ASTER -- Agentic Science Toolkit for Exoplanet Research

This paper introduces ASTER, an agentic science toolkit that leverages large language models to orchestrate complex, multi-step exoplanet research workflows—such as data retrieval, radiative transfer modeling, and Bayesian analysis—by integrating specialized tools and providing intelligent guidance to streamline the characterization of exoplanet atmospheres.

The Gist

Imagine you are trying to bake a very complex, multi-layered cake, but the recipe is scattered across ten different cookbooks, the ingredients are locked in a high-security vault, and the oven requires you to program it with a language only a few experts speak. Furthermore, you have to do this while wearing oven mitts that make your fingers clumsy.

This is currently what it feels like to study exoplanet atmospheres (the air around planets orbiting other stars). Scientists have to manually hunt for data, download files, run complex physics simulations, and perform statistical math, often switching between different software programs that don't talk to each other.

Enter ASTER.

Think of ASTER not as a robot that replaces the scientist, but as a super-smart, highly organized sous-chef who has read every cookbook in the universe and knows exactly how to use every kitchen tool.

Here is how ASTER works, broken down into simple concepts:

1. The "Brain" (The Agent)

ASTER is built on a "Large Language Model" (like the AI you might chat with), but it's been given a special job. Instead of just chatting, it acts as a project manager.

• The Analogy: Imagine a conductor in an orchestra. The conductor doesn't play every instrument, but they know exactly when the violins should start, when the drums should hit, and how to keep the whole performance in sync. ASTER is the conductor for exoplanet research.

2. The "Hands" (The Tools)

The AI brain needs hands to do the work. ASTER is equipped with a set of digital "tools" that let it interact with the real world of science:

• The Librarian: It can instantly go to the "NASA Exoplanet Archive" (a giant digital library) and pull out the exact stats for a planet like WASP-39b (its size, mass, temperature) without the human having to search through pages of data.
• The Simulator: It can run a "virtual wind tunnel" (using a tool called TauREx) to simulate what the planet's atmosphere should look like based on physics.
• The Detective: It can download real observation data from space telescopes (like JWST), clean it up, and prepare it for analysis.
• The Mathematician: It runs complex "Bayesian retrieval" (a fancy way of saying "guessing the ingredients by tasting the cake") to figure out what gases are actually in the atmosphere based on the data.

3. The "Safety Net" (Guardrails)

You wouldn't let a robot chef run a kitchen without supervision, right? ASTER has built-in safety checks.

• The Analogy: Before the AI executes a dangerous command (like deleting a file or running a complex calculation), it has a "safety manager" that double-checks the action. If the AI tries to do something silly or dangerous, the safety manager stops it and asks the human, "Are you sure you want to do this?" This ensures the AI doesn't accidentally break the lab.

4. The "Assistant" (Not a Replacement)

The paper emphasizes that ASTER isn't there to replace the scientist. It's there to handle the boring, repetitive, and technical stuff so the scientist can focus on the big ideas.

• The Analogy: If you are writing a novel, you might use a spellchecker and a grammar tool. You don't let the spellchecker write the story for you; it just fixes the typos so you can focus on the plot. ASTER fixes the "typos" in the scientific workflow (downloading files, formatting data, running code) so the scientist can focus on interpreting what the atmosphere means.

The "WASP-39b" Test Drive

To prove it works, the authors gave ASTER a test case: WASP-39b, a hot, puffy planet that has been studied a lot.

1. The Ask: The human simply asked, "Can you get the data for WASP-39b, run a model, and tell me what's in its atmosphere?"
2. The Action: ASTER didn't just say "Okay." It actually:
   • Found the planet's stats.
   • Downloaded real telescope data.
   • Ran a simulation.
   • Performed the complex math to find the best fit.
   • Drew the graphs.
3. The Result: ASTER successfully recreated results that match what human experts have published, but it did it all in a seamless, automated conversation.

