Source-linked AI summary
A computable representation of the physical laboratory enables verifiable workflows
Xiaobo Li, Luyao Ge, Xiaohui Li, Lulu Guo, Ming Mao, Jiwang Zheng, Wenting Guan, Xin Yang, Yi Luo, Jun Jiang, Linjiang Chen
TL;DR
The paper addresses the need for a computable representation of the physical laboratory alongside machine-readable scientific knowledge. It combines typed research objects, capability-bound operations, compositional workflows, executable Function Skills, and stateful simulation to generate and verify workflows for diverse scientific intents.
Problem
Science requires representations of both scientific knowledge and the physical laboratory where scientific claims are tested.
Method
The framework combines typed research objects, capability-bound operations, a compositional workflow algebra, executable Function Skills, and stateful simulation for pre-dispatch verification.
Results
Four capability-relative workflows used registered capabilities and passed structural analysis and state simulation without mandatory errors or advisory warnings.
Takeaways & Limitations
The framework provides a computational interface between agent reasoning and capability-bound physical transformations for increasingly complex autonomous experiments.
Takeaways & Limitations
Generality is capability-relative and depends on each laboratory registering its local operations, constraints, and adapters.
Abstract
from arXiv · showhide
Making science computable requires representations of both scientific knowledge and the physical world in which scientific claims are tested. A computable representation of the physical laboratory is established through typed research objects, capability-bound operations and a compositional workflow algebra. It provides the physical-world counterpart to machine-readable knowledge, expressing workflows as programs over evolving laboratory states with explicit dependencies, decisions, iteration and concurrency. The representation was implemented in a modular agentic robotic laboratory by binding formal operations to executable Function Skills. For diverse scientific intents, capability-relative workflows were generated, while stateful simulation propagated object transformations and verified operation preconditions and laboratory constraints before dispatch. The proposed representation and its engineering framework jointly establish a general computational interface between agent reasoning and capability-bound physical transformations, providing a foundation for end-to-end autonomous scientific discovery.
Workflows as compositional programs
The framework expresses laboratory procedures as compositional programs over evolving typed object states. Its operators and checks separate scientifically plausible plans from operations admissible under registered capabilities.
- Workflow operators: Five operators compose laboratory workflows: sequence, bounded loop, conditional, parallel fork–join, and protected interval.They support dependencies, bounded iteration, evidence-based branching, concurrency, and resource-sensitive ordering.
- Workflow operators: Workflow operators are nested into programs over evolving laboratory state rather than idealized chemical procedures.
- Workflow checks: Type boundaries, loop bounds, branch guards, conflicting writes, and protected ordering are checked during composition.
- Capability-relative scope: Generality is capability-relative: each laboratory must register its local operations, constraints, and adapters.
Engineering executable semantics
The implementation compiles registered laboratory capabilities into executable Function Skills and validates generated Python workflows before dispatch. Stateful simulation uses operation contracts to commit valid transformations and localize mandatory failures.
- Capability compilation: 49 workstation specifications—45 physical and 4 virtual—were compiled into 63 agent-readable Function Skills.The functions cover laboratory handling, processing, characterization, analysis, and control-flow capabilities.
- Workflow admission: Generated Python workflows are restricted to registered calls and admitted through allow-listed abstract-syntax-tree analysis.
- Validate-then-commit: Stateful simulation evaluates mandatory and advisory conditions before applying effects, committing successful object-state updates.
- Validate-then-commit: Mandatory failures block updates and return localized diagnostics identifying conditions and suggested remediation.
Capability-relative generation from scientific intent
The system generated capability-relative workflows from scientific objectives and control logic across varied laboratory procedures. All four workflows used registered capabilities and passed structural and state-simulation checks without reported errors or warnings.
- Generated workflows: Four prompts generated workflows spanning sequence, branching, bounded iteration, fork–join parallelism, and protected execution.The cases covered catalyst synthesis, coupled synthesis and analysis, reaction optimization, and parallel characterization.
- Generated workflows: The generated workflows ranged from 20 to 27 nodes and used 11 to 15 workstations across the reported cases.The first case had 27 nodes across 11 workstations; the fourth had 27 nodes across 15 workstations.
- Validation outcomes: All four accepted workflows used only registered capabilities and passed structural analysis and state simulation with no mandatory errors or advisory warnings.
- Validation outcomes: The workflows were compositions of shared operations selected through object transitions and evidence dependencies rather than hard-coded protocol templates.
A representation layer for agentic laboratories
The paper unifies typed research-object state, capability-bound transitions, and executable checking into one representation layer for agentic laboratories. Its checks remain bounded by workstation contracts and simulated-state fidelity, leaving several execution realities to additional evidence or models.
- Unified representation: The contribution unifies typed physical research objects, compositional capability-bound transitions, and executable state checking.
- Unified representation: Object state remains an explicit operand throughout workflow generation, composition, and checking.
- Unified representation: The framework connects agent reasoning to physical action through an explicit computational object rather than prose alone.
- Bounded checks: Validity depends on complete workstation contracts and faithful simulated state, while scheduler isolation, telemetry, faults, repeatability, and chemical outcomes require further evidence or models.
Methods
The methods define typed research objects, capability-bound operations, and compositional workflows, then analyze and simulate generated Python workflows against laboratory constraints. Four workflow constructions were evaluated without wet-laboratory execution or inferential statistical analysis.
