AI FM Design Studio · NASA ODSI

Why it matters

Designing a science foundation model is still expert-intensive, with no standard method. The studio makes it a repeatable, specification-driven process.

Scope, datasets, architecture and acceptance criteria are usually settled in conversations and workshops, and the reasoning stays with people rather than in documents a team can reuse. Two problems are easiest to fix during design, before time and compute go into training.

Scope before data

Check the data before fixing the scope

Teams can settle on a model's scope before checking whether the data capture the phenomena they want to study. The scale feasibility map checks that first.

Goals without tests

Every goal comes with its test

Teams can agree on broad goals without defining how they will be tested. Here each scientific requirement is linked to the tests and held-out data that will assess it.

Expert time

Expert time goes to judgment

Agents gather requirements, examine datasets and compare options. Each expert gives about 10 to 15 minutes up front, then reviews grouped questions.

The studio

Turn scientific domain expertise into an actionable engineering specification.

An AI design partner interviews the science team stage by stage—supporting both four-phase domain discovery and NASA ODSI's full eight-stage specification workflow—synthesizing schema-checked, versioned documents that the team reviews and signs off.

Guided design sessions

Structured, repeatable discovery.

Each stage unlocks sequentially with its own goal, reference context (PDFs, CSVs, code), and target schema. Stages can dispatch specialist sub-agents or run in-browser Python and JavaScript tools behind a fail-closed approval gate.

Atmosphere–land Spectral biogeochemistry Mars surface Lunar surface

AtmoHydro-X · Coupled Atmosphere–Land FMPhase 1 of 4

1 · Project Scope

In progress

2 · Use Cases

3 · Pretraining Data

4 · Architecture

What coupled land–atmosphere problem should the model address — water and energy exchange, or hydrometeorological extremes — and for whom?

Flash-drought onset. Water managers need two to six weeks of warning.

Good. What would count as success — which skill score, at what lead time?

Reply to the design partner…

Team · AtmoHydro-X4 of 5 filed

Science lead

Scale map · Project envelope

Filed

Hydrology expert

Use-case deep-dive

Filed

Remote-sensing expert

Data report

Filed

Atmospheric modeler

SME intake

Working

Merge stagewaiting on 1 expert

Whole-team collaboration

Parallel expert contributions, unified by the lead.

The science lead defines the project scope and invites domain specialists—each contributing their section from their own browser (or uploading existing drafts directly) before the merge stage synthesizes a single cohesive design.

Signed specifications

Versioned specifications with formal sign-off.

Every stage output—from Project Scope, Scale Map, and SME Data Reports to Architecture & Freeze and Validation—is schema-checked, versioned, and downloadable individually, as a ZIP archive, or as one combined Markdown specification.

L0_Requirements.md · v3Approved by team

Project Scope — AtmoHydro-X

Science problem

Capabilities

Success criteria

Use Cases → Data Design → Architecture →

The FM Design process

From the program mandate to a frozen architecture, in eight stages.

The studio runs NASA-IMPACT's agentic FM design process: eight stages and specialist AI agents, with each scientist working on their own laptop and schedule. Hand-offs travel through the team's shared, capped access key. It runs in the browser with nothing to install; one link per team gives access, with a spending cap and an expiry date.

Envelope

leadership · mandate and hard constraints

SME Intake

each scientist · 10 to 15 minute interview, or an upload

Data Report

each scientist · AI characterizes, the scientist approves

Scale Map

lead · merged and checked; the lead decides

Use-Case Deep-Dive

each scientist · one spec per agreed task

Data Design

team · pretraining data catalog and access

Architecture & Freeze

team · backbone, compute, freeze

Validation

team · traceability and benchmarks

leadership or leadeach scientist, on their own laptopthe team

How a team uses it

Leadership records the mandate and hard constraints once, and that Project Envelope reaches every expert.
Each scientist, independently, does a short intake interview, reviews an AI-produced data report for their tasks in plain language, and sends a confirmed data packet to the team in one click.
The lead's Scale Map starts itself once every packet is in. It computes what is observable at which scales and asks the lead only for decisions.
Each scientist then writes a use-case specification for one agreed task, followed by data design, architecture and freeze, and validation.

What it produces

A Scale Feasibility Map: a machine-checked document and chart saying, for every science target and dataset, whether it can be captured and why.
An agreed pretraining envelope: core, auxiliary and anchor datasets, the shared record period, and which tasks the model serves, each decision attributed.
Scientific Requirements, the final FM scope, out-of-scope items and a decisions log, then per-task Use-Case and Training-Data specifications, versioned and schema-validated.
An architecture plan that considers adapting existing models before committing to a new one, and a validation plan that traces each requirement to its tests, held-out data and leakage checks.

Why it can be trusted

Verdicts come from a deterministic checker, never from the model: the AI proposes, the checker decides. Experts can revise a finding, and their explanation is recorded.
Every fact carries its source: stated by a scientist, verified against primary documentation, or inferred. Nothing is approved without an explicit yes.
Scientists confirm their own data, so no one's expertise enters the record without their approval.

Demonstrated, October 2026. Run end to end on the live system with a lead and three domain scientists, covering forest biomass and canopy height, harmful algal blooms and water quality, and tornado damage detection. It produced an approved scope covering six tasks, with every document passing schema validation.

Open the AI FM Design Studio →

What the team sees

One studio, each person's view.

The studio's home: the team's foundation-model designs, each a staged process.

A design open at its first stage: the stage rail, the assistant's opening question, and the panels for documents, context, tracing and the team.

The Team tab: the lead, the experts expected, and the documents each person has filed.

Open and inspectable: the process, prompts, schemas and checker come from NASA-IMPACT's agentic FM design repository, and the studio is NASA-IMPACT/care-studio-lite, tested by over a thousand automated browser tests.

Open the AI FM Design Studio →

Design a foundation model with your science team.

Designing a science foundation model is still expert-intensive, with no standard method. The studio makes it a repeatable, specification-driven process.

Turn scientific domain expertise into an actionable engineering specification.

Structured, repeatable discovery.

Parallel expert contributions, unified by the lead.

Versioned specifications with formal sign-off.

From the program mandate to a frozen architecture, in eight stages.

How a team uses it

What it produces

Why it can be trusted

One studio, each person's view.

Open the studio