GROWFLOW: Low-Latency Computational Fluid Dynamics at the Edge for Real-Time Digital Agriculture
In a new paper accepted to the IEEE Smart Connected Systems and Emerging Technologies Conference (IEEE SCSETech 2026), Liubov Kurafeeva, Benjamin Carter, Ryan Hartung, Alan Subedi, Avhishek Biswas, Michael Fay, Shantenu Jha, Chandra Krintz, Andre Merzky, Douglas Thain, Mehmet Can Vuran, and Rich Wolski introduce a distributed modeling architecture that bridges the gap between resource-constrained edge sensors and high-performance computing (HPC) clusters.
In their paper, GROWFLOW: Low-latency Computational Fluid Dynamics at the Edge for Real-time Agricultural Decision Support, they demonstrate how asynchronous model refreshing enables sub-second physical inference in dynamic, field-scale environments.
The Problem: The Time-Scale Mismatch in Smart IoT Systems
Modern cyber-physical systems—such as precision agriculture, smart infrastructure, and environmental monitoring—increasingly rely on high-fidelity physics models to infer conditions across large spaces from sparse sensor networks. However, deploying these models at the edge creates a fundamental systems bottleneck:
- Computationally Expensive Physics: High-resolution simulations like Computational Fluid Dynamics (CFD) require massive remote HPC resources and take tens of minutes to execute, making them far too slow for real-time edge decision-making.
- Intermittent Edge Connectivity: Field deployments often rely on bandwidth-limited or intermittent wireless links, meaning local edge nodes cannot block on remote data centers.
- Model Staleness: Because environmental conditions (such as wind speed and direction) shift continuously, static surrogate models trained on historical data rapidly lose accuracy over time.
The Solution: An Asynchronous Online/Offline Loop
Built atop the open-source xGFabric distributed platform and the CSPOT log-based runtime, GROWFLOW resolves this time-scale mismatch by splitting execution into two continuous, decoupled loops:
- Online Edge Inference: Deployed on an on-site Intel NUC cluster, the online path continuously ingests local weather station streams and serves spatial wind-field predictions using a lightweight Principal Component Regression (PC-R) surrogate model. Because inference never waits on remote HPC jobs, predictions remain non-blocking and resilient to network disruptions.
- Offline HPC Simulation & Retraining: In parallel, remote HPC nodes continuously ingest newly captured sensor boundary conditions to run parallel OpenFOAM CFD simulations (generating ~6.5 million 3D wind vectors), map them onto a 12,920-point spatial grid, and retrain updated PC-R surrogate models.
- First-Class Staleness Management: Rather than retraining on a rigid schedule, each offline iteration begins immediately upon the completion of the last, publishing fresh surrogate models asynchronously back to the edge to keep model staleness bounded.
Field-Scale Deployment & Key Findings
The team deployed and evaluated GROWFLOW at a commercial 4-acre Citrus Under Protective Screens (CUPS) facility in California, where dense mesh enclosures create complex internal microclimates that impact spray planning and frost mitigation.
Key Findings:
- Over 1,000x Latency Reduction: Once trained, the PC-R surrogate computes wind-field predictions across all 12,920 spatial locations in under one second on edge hardware—compared to ~22.7 minutes for a direct CFD simulation.
- Continuous 31-Minute Refresh Cadence: Across continuous multi-node HPC execution, the end-to-end offline simulation and retraining pipeline completed in an average of 31.2 minutes, keeping operational model staleness between 0.5 and 1 hour.
- Bounded Prediction Error: With continuous asynchronous refreshing, surrogate prediction error (MAE of ~0.65–1.0 m/s) remained comparable to the manufacturer-reported uncertainty of the physical weather sensors (0.44–0.88 m/s). Without refresh, error degraded substantially over 24 hours.
By treating model staleness as a measurable systems property and decoupling edge execution from remote simulation, GROWFLOW provides a scalable blueprint for real-time, adaptive intelligence across the edge-to-HPC continuum.