoning with rigorous computational validation, ARCHE advances autonomous mechanistic discovery and establishes a foundation for broader machine-assisted chemical research. The code for ARCHE is publicly available at https://github.com/JetAstra/Arche-Harness.

Introduction

ARCHE addresses the difficulty of autonomously investigating unresolved reaction mechanisms by coupling open-ended hypothesis generation with computational validation and iterative refinement. It demonstrates this framework across stereochemical reconstruction, radical-mechanism discovery, and interpretable selectivity analysis.

  • Motivation: Existing chemistry agents mainly execute predefined protocols, limiting open-ended hypothesis generation, discriminative calculation design, and iterative mechanistic refinement.Autonomous investigation remains constrained because workflow execution, input/output checking, and mechanistic reasoning are not fully integrated.
  • ARCHE framework: ARCHE closes the proposal–execution–reflection loop by evaluating computational evidence against mechanistic expectations without human intervention.Reflection modules verify logical coherence, validate Gaussian inputs, analyze outputs, and refine hypotheses.
  • ARCHE framework: ARCHE combines a generalist reasoning model, ARCHE-Chem, and a structured tool registry to support autonomous, self-validating mechanistic discovery.The generalist model proposes hypotheses and workflows, while ARCHE-Chem refines computational parameters, checks consistency, interprets outputs, and diagnoses failures.
  • Validation scenarios: In an asymmetric organocatalytic benchmark, ARCHE reconstructs stereocontrolling transition states and reproduces experimental selectivity trends.This scenario tests methodological reliability against a previously studied reaction.
  • Validation scenarios: For an unpublished radical transformation, ARCHE iteratively evaluates competing hypotheses and converges on a plausible mechanistic explanation.The case demonstrates mechanistic discovery rather than reproduction of an established pathway.
  • Validation scenarios: For nickel-catalysed migratory cross-coupling, ARCHE extracts a compact descriptor governing ligand-controlled regioselectivity and resolves a mechanistic ambiguity.The descriptor abstracts complex catalytic behavior into an interpretable chemical principle.

ARCHE System Architecture

ARCHE couples generalist mechanistic reasoning, domain-specialized Gaussian expertise, and a structured tool registry in an iterative reasoning–validation loop. Across its validation cases, the system supports computationally grounded mechanism reconstruction, radical-pathway identification, and selectivity-descriptor abstraction.

  • Core architecture: ARCHE combines a dual-model reasoning layer with a structured tool registry and execution interface to support iterative reasoning–validation.A generalist model explores pathways and plans workflows, while ARCHE-Chem reviews Gaussian settings, interprets outputs, and diagnoses failures.
  • Closed-loop workflow: Competing hypotheses are generated and prioritized, translated into computational workflows, executed, and evaluated for logical coherence, consistency, and mechanistic plausibility.Detected inconsistencies trigger revisions to hypotheses, parameters, or workflow configuration.
  • ARCHE-Chem evaluation: ARCHE-Chem achieved 70.65%, 65.60%, 80.30%, and 87.60% accuracy across CGB, GRG, GRA, and GEH, respectively.GPT-5.5 led CGB at 82.71%, whereas ARCHE-Chem led GRG, GRA, and GEH; its largest route-generation margin was 65.60% versus 44.80% for GPT-o3-mini.
  • Mechanistic validation: In the asymmetric aldol case, ARCHE identified 12 transition states, recovered the reported stereocontrolling structures, and matched the experimental selectivity gap.The computed activation-free-energy difference was 1.9 kcal/mol versus 1.6 kcal/mol inferred from 88% ee, with analysis completed within 24 hours without manual intervention.
  • Mechanistic validation: For the unresolved α-iodoboronate reaction, ARCHE identified a CsPPh2-complex whose approximately 64.0 kcal/mol excitation matches 450 nm irradiation and supports homolytic C–I cleavage.The refined radical pathway was consistent with available experimental evidence, including TEMPO inhibition.
  • Mechanistic abstraction: In nickel-catalysed migratory cross-coupling, ARCHE identified Br–N–N–H torsional distortion energy as an interpretable descriptor of ligand-dependent selectivity.Ortho substitution lowers the torsional energy penalty and favours hydride migration, whereas other substitution patterns favour the non-migratory pathway.

Discussion

ARCHE couples general reasoning with domain-specialized computational expertise and iterative validation to organize autonomous mechanistic inquiry. Across three scenarios, it reproduces established mechanisms, identifies a plausible radical pathway, and extracts a chemically interpretable selectivity descriptor.

