[Paper Review] Deep Learning on Real Geophysical Data: A Case Study for Distributed Acoustic Sensing Research

Vincent Dumont, Verónica Rodrı́guez Tribaldos+2 more|arXiv (Cornell University)|Oct 15, 2020

Seismic Waves and Analysis4 references4 citations

TL;DR

This paper presents a hyperparameter-optimized deep learning framework for identifying usable seismic energy in real Distributed Acoustic Sensing (DAS) data using ResNet architectures on HPC systems. By training on only 1,000 labeled images, the method achieves high accuracy in classifying coherent waves and noise, with a 100x speedup via GPU scaling and identification of four highly interpretable models matching ground-truth probability expectations across complex signal patterns.

ABSTRACT

Deep Learning approaches for real, large, and complex scientific data sets can be very challenging to design. In this work, we present a complete search for a finely-tuned and efficiently scaled deep learning classifier to identify usable energy from seismic data acquired using Distributed Acoustic Sensing (DAS). While using only a subset of labeled images during training, we were able to identify suitable models that can be accurately generalized to unknown signal patterns. We show that by using 16 times more GPUs, we can increase the training speed by more than two orders of magnitude on a 50,000-image data set.

Motivation & Objective

• To develop a scalable and interpretable deep learning classifier for identifying usable seismic energy in high-density, noisy DAS data.
• To address the challenge of classifying complex, non-coherent, and saturated signal patterns that are not easily labeled or generalized.
• To optimize hyperparameters across hardware, data representation, and model architecture for efficient training on large-scale geophysical datasets.
• To establish a methodology for model interpretability and reliability assessment in the absence of labeled data for complex signal types.
• To evaluate GPU scaling efficiency for training deep learning models on terabyte-scale geophysical data.

Proposed method

• A 9-dimensional hyperparameter search was conducted across HPC hardware (CPUs and GPUs), data representation (grayscale and RGB), model architecture (ResNet with varying depth and bottleneck size), and training settings (learning rate, epochs, batch size).
• The study used the Cori supercomputer at NERSC to train 6,460 distinct models, with a focus on GPU-based training using up to 32 V100 GPUs across 4 nodes.
• A parallel coordinates plot was used to visualize and compare all hyperparameter combinations, with the best model identified by proximity to ground-truth probability values.
• Model interpretability was assessed by evaluating output probabilities against expected ranges for four reference signal types: white noise, coherent waves, non-coherent waves, and saturated signals.
• A training throughput mapping was performed to determine optimal GPU scaling, showing a more than two orders of magnitude speedup when increasing from 16 to 32 GPUs on a 50,000-image dataset.
• Binary classification was used: distinguishing coherent surface waves from white noise, with models evaluated for consistency in probability outputs across diverse signal patterns.

Experimental results

Research questions

• RQ1 What hyperparameter configurations yield the most accurate and generalizable deep learning models for classifying seismic energy in real DAS data?
• RQ2 How does GPU scaling affect training throughput and efficiency for large-scale geophysical deep learning workloads?
• RQ3 Can models trained only on clean signal patterns (coherent waves and white noise) reliably generalize to predict usable energy in complex, non-labeled signal regions?
• RQ4 How can model interpretability be assessed when ground-truth labels are unavailable for non-coherent and saturated signal patterns?
• RQ5 What is the optimal balance between model depth, learning rate, batch size, and data representation for high-accuracy seismic signal classification?

Key findings

• Only four out of 1,440 trained models produced output probabilities that matched the expected ground-truth ranges for all four reference signal types, indicating high reliability and interpretability.
• The best-performing models were a 14-hidden-layer ResNet with a learning rate of 0.001, batch size of 128, and trained on a 1,000-image subset, consistently outputting probabilities of 0.9163 for coherent waves and 0.0237 for white noise after 40 epochs.
• For non-coherent waves, the best models predicted probabilities of 0.8592–0.8628, which fell within the expected range of 0.7–0.9.
• For saturated signals, the best models predicted probabilities of 0.2048–0.2975, aligning with the expected 0.2–0.3 range.
• Training throughput increased by more than two orders of magnitude when scaling from 16 to 32 GPUs on a 50,000-image dataset, demonstrating strong GPU scaling efficiency.
• The study identified that GPU-based training significantly outperformed CPU-based parallelization in speed and scalability for large-scale DAS data.