SciLink is a multi-agent AI platform that helps researchers automate, accelerate, and scale scientific workflows. Designed for complex experimental and computational campaigns, SciLink connects scientific knowledge, structured data, AI agents, analysis tools, simulation codes, and laboratory automation systems into a coordinated research environment.

SciLink supports scientists across the full discovery loop, from hypothesis generation and experiment design to data analysis, simulation, verification, and iterative refinement.

As part of the Center for Robotics and Autonomy @PNNL’s capabilities, SciLink provides an AI-enabled foundation for integrated learning across experiments, simulations, data, and decision-making.

What SciLink Does

SciLink combines large language model–based agents with scientific tools, user-supplied knowledge, structured data, and executable workflows. Its capabilities include:

Knowledge-Aware Scientific Reasoning

SciLink uses retrieval-augmented generation over user-provided papers, project notes, instrument manuals, prior results, and other research materials. This allows agents to generate hypotheses, design experiments, and make recommendations grounded in the scientific context of a project.

Structured Data Querying

SciLink complements document-based reasoning with agentic querying of structured data, including tabular files and record databases. Agents can dynamically generate and execute query code without requiring researchers to define a schema in advance.

Tool and Code Integration

SciLink can combine pre-built tools, user-provided models, and on-the-fly code generation to produce runnable analysis scripts, simulation inputs, or laboratory automation protocols. These workflows can be executed locally, on high-performance computing systems, or on connected laboratory instruments.

Pluggable Scientific Skills

SciLink can be extended through modular skill bundles that add support for new instrument data types, simulation methods, or domain-specific workflows. Domain experts can contribute self-contained markdown files and optional Python helpers that SciLink automatically discovers and routes.

Configurable Levels of Autonomy

SciLink supports three levels of autonomy:

• Co-pilot: The human researcher reviews and approves agent actions
• Autopilot: The system advances routine steps while escalating key decisions
• Autonomous: The agent system holds the acceptance gate for defined workflows.

This allows researchers to select the right level of oversight for each scientific task.

How SciLink Works

SciLink is organized as a multi-tier team of AI agents modeled after the structure of a research organization. At the top are three mode orchestrators that serve as project leads for:

• Experimental analysis
• Campaign planning
• Computational simulation.

Each orchestrator delegates work to specialist agents responsible for scientific tasks within their domain. The specialists are supported by role-specific agents that help generate, execute, verify, and refine outputs.

Within this workflow, dissent, verification, and reconciliation are built into the system. Separate agents review results, challenge assumptions, evaluate quality, and help determine whether a workflow should proceed, be revised, or escalate to a human researcher.

Supporting the Full Scientific Campaign Loop

SciLink is designed to support integrated research campaigns across experimental, analytical, and computational modes.

Campaign Planning

SciLink can support closed-loop optimization campaigns, moving from hypothesis to experiment, measurement, revision, and next-step selection. Multi-objective Bayesian optimization can guide the campaign, with optional generation of laboratory automation code when experiments are performed on robotic systems.

Experimental Analysis

SciLink can ingest and interpret data across major materials and life-science modalities, including:

• One-dimensional spectroscopy and kinetics, such as X-ray diffraction, Raman, UV-Vis, differential scanning calorimetry, enzyme kinetics, dose-response curves, plate-reader curves, quantitative polymerase chain reaction, and chromatography
• Two-dimensional imaging, such as optical microscopy, scanning electron microscopy, transmission electron microscopy, atomic force microscopy, atomic-resolution scanning transmission electron microscopy, fluorescence microscopy, and histology
• Three-dimensional hyperspectral datacubes, such as electron energy loss spectroscopy, energy-dispersive spectroscopy, and Raman imaging, with spectral unmixing workflows.

Computational Simulation

SciLink supports atomistic simulations across scales—from quantum (density functional theory) calculations to classical and machine-learning-driven molecular dynamics, including biomolecular systems—using established engines such as VASP and LAMMPS. It supports the full workflow, from building structures to preparing inputs, automatically assessing results and re-running with improvements when needed.

Adaptive Agentic Pipelines

SciLink includes simulated-annealing agentic pipelines that balance scientific rigor with adaptive problem solving. These pipelines initially hold scientific priors and domain rules tightly, then progressively relax implementation constraints only when iterative refinements fail to converge.

Verifier-driven acceptance steps help determine whether proposed changes improve the outcome before the system proceeds.

Why It Matters

SciLink turns scientific intent into actionable research: its AI agents design experiments, analyze multimodal data, and run simulations all while reasoning about the science, grounding decisions in the literature, and verifying their own work as they go. Adjustable autonomy provides flexible collaboration. Scientists and SciLink work side-by-side, pairing human intuition and judgment with the speed and reach of AI.