  • The framework couples a general reasoning model for hypothesis exploration with a domain-specialized computational chemistry model that enforces methodological validity.An iterative reasoning–validation cycle supports systematic exploration while maintaining computational reliability.
  • Across three scenarios spanning reliability, discovery, and mechanistic abstraction, ARCHE’s results highlight the value of embedding domain-level expertise in mechanistic analysis.The scenarios cover established stereocontrol, a newly identified radical pathway, and selectivity abstraction.
  • ARCHE reproduces established stereocontrolling transition states with quantitative agreement to experiment, identifies a previously unreported radical pathway, and extracts a minimal torsional descriptor governing migratory selectivity.
  • By integrating computational expertise with systematic hypothesis evaluation, ARCHE supports more reliable and chemically grounded mechanistic analysis than general-purpose models or workflow automation alone.
  • The study frames reliable scientific AI as requiring specialised knowledge, rigorous validation, and scientifically interpretable conclusions under real methodological and practical constraints.
  • ARCHE currently operates within a Gaussian-centric workflow, although its modular registry and dual-model architecture are presented as a blueprint for extension to other software and closed-loop laboratory systems.

Methods

ARCHE combines generalist reasoning, domain-specialized computational chemistry, and a structured tool registry into a closed-loop system for autonomous reaction-mechanism discovery. It generates hypotheses, executes computational workflows, and revises conclusions using calculated evidence.

  • System architecture: ARCHE couples a generalist reasoning model with the domain-specialized ARCHE-Chem model and a structured tool registry for executable mechanistic workflows.The registry specifies operations, required inputs, expected outputs, and dependencies so hypotheses become schema-constrained computational procedures.
  • System architecture: Starting from a scientific question, ARCHE retrieves relevant knowledge, generates and ranks competing hypotheses, and translates them into workflows targeting decisive evidence.The workflows are executed through external computational chemistry tools.
  • Closed-loop validation: The system performs a reasoning–validation cycle in which hypotheses are formulated, computationally evaluated, and refined from calculated evidence.Reflection revises the hypothesis, workflow, or both when calculations fail, evidence conflicts with a claim, or computed results disagree with observed reactivity.
  • Domain-specialized model: ARCHE-Chem supports execution-critical chemistry by refining Gaussian settings, diagnosing calculation failures, and interpreting computational outputs.It was developed for domain terminology, theoretical concepts, Gaussian route generation, output analysis, and error diagnosis.
  • Training and independence: ARCHE-Chem was trained on curated computational-chemistry materials and approximately 50,000 instruction pairs, with no task-specific evaluation information included.The evaluation scenarios, hypotheses, structures, procedures, energetic results, and benchmark instances were excluded from training data.
  • Evaluation: The system was evaluated at model and system levels, including four computational-chemistry task categories and three end-to-end mechanistic demonstrations.The demonstrations covered established-mechanism reproduction, an unresolved mechanistic question, and extraction of an interpretable descriptor.

A Supplementary Methods S4

The supplementary methods document ARCHE’s execution flow, training pipeline, tool registry, computational protocols, and evaluation procedures. These materials support agent coordination, workflow construction, and three value demonstrations.

  • Architecture and execution: Supplementary methods specify agent-level architecture, execution flow, inter-agent communication, and cross-stage consistency.
  • Model development: They describe ARCHE-Chem’s development through training-pipeline, corpus, and instruction-dataset configurations.
  • Tool registry: The tool-registry materials define its schema, representative entries, workflow-construction mechanism, and example generated workflow.
  • Computational protocols: Supplementary protocols cover Gaussian calculations for asymmetric catalysis, a photochemical reaction, and descriptor-based mechanistic analysis.
  • Evaluation protocol: The evaluation protocol includes three value demonstrations, comparisons with external models, and success criteria for the demonstrations.

B Supplementary Discussion S31

The supplementary discussion organizes the three autonomous reasoning and validation case studies and provides their associated hypotheses, workflows, and supplementary results. It covers asymmetric catalysis, radical deiodoborylation, and descriptor-based migratory cross-coupling analysis.