- Research-object and operation representation: Typed research objects capture identity, sample, container, location, evidence, provenance, and simulator-tracked operational fields.Workstation contracts declare which object fields each operation requires.
- Research-object and operation representation: Function Skills bind physical operations to workstation-specific requirements, parameters, constraints, state changes, returns, and diagnostics.Transitions are admitted only when all mandatory predicates hold for the relevant object states.
- Research-object and operation representation: Workflow composition uses sequence, bounded loops, conditionals, parallel fork–join, and protected intervals to express dependencies, iteration, decisions, concurrency, and resource-sensitive blocks.Sequential boundaries require type compatibility, while loop bounds can produce non-convergence diagnostics and joins require compatible branch writes.
- Structural analysis and stateful simulation: Generated workflows were parsed as Python abstract syntax trees and restricted to registered imports, allow-listed calls, supported statements, and a single entry point.Constant expressions and inactive branches were simplified during normalization.
- Structural analysis and stateful simulation: Stateful simulation reset the shared state, injected registered workstation functions, evaluated all applicable checks, and applied effects only after validation.Checks included container, sample, volume, mass, parameter, file, and workstation-dependency constraints, with diagnostics recording observed and expected values.
- Case construction and evaluation: The study evaluated four workflow constructions and the representation and software realization, without wet-laboratory outcome comparison or inferential statistical analysis.The four prompts covered coordination-environment modulation, layered double hydroxide synthesis, furfural hydrogenation optimization, and ZnO synthesis with adaptive dilution.
Data availability
Supplementary materials provide the registry, formal definitions, workflow cases, validation summaries, evidence manifest, and complete records for the four reported workflows.
- Data availability: Supplementary Data 1–4 provide the capability registry, formal definitions, four prompts, canonical workflows, validation summaries, and evidence manifest.Complete workflow and validation records for the four cases in Extended Data Fig. 1 are supplied as Additional File 3.
Code availability
Source code and the frozen Function-Skill library are supplied as additional files, while the provenance record is not presented as a public repository or persistent public access route.
- Code availability: Additional File 1 contains source code for Function-Skill conversion, workflow analysis, simulation, and serialization, and Additional File 2 contains the frozen Function-Skill library.Additional File 4 contains an internal release-page record for provenance only.
Additional information
The paper supplies supplementary materials defining typed research objects, operation contracts, workflow composition, simulation semantics and validation results. These materials document how the executable representation preserves object identity and provenance while checking laboratory constraints and state transitions.
- Scope: The simulator implements only a subset of the richer laboratory-state definition needed for workstation checking, and workflow generation reports no new wet-laboratory measurements.The implementation is therefore a bounded operational projection rather than a replacement for the full conceptual state.
- Research-object representation: Typed research objects record sample, container, evidence, provenance and handling fields while preserving persistent object identity across operations.Operations can update approved fields and append evidence without discarding identity, supporting provenance, branching and convergence.
- Operation contracts: Primitive operations are workstation-bound function calls with typed arguments, preconditions, constraints, state changes, returns and failure categories.Mandatory rules block acceptance and advisory rules generate warnings; state changes commit only when all mandatory rules pass.
- Execution and verification: The executable representation preserves workflow hierarchy as an abstract syntax tree and admits only registered, allow-listed operations before simulation.The simulator evaluates encoded conditions, applies successful state transitions and records execution diagnostics.
- Validation results: Validation cases passed structural analysis and stateful simulation with zero mandatory errors and zero advisory warnings across workflows using diverse control-flow patterns.Reported simulations included 27 steps, 31 steps, 23 executed steps and 30 steps, with workflow-generation templates returned for the cases.
Supplementary Figures
The supplementary figures document the representation’s stateful workflow operators, capability taxonomy, quality-control logic, simulation safeguards, and diverse laboratory workflow applications.
- Stateful workflows: The stateful workflow model tracks research-object states across repeated loop iterations and exits when the pH condition is satisfied.States S4(k)–S9(k) belong to the repeated loop body, while S10 follows the exit condition.
- Workflow operators: The five workflow operators have mandatory structural invariants; violations produce localized diagnostics and prevent the corresponding execution pattern.The operators cover sequence, loop, conditional, parallel, and protected-interval execution.
- Capability inventory: The capability inventory organizes 49 workstation specifications and representative operations across synthesis, testing, characterization, and virtual control-flow categories.The taxonomy spans material and container management, sample handling, processing, separation, reaction, and characterization.
- Skill assessment: Function-Skill quality assessment combines structural checks, semantic scoring, special-case handling, score aggregation, and threshold-based admission.The rubric assigns a total score of 100 points across required and optional items.
- Recipe validation: Static analysis admits, normalizes, and conservatively optimizes generated Python recipes before simulation, with O0–O2 enabled and O3–O4 disabled.Rejected recipes are returned for regeneration, while accepted recipes are serialized for simulation and review.
- Workflow applications: Supplementary workflows demonstrate coordinated synthesis, characterization, electrochemical testing, adaptive spectroscopy, chromatographic optimization, and archived simulation validation.Examples include Fe coordination-environment comparison, LDH HMF oxidation, furfural hydrogenation, and ZnO spectroscopy workflows.