Why Does This Matter?

Currently, studying exoplanets is like trying to build a house with a hammer, a saw, and a screwdriver, but you have to walk to a different room to get each tool every time you need it.

ASTER builds a fully stocked workshop right next to the house. It brings all the tools together, organizes the blueprints, and helps you build the house faster and with fewer mistakes. It makes advanced space science accessible to more people, not just the few experts who know how to code all the different software programs.

In short: ASTER is the ultimate research assistant that turns a chaotic, multi-step scientific puzzle into a smooth, conversational workflow.

Technical Summary

1. Problem Statement

The characterization of exoplanet atmospheres, particularly via transmission spectroscopy, has become increasingly complex due to the proliferation of specialized modeling tools, vast observational datasets, and the need for sophisticated statistical inference (Bayesian retrieval).

• Fragmentation: Current workflows are fragmented across multiple notebooks, software packages, and manual steps (archival queries, forward modeling, data downloading, retrieval).
• High Barrier to Entry: Coordinating these steps requires substantial expertise in radiative transfer, coding, and statistical inference, limiting accessibility for early-career researchers and non-specialists.
• Reproducibility: Manual coordination of tools often leads to inconsistencies and difficulties in reproducing analysis pipelines.
• Need for Automation: There is a critical need for a flexible, user-friendly system that can orchestrate these multi-step tasks without replacing the scientific judgment of the researcher.

2. Methodology: The ASTER Framework

The authors introduce ASTER (Agentic Science Toolkit for Exoplanet Research), an AI agent built upon the Orchestral AI framework (Roman & Roman 2026). ASTER acts as a workflow coordinator that integrates Large Language Models (LLMs) with domain-specific tools.

Core Architecture

• Orchestral Framework: Provides the underlying infrastructure for context management, tool schema validation, and controlled execution. It supports provider-agnosticism (switching between OpenAI, Anthropic, Google, etc.) and synchronous execution for easier debugging.
• Agent Loop: The agent operates in an iterative loop:
  1. Receives context (user prompt + history).
  2. LLM reasons and decides on an action (natural language response or tool call).
  3. Tool Execution: If a tool is called, arguments are validated, safety hooks are applied, and the tool executes in a controlled environment.
  4. Feedback: The tool's output is inserted back into the context, and the loop repeats until the task is complete.
• Safety & Security: The system employs a hybrid safety mechanism:
  • Heuristic Checks: Block obviously dangerous commands (e.g., destructive shell commands).
  • Safeguard LLM: A secondary model classifies ambiguous actions as safe, unsafe, or uncertain.
  • User Approval: Uncertain or high-risk actions require explicit user confirmation.
• Tools vs. Skills: The authors distinguish between Tools (predefined, robust functions for specific operations like running a model) and Skills (natural language instructions or code snippets guiding the agent's behavior, e.g., "best practices for retrieval bounds"). This hybrid approach balances reliability with flexibility.

Domain-Specific Tool Implementation

ASTER extends the base Orchestral toolset with four key exoplanet-specific modules:

1. Data Acquisition:
   • Planetary Parameters: Downloads composite planetary and stellar parameters (radius, mass, temperature, metallicity) from the NASA Exoplanet Archive via the TAP service.
   • Observational Data: Downloads transmission spectroscopy datasets (e.g., JWST NIRSpec) from the NASA Exoplanet Archive. Since direct TAP access for legacy spectra is retired, the agent guides the user to generate wget commands, which the agent then executes to download and parse CSV files into standardized spectrum.dat and metadata.json formats.
2. Forward Modeling:
   • TauREx Integration: Interfaces with the TauREx radiative transfer code. The agent sets up paths for opacity (ExoMol) and Collision-Induced Absorption (CIA) files, configures isothermal temperature profiles, and generates synthetic transmission spectra based on retrieved planetary parameters.
3. Bayesian Retrieval:
   • TauREx Retrieval: Executes atmospheric retrievals using nested sampling optimizers (Nestle, MultiNest, UltraNest).
   • Configuration: Supports "reduced" (predefined molecules), "full" (custom molecules), and "equilibrium" (thermochemical equilibrium via ACE) modes. The agent manages parameter bounds (e.g., mixing ratios, temperature) based on embedded "best practice" skills.
4. Visualization:
   • Generates corner plots (using the corner library) to visualize posterior distributions and parameter correlations, overlaying results from multiple retrievals for comparison.