  • Case studies: The supplementary case studies address asymmetric aldol catalysis, blue-light-driven deiodoborylation, and descriptor abstraction for migratory cross-coupling.
  • Supplementary results: Supplementary results are organized separately for each of the three case studies.
  • Agent prompts: The supplementary materials include prompts for retrieval, hypothesis, planning, and Gaussian-expert agents.
  • Hypotheses and formulations: Problem formulations and example hypothesis objects are provided for asymmetric catalysis, photochemical radical reaction validation, and dihedral-based selectivity validation.
  • Workflow examples: Example workflow objects are supplied for asymmetric catalysis, photochemical radical reaction validation, and dihedral-based selectivity validation.

A.1.1 Agent-level architecture

ARCHE implements mechanistic inquiry as a closed-loop system that connects retrieval, hypothesis generation, workflow planning, computation, and reflection. Its modular agents separate reasoning, execution, and interpretation while using computed evidence to revise mechanistic conclusions.

  • A.1.1 Agent-level architecture: Reflection evaluates structured computational outputs to determine whether evidence supports, refines, or rejects mechanisms, while provenance tracking preserves an auditable question-to-conclusion chain.Incomplete or non-diagnostic results are recorded rather than promoted into unsupported interpretations.
  • A.1.1 Agent-level architecture: ARCHE couples retrieval, hypothesis generation, planning, execution, and reflection in an iterative loop that uses computed evidence for subsequent revision.Retrieval informs hypotheses, hypotheses determine plans, plans enable execution, and execution supplies evidence for reflection.
  • A.1.1 Agent-level architecture: The architecture distributes scientific reasoning, workflow construction, computation, and interpretation across coordinated components rather than a single monolithic agent.This partitioning distinguishes the responsibilities and input/output types of the system's modules.
  • A.1.1 Agent-level architecture: The Retrieval Agent assembles literature and local knowledge into searchable context, which the Hypothesis Agent converts into ranked, testable alternative mechanisms.Hypotheses may differ in reactive species, elementary-step order, bond-making or bond-breaking events, and the origin of selectivity or reactivity.
  • A.1.1 Agent-level architecture: The Planner Agent maps prioritized hypotheses to executable workflows with ordered operations, dependencies, required tools, and expected intermediate outputs.Possible operations include conformer sampling, transition-state searches, optimization, frequency analysis, IRC calculations, and single-point energies.
  • A.1.1 Agent-level architecture: A general reasoning model checks workflow coherence, while ARCHE-Chem reviews domain-specific computational settings before execution.The two-stage review checks logical consistency and the chemical and technical appropriateness of methods and Gaussian keywords.

A.4.4 Descriptor-based mechanistic analysis

ARCHE evaluates competing mechanistic descriptors for nickel-catalysed migratory cross-coupling by requiring both predictive agreement and a direct connection to the underlying structural distortion. The retained dihedral descriptor provides a chemically interpretable account of ligand-dependent selectivity, whereas alternative geometric and electronic descriptors were less direct or consistently predictive.

  • A.4.4 Descriptor-based mechanistic analysis: The retained descriptor satisfies three criteria: tracking migratory-coupling selectivity, connecting directly to migration-related distortion, and enabling chemically interpretable ligand comparisons.These criteria were applied without assuming in advance that one geometric variable controlled selectivity.
  • A.4.4 Descriptor-based mechanistic analysis: The Br–N–N–H dihedral descriptor reproduces the ligand-dependent trend across the bipyridine series and directly reports the distortion required for hydride migration.Its torsional energy profiles capture the observed ligand-dependent behavior.
  • A.4.4 Descriptor-based mechanistic analysis: The descriptor links increased distortion energy along the relevant coordinate to a higher energy penalty for migration and a distinct reactivity pattern.This provides a mechanistically proximal and quantitatively interpretable account of the selectivity-determining process.
  • A.4.4 Descriptor-based mechanistic analysis: Bite-angle and electronic Ni–H bond descriptors were examined but rejected because they did not provide an equally direct or consistently predictive account of migration-relevant distortion.The alternatives captured some structural or electronic differences but did not explain ligand-dependent behavior as directly.
  • A.4.4 Descriptor-based mechanistic analysis: The demonstrations assess ARCHE at complementary levels of reconstruction, convergence, and abstraction rather than treating calculation completion alone as success.The broader validation spans recovery of known mechanisms, inference under unresolved prior knowledge, and extraction of mechanistic principles.

B.1.1 Case Study 1: Asymmetric aldol reaction

ARCHE autonomously refined competing mechanistic hypotheses into validated transition-state workflows for an asymmetric aldol reaction. It recovered reported stereocontrolling structures, identified additional candidates, and matched the observed enantioselectivity energetically.

