Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Model Recovery at the Edge under Resource Constraints for Physical AI

Bin Xu, Ayan Banerjee, Sandeep K. S. Gupta

TL;DR

The paper tackles the challenge of deploying Model Recovery (MR) for physical AI on resource-constrained edge devices. It introduces MERINDA, an FPGA-accelerated MR framework that replaces iterative ODE solvers with a GRU-based neural-flow architecture equivalent to NODEs, trained with an ODE loss to reproduce dynamics. The work provides a theoretical basis via equivalence proofs and the MACE theorem, and develops explicit memory and energy models to guide hardware design. Empirical results across simulated and real-world datasets show MERINDA achieves comparable accuracy to state-of-the-art MR methods while delivering substantial reductions in DRAM usage and energy, plus significant runtime benefits on FPGA relative to GPUs. The findings highlight a fundamental energy–memory tradeoff and demonstrate the practicality of edge-ready, reconfigurable MR for mission-critical autonomous systems.

Abstract

Model Recovery (MR) enables safe, explainable decision making in mission-critical autonomous systems (MCAS) by learning governing dynamical equations, but its deployment on edge devices is hindered by the iterative nature of neural ordinary differential equations (NODEs), which are inefficient on FPGAs. Memory and energy consumption are the main concerns when applying MR on edge devices for real-time operation. We propose MERINDA, a novel FPGA-accelerated MR framework that replaces iterative solvers with a parallelizable neural architecture equivalent to NODEs. MERINDA achieves nearly 11x lower DRAM usage and 2.2x faster runtime compared to mobile GPUs. Experiments reveal an inverse relationship between memory and energy at fixed accuracy, highlighting MERINDA's suitability for resource-constrained, real-time MCAS.

Model Recovery at the Edge under Resource Constraints for Physical AI

TL;DR

The paper tackles the challenge of deploying Model Recovery (MR) for physical AI on resource-constrained edge devices. It introduces MERINDA, an FPGA-accelerated MR framework that replaces iterative ODE solvers with a GRU-based neural-flow architecture equivalent to NODEs, trained with an ODE loss to reproduce dynamics. The work provides a theoretical basis via equivalence proofs and the MACE theorem, and develops explicit memory and energy models to guide hardware design. Empirical results across simulated and real-world datasets show MERINDA achieves comparable accuracy to state-of-the-art MR methods while delivering substantial reductions in DRAM usage and energy, plus significant runtime benefits on FPGA relative to GPUs. The findings highlight a fundamental energy–memory tradeoff and demonstrate the practicality of edge-ready, reconfigurable MR for mission-critical autonomous systems.

Abstract

Model Recovery (MR) enables safe, explainable decision making in mission-critical autonomous systems (MCAS) by learning governing dynamical equations, but its deployment on edge devices is hindered by the iterative nature of neural ordinary differential equations (NODEs), which are inefficient on FPGAs. Memory and energy consumption are the main concerns when applying MR on edge devices for real-time operation. We propose MERINDA, a novel FPGA-accelerated MR framework that replaces iterative solvers with a parallelizable neural architecture equivalent to NODEs. MERINDA achieves nearly 11x lower DRAM usage and 2.2x faster runtime compared to mobile GPUs. Experiments reveal an inverse relationship between memory and energy at fixed accuracy, highlighting MERINDA's suitability for resource-constrained, real-time MCAS.

Paper Structure

This paper contains 22 sections, 1 theorem, 13 equations, 4 figures, 7 tables, 2 algorithms.

Table of Contents

  1. Theoretical background
  2. Basics of Model Recovery
  3. Neural flows and equivalent architectures to NODE
  4. MERINDA Architecture
  5. Proof of Equivalence
  6. Energy-Memory tradeoff in Physics Guided MR
  7. MERINDA Memory model
  8. MERINDA Energy Model
  9. Proof of MACE theorem
  10. Accelerator Design Exploration
  11. Computation Optimization
  12. Memory Optimization
  13. Evaluation and Results
  14. Case Studies and Data
  15. Implementation Details
  16. ...and 7 more sections

Key Result

Theorem 1

MACE theorem: An acceleration strategy for physics informed model recovery can only guarantee two out of three constraints of memory footprint, recovery accuracy, and computational energy efficiency.

Figures (4)

  • Figure 1: FPGA acceleration strategy using neural flow based equivalent architecture to neural ODEschen2018neural.
  • Figure 2: Generic architecture for physics-guided inferencing to be used in physical AI.
  • Figure 3: MERINDA: GRU NN-based MR architecture.
  • Figure 4: Trade-off between memory and energy. The left plot corresponds to SINDY systems, while the right plot corresponds to MR systems. Each dot is annotated with its configuration in the format: (Nonlinear Terms (NL), Polynomial Order (PO), and State Variables (SV)). System complexity details are provided in Table \ref{['tbl:complexity']}.

Theorems & Definitions (1)

  • Theorem 1