An Agentic AI Framework to Accelerate Scientific Discovery in Plant Phenotyping¶

🕒 Published (v1): 2026-06-30 15:37 UTC · Source: Arxiv · link

Why this paper was selected

ORNL national lab; scientific discovery agent framework from a high-credibility group

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

This paper presents an end-to-end agentic AI framework deployed at Oak Ridge National Laboratory's Advanced Plant Phenotyping Laboratory (APPL) that bridges a conversational Co-Scientist Agent (cloud-hosted) with a Compute Agent running on the Frontier exascale supercomputer via a secure streaming channel. The system transforms multi-week, expert-bound trait extraction workflows into interactive, seconds-scale query loops. Provenance is captured end-to-end as a first-class output for reproducibility and foundation model curation.

Problem¶

APPL's automated imaging stations generate terabytes of multimodal imagery per experiment (up to 10,400 plants across 8 imaging modality groups), but downstream trait extraction remains manual, expert-bound, single-modality, and strictly post-hoc—making analysis, not acquisition, the bottleneck. Existing cloud-native agentic frameworks ignore the federation, data-movement, and security-domain realities of HPC-hosted scientific instruments.

Method¶

The framework is organized into four layers (Data, Compute, Agent, UI) spanning two security domains:

AI-Readiness Pipeline: Modality-specific readers normalize raw imagery (RGB, hyperspectral VNIR/SWIR, fluorescence FC1, thermal IR1, 3D laser, belowground NIR) into standardized NumPy arrays with a shared metadata schema keyed on {experiment_id}__{plant_id}__{round_id}.
ViT Segmentation Service: A vit_base encoder (patch size 8, window 448, pretrained on ImageNet, fine-tuned on APPL imagery) with a convolutional upsampling decoder performs plant/background segmentation via a sliding-window strategy with 50% stride and Hann-window blending; connected-component pruning removes spurious detections. Deployed on Frontier via Parsl-orchestrated GPU nodes.
Trait Extraction: Per-plant scalars are computed from segmented masks—morphological (height, projected leaf area, width, compactness, roundness from RGB), physiological (\(F_v/F_m\) from fluorescence, leaf temperature from thermal, chlorophyll proxies from RGB color).
Multi-Agent Architecture: A Co-Scientist Agent (Chainlit + LiteLLM, on AWS EKS) translates natural-language scientific questions into structured analysis plans. A Compute Agent (on Frontier) executes plans by dispatching ViT inference jobs. The two agents communicate over S3M, a secure token-authenticated streaming channel. FlowCept captures end-to-end provenance—prompts, parameters, model versions, intermediate artifacts—queryable via an MCP server. The interaction loop: context specification → availability check / on-demand inference → aggregation + report generation → follow-up with context reuse.

Key Contributions¶

End-to-end agentic framework that integrates cloud-hosted LLM/MCP services with Slurm-scheduled HPC workflows on Frontier, designed for real federated DOE facility constraints.
AI-readiness pipeline handling 8 multimodal imaging groups at terabyte-per-day scale with a unified metadata schema.
ViT-based segmentation and trait-extraction service producing time-resolved, per-plant trait tables across multiple modalities.
S3M secure, token-authenticated streaming channel enabling cross-security-domain agent coordination without persistent coupling.
End-to-end provenance system (FlowCept) treating auditability as a first-class output, supporting both reproducibility and training-data curation for downstream biology foundation models.
Interactive query loop reducing days-to-weeks analysis latency to seconds for follow-up queries within the same session.

Results¶

No quantitative benchmark numbers (accuracy, throughput SLAs, latency measurements) are reported in the provided text.
The system is deployed in production at APPL/ORNL; qualitative claims state the framework turns days-to-weeks workflows into seconds-scale interactive loops.
Covers 70+ tracked experiments spanning poplar (Populus trichocarpa), switchgrass (Panicum virgatum), pennycress (Thlaspi arvense), and Arabidopsis thaliana.
APPL scale: up to 520 plants per cycle, 700-foot track, 10,400 plants per multi-week campaign, terabytes of data per experiment.

Limitations¶

Several ecophysiologically important traits are not yet image-derived: photosynthetic rate, stomatal conductance, aboveground biomass, leaf mass per area, total protein, volatile emissions (isoprene, methanol)—still requiring destructive sampling or gas-exchange instruments.
No quantitative evaluation of segmentation accuracy, trait extraction error, or end-to-end latency is provided in the paper.
The architecture is tightly coupled to DOE/ORNL infrastructure (Frontier, OPAL data lakehouse, Slurm/Parsl); generalizability to other HPC environments is not demonstrated.
Multimodal foundation model fine-tuning for APPL (needed to integrate missing traits) is listed as future work, not yet delivered.
Hyperspectral and belowground modality integration details for trait extraction beyond segmentation are not fully described.

Relevance to Harnesses / Meta-Harnesses¶

This paper is a concrete production deployment of a meta-harness pattern: a Co-Scientist Agent acts as an orchestrator that translates user intent into structured plans, dispatches work to a domain-specific Compute Agent (itself a harness over Parsl/Slurm/ViT jobs), and closes the loop with provenance capture and result synthesis. The S3M inter-agent channel and FlowCept provenance layer illustrate how harnesses must be engineered for federated, multi-security-domain environments—a gap in cloud-only agentic frameworks. The four-layer decomposition (Data / Compute / Agent / UI) and context-reuse loop are directly reusable design patterns for harness builders targeting scientific HPC workflows. The MCP server exposing provenance queries to the Co-Scientist Agent exemplifies how harnesses can expose tool APIs over standard protocols to enable compositional agent behavior.