  • The workflow generated five hypotheses, converted them into sampling, transition-state optimization, validation, and free-energy comparison workflows, and iteratively revised the computational protocol.The initial protocol was replaced with M06-2X/def2-TZVPP//B3LYP/6-31G(d) as better suited to transition-state ranking.
  • Filtering removed candidates that failed saddle-point convergence, relaxed to minima, or failed IRC validation, leaving chemically coherent and energetically plausible structures.
  • ARCHE identified 12 verified transition-state structures, including nine reported structures and three additional higher-energy candidates.The additional structures were chemically reasonable but not competitive.
  • The lowest-energy structures leading to opposite enantiomers corresponded to the previously reported competing stereocontrolling pathways.Their backbone RMSDs were 0.01 Å and 0.15 Å for the major and minor transition states, respectively.
  • A 1.9 kcal·mol−1 computed free-energy difference between the two lowest competing pathways closely matched the value inferred from experimental enantiomeric excess.The analysis was completed within 24h without manual intervention.

B.1.2 Case Study 2: Blue-light-driven deiodoborylation

ARCHE explored and computationally tested multiple mechanisms for blue-light-driven α-iodoboronate deiodination. Iterative elimination and reassessment identified a cooperative, photochemically accessible radical pathway consistent with the experimental conditions.

  • ARCHE generated 25 candidate strategies and consolidated them into six mechanistic hypotheses spanning concerted, direct-photoexcitation, and phosphine-associated activation scenarios.
  • A representative concerted C–I cleavage pathway had a 46.0 kcal·mol−1 barrier, excluding it as the dominant mechanism at room temperature.
  • Direct substrate excitation and a substrate-associated adduct required 102.8 and 75.8 kcal·mol−1, respectively, exceeding the 63.6 kcal·mol−1 energy supplied by 450 nm irradiation.
  • A cooperative CsPPh2-containing complex provided the validated activation mode, with −3.0 kcal·mol−1 association free energy and 64.0 kcal·mol−1 excitation energy matching 450 nm light.Excitation can promote homolytic C–I cleavage and generate radical intermediates.
  • The refined mechanism agreed with observations that product formation required blue light, base, and phosphine and was suppressed by TEMPO.

B.2.1 Supplementary results for case study 1

Supplementary conformational exploration recovered the reported transition-state ensemble while extending it with additional chemically reasonable motifs. RMSD comparisons supported faithful recovery of known structures.

  • The search identified multiple transition-state geometries spanning crown, chair-boat, boat-chair, and twist-derived conformations.
  • All previously reported low-energy transition-state structures, TS-2b through TS-2i, were recovered within the explored conformational ensemble.
  • Relative activation free energies were computed relative to the lowest-energy stereoisomeric transition structure to rank chemically relevant conformers.
  • RMSD analysis confirmed that the protocol faithfully recovered known transition-state structures, while newly identified conformers extended the accessible conformational space.

B.2.2 Supplementary results for case study 2

Supplementary calculations eliminated closed-shell, alternative bond-cleavage, and mismatched photochemical mechanisms using energetic, convergence, and wavelength constraints. The retained mechanism used 450 nm-accessible S2 excitation of a cooperative complex.

  • The closed-shell pathway had a 46.0 kcal·mol−1 barrier, indicating an extremely slow reaction under the experimental conditions.
  • Mechanisms B–D received no computational support because transition-state searches failed to locate the expected saddle points.
  • S1-based radical hypotheses agreed qualitatively with dark controls and radical inhibition but predicted ultraviolet rather than 450 nm activation.
  • The retained mechanism used S2 excitation at 447 nm with oscillator strength f=0.0567, matching the experimental 450 nm irradiation.S1 excitation at 649 nm had oscillator strength f=0.0044 and was predicted to be inefficient.
  • Formation of the CsPPh2 complex lowered free energy by 3.0 kcal·mol−1, supporting the mechanism as the most plausible pathway.

B.2.3 Supplementary results for case study 3

The study evaluates bite-angle and Ni–H electronic descriptors for ligand-controlled selectivity in nickel-catalysed migratory cross-coupling. Neither alternative directly captures the selectivity-controlling hydride-migration distortion, whereas the dihedral descriptor provides a more direct mechanistic account.