3. Key Contributions

• Unified Orchestration: ASTER provides the first end-to-end agentic framework specifically designed for exoplanet atmospheric analysis, seamlessly linking database queries, radiative transfer modeling, and Bayesian inference.
• Autonomous Workflow Management: The agent can decompose high-level user requests (e.g., "Retrieve the atmosphere of WASP-39b") into a sequence of sub-tasks (download parameters $\to$ download data $\to$ run forward model $\to$ run retrieval $\to$ plot results) without explicit step-by-step prompting.
• Adaptive Intelligence: The system demonstrates the ability to infer missing context (e.g., deducing spectral resolutions for instruments like Ariel or JWST when not explicitly provided) and correct errors (e.g., detecting missing directory structures and prompting for paths).
• Modularity and Extensibility: The toolkit is designed so that new tools (e.g., for new instruments or models) can be added as independent modules without altering the core architecture.

4. Results: The WASP-39b Case Study

The authors validated ASTER using WASP-39b, a well-studied hot Jupiter with extensive JWST data.

• Automated Setup: The agent successfully retrieved stellar and planetary parameters from the NASA Exoplanet Archive.
• Forward Modeling: It generated synthetic transmission spectra at native resolution and successfully rebinned them to simulate JWST (NIRSpec, NIRISS, MIRI) and Ariel resolutions, despite not being explicitly programmed with instrument specifications.
• Data Handling: The agent downloaded nine distinct NIRSpec datasets from the archive, parsed them, and visualized them.
• Retrieval Execution:
  • The agent performed two independent Bayesian retrievals using different datasets (JWST Transiting Exoplanet Community Early Release Science Team et al. 2023 vs. Rustamkulov et al. 2023).
  • It retrieved parameters including planet radius, temperature, and abundances of H$_2$O, CH$_4$, CO, CO$_2$, and NH$_3$.
  • Comparison: The agent generated corner plots comparing the two retrievals. It correctly identified that the broader wavelength coverage of the Rustamkulov et al. dataset led to tighter constraints (narrower posteriors) compared to the limited range of the first dataset, and noted the offset in best-fit values due to different atmospheric layers being probed.
• Outcome: The retrieved parameters (e.g., $R_p \approx 1.29-1.30 R_{Jup}$, $T \approx 666-1318$ K) were consistent with literature values, demonstrating the agent's ability to produce scientifically valid results autonomously.

5. Significance and Future Directions

• Accessibility: ASTER significantly lowers the technical barrier to entry for complex exoplanet analyses, enabling non-specialists to perform state-of-the-art retrievals.
• Reproducibility: By standardizing data ingestion, model execution, and output formatting, ASTER enhances the reproducibility of scientific workflows.
• Scientific Assistant Role: Rather than replacing human expertise, ASTER acts as a "scientific assistant" that handles technical implementation, error diagnosis, and routine tasks, allowing researchers to focus on interpretation and hypothesis generation.
• Future Potential: The framework is designed to expand beyond transmission spectroscopy to other areas of exoplanet science (e.g., emission spectroscopy, phase curves) and can integrate new tools as the field evolves. The authors call for community contributions to expand the tool library.

In conclusion, ASTER represents a paradigm shift in exoplanet research, demonstrating that agentic AI systems can effectively orchestrate complex, multi-step scientific pipelines, bridging the gap between raw data and physical interpretation.