Two Teachers Better Than One: Hardware-Physics Co-Guided Distributed Scientific Machine Learning

Scientific machine learning (SciML) is increasingly applied to in-field processing, controlling, and monitoring; however, wide-area sensing, real-time demands, and strict energy and reliability constraints make centralized SciML implementation impractical. Most SciML models assume raw data aggregation at a central node, incurring prohibitively high communication latency and energy costs; yet, distributing models developed for general-purpose ML often breaks essential physical principles, resulting in degraded performance. To address these challenges, we introduce EPIC, a hardware- and physics-co-guided distributed SciML framework, using full-waveform inversion (FWI) as a representative task. EPIC performs lightweight local encoding on end devices and physics-aware decoding at a central node. By transmitting compact latent features rather than high-volume raw data and by using cross-attention to capture inter-receiver wavefield coupling, EPIC significantly reduces communication cost while preserving physical fidelity. Evaluated on a distributed testbed with five end devices and one central node, and across 10 datasets from OpenFWI, EPIC reduces latency by 8.9$\times$ and communication energy by 33.8$\times$, while even improving reconstruction fidelity on 8 out of 10 datasets.

💡 Research Summary

Scientific machine learning (SciML) has shown great promise for real‑world field applications such as seismic imaging, medical ultrasound, and environmental monitoring. However, most existing SciML pipelines assume a centralized architecture: raw sensor data are streamed from distributed receivers to a powerful central server where a deep neural network performs inference. In practice, especially in remote or harsh environments, this assumption quickly breaks down. High‑resolution waveforms generate massive data volumes, and transmitting them over limited‑bandwidth links (Wi‑Fi, 4G, satellite) leads to severe communication latency and energy consumption. Moreover, many SciML tasks, including full‑waveform inversion (FWI), rely on global physical coupling governed by partial differential equations; naïvely partitioning the computation across devices can violate these physical dependencies and degrade the quality of the reconstructed model.

The authors first quantify the bottleneck by deploying the state‑of‑the‑art InversionNet (a data‑driven FWI model) on a six‑node testbed (five edge devices and one central node). Under a realistic 4G emulation (15 Mbps uplink, 50 ms latency, 0.5 % packet loss), communication accounts for 93 % of total latency, confirming that raw‑data transmission is the primary obstacle. They then evaluate two off‑the‑shelf distributed learning paradigms: a federated‑learning‑style approach (FLA) where each edge runs the full InversionNet and sends its local velocity map, and a split‑learning‑style approach (SLA) where each edge runs only the encoder and sends latent features to a central decoder. Both methods dramatically reduce communication time (by >95 %) but suffer notable drops in reconstruction quality (SSIM decreases from 0.85 to 0.76–0.82). Visual inspection reveals boundary artifacts in FLA and loss of fine‑scale detail in SLA.

Through careful analysis the authors identify two physics‑driven root causes. First, acoustic waves obey the superposition principle: a source’s wavefield spreads throughout the entire medium, so receivers on one side still capture significant energy from structures on the opposite side. When each edge processes only its own local waveforms (as in FLA), crucial cross‑region information is lost, leading to artifacts. Second, the amplitude of reflected waves is strongly position‑dependent; some receivers record much stronger signals for a given region than others. SLA treats all latent vectors equally, ignoring this spatial weighting, which explains its inability to fully recover fine details.

Guided by these insights, the paper proposes EPIC (Edge‑compatible and Physics‑Informed distributed Scientific Machine Learning). EPIC’s design is driven by two “teachers”: a hardware teacher that pushes computation to the edge to minimize transmitted data, and a physics teacher that ensures the central decoder respects wave coupling. Concretely, each edge device runs a lightweight encoder that compresses raw waveforms into a compact latent vector (orders of magnitude smaller than the original data). The central node hosts a physics‑aware decoder built around a cross‑attention mechanism. The attention module receives all latent vectors, together with positional embeddings that encode receiver locations, and learns to weight each vector according to its physical relevance. This cross‑attention effectively reconstructs the global wavefield while preserving the underlying PDE constraints.

EPIC is organized into four system components:

1. EPIC‑Infra – formalizes the distributed sensing infrastructure, defining the set of edge devices, their assigned receivers, and network constraints (bandwidth b, latency l, packet loss p) together with a real‑time deadline T. It provides a multi‑objective formulation that balances latency, energy, and reconstruction fidelity.

2. EPIC‑Net – automatically constructs the neural architecture based on EPIC‑Infra parameters (e.g., number of edges, bandwidth). It includes the edge encoder, the central cross‑attention decoder, and optional self‑attention blocks for refinement.

3. EPIC‑Depl – handles model deployment, converting the trained EPIC‑Net into a format suitable for resource‑constrained edge hardware (quantization, pruning) and orchestrating the distribution of encoder weights.

4. EPIC‑Mgmt – runs at runtime, monitoring network conditions and inference latency. If the deadline is at risk, EPIC‑Mgmt can adapt encoder compression ratios or trigger a lightweight re‑training of attention weights, ensuring robustness to variable wireless conditions.

Experimental evaluation on the same six‑node testbed, using ten OpenFWI datasets, demonstrates the practical impact of EPIC. Compared with the centralized baseline, EPIC reduces end‑to‑end latency by 8.9× (average 598 ms vs. 5325 ms) and communication energy by 33.8× (≈71 mJ vs. 2417 mJ). Remarkably, EPIC improves SSIM on eight datasets (average increase of 0.01–0.03) and only marginally lowers it on the remaining two (by 0.3 %). The authors attribute this gain to the physics‑informed attention, which regularizes the decoder toward physically plausible solutions despite aggressive compression.

In summary, the paper makes three key contributions:

• It identifies and rigorously quantifies the communication bottleneck and the physical information loss that plague naïve distributed SciML approaches.
• It introduces a novel co‑guided design paradigm—hardware‑physics co‑guidance—that simultaneously addresses bandwidth constraints and preserves the underlying physics of wave propagation.
• It delivers a complete, end‑to‑end framework (EPIC) that spans infrastructure modeling, neural architecture synthesis, deployment, and adaptive management, and validates it on realistic hardware and benchmark datasets.

The work opens a clear path toward scalable, low‑latency, energy‑efficient scientific AI for field deployments, and its principles are likely transferable to other PDE‑based inverse problems beyond seismic imaging.