  • Bite-angle modulation: The ortho-substituted ligand has the smallest energetic penalty in constrained bite-angle scans, but bite angle does not directly map onto the migratory coupling distortion.The relevant change is an out-of-plane displacement accompanying hydride migration, captured by a dihedral coordinate.
  • Electronic modulation of the Ni–H bond: Wiberg bond indices show the largest Ni–H bond order for the ortho-substituted ligand, contradicting the predicted bond weakening.The result does not support the proposed electronic explanation based on Ni–H bond weakening.
  • Electronic modulation of the Ni–H bond: Ni–H bond dissociation energies are directionally consistent with easier cleavage in the ortho-substituted system but do not directly report the migratory barrier.BDEs compare bond-cleavage thermochemistry between initial and final states rather than energy evolution along the operative reaction coordinate.
  • Electronic modulation of the Ni–H bond: Together, the Wiberg index and BDE provide no consistent electronic explanation for the ligand-dependent selectivity trend.The two quantities probe different aspects of Ni–H bonding and neither directly represents the energetic requirement of migration.
  • Alternative mechanistic descriptors: The dihedral descriptor provides a more direct mechanistic description of the selectivity-controlling process than bite-angle or Ni–H bonding metrics.It explicitly probes the out-of-plane geometric change coupled to hydride migration.

C.8.2 Hypothesis for the radical reaction validation

The radical-reaction validation compares ionic, photochemical, SET, and radical-chain hypotheses using calculated barriers, excitation energies, bond strengths, and reaction pathways. The refined hypotheses converge on photochemical or SET initiation that generates a shared radical intermediate and supports the observed deiodination under mild conditions.

  • Alternative ionic and concerted SN2 hypotheses evaluate deprotonation, substitution barriers, transition-state connectivity, and overall free-energy feasibility.The direct substitution pathway is assessed against a typical thermal feasibility threshold, while the ionic route combines deprotonation and substitution free energies.
  • Additional mechanisms include iodine-atom transfer, radical-polar crossover, photoexcited phosphide activation, exciplex-mediated electron transfer, and PCET.These proposals provide alternative routes for C–I activation, radical generation, hydrogen delivery, or reaction termination.
  • A radical-chain hypothesis tests 450 nm C–I homolysis, α-boryl radical formation, hydrogen-atom transfer from HPPh2, and chain propagation.The proposed calculations include excitation energies, C–I bond dissociation energies, P–H bond strengths, HAT barriers, and radical stability.
  • Photochemical or SET initiation can generate the same key radical intermediate, explaining reactivity under photoredox–base conditions.The supplied passages describe this as a synergy that can account for the observed reaction.
  • The integrated photochemical/SET hypothesis accommodates initiation by visible light and a subsequent radical pathway for α-boryl radical formation.The proposed mechanism is framed as an integrated evaluation rather than a single isolated elementary step.
  • The resulting mechanism rationalizes the need for 2 equivalents of HPPh2, the observed P–P coupling product, and visible-light initiation feasibility.The proposed outcome is hydrodeiodination with PPh2–PPh2 formation while the carbon-centered radical remains transient and controlled.

C.8.3 Hypothesis for the dihedral-based selectivity validation

The dihedral-based selectivity validation evaluates how ligand substitution changes nickel coordination geometry, Ni–H and Br–Ni bond properties, and hydride-migration energetics. The unified hypothesis is that ortho-methyl substitution combines planar distortion with bromide lability to favor migratory product formation.

  • Relaxed scans and energy surfaces test whether ortho-substituted ligands have lower penalties for Ni–H displacement, bite-angle changes, and Br–Ni elongation.The workflow compares one- and two-dimensional energy profiles across ligand substitution patterns.
  • The hypotheses predict increased hydride→Ni donation, lower Ni–H bond dissociation energies and bond indices, and higher migratory product formation for ortho-substituted ligands.These descriptors are evaluated using distorted geometries and NBO analysis.
  • The proposed descriptor is a steric–electronic synergy in which pseudo-tetrahedral flexibility and Br–Ni lability jointly control migratory selectivity.The combined framework includes planar distortion, hydride departure, coordination flexibility, and halide rearrangement.
  • Structural pre-distortion and destabilizing sterics correlate with experimentally observed preference for migratory product formation.The proposed unified mechanism links ligand-induced geometry changes to migratory selectivity.
  • Ortho-methyl substitution simultaneously distorts the Ni(II) coordination environment and increases bromide mobility, facilitating hydride migration.The combined effects lower the geometric penalty for Ni–H displacement and ease halide repositioning or dissociation.
Loading 2609.11147v1…