Parameter-Efficient Transfer Learning for Microseismic Phase Picking Using a Neural Operator
Authors: Ayrat Abdullin, Umair Bin Waheed, Leo Eisner, Naveed Iqbal
Organizations: Department of Geosciences, King Fahd University of Petroleum and Minerals, Dhahran, 31261, Saudi Arabia · Seismik s.r.o., Prague, 18200, Czech Republic · Department of Electrical Engineering, King Fahd University of Petroleum and Minerals, Dhahran, 31261, Saudi Arabia
Seismic phase picking is fundamental for microseismic monitoring and subsurface imaging. Manual processing is impractical for real-time applications and large sensor arrays, motivating the use of deep learning-based pickers trained on extensive earthquake catalogs. On a broader scale, these models are generally tuned to perform optimally in high signal-to-noise and long-duration networks and often fail to perform satisfactorily when applied to campaign-based microseismic datasets, which are characterized by low signal-to-noise ratios, sparse geometries, and limited labeled data. In this study, we present a microseismic adaptation of a network-wide earthquake phase picker, Phase Neural Operator (PhaseNO), using transfer learning and parameter-efficient fine-tuning. Starting from a model pre-trained on more than 57,000 three-component earthquake and noise records, we fine-tune it using only 200 labeled and noisy microseismic recordings from hydraulic fracturing settings. We present a parameter-efficient adaptation of PhaseNO that fine-tunes a small fraction of its parameters (only 3.6%) while retaining its global spatiotemporal representations learned from a large dataset of earthquake recordings. We then evaluate our adapted model on three independent microseismic datasets and compare its performance against the original pre-trained PhaseNO, a STA/LTA-based workflow, and two state-of-the-art deep learning models, PhaseNet and EQTransformer. We demonstrate that our adapted model significantly outperforms the original PhaseNO in F1 and accuracy metrics, achieving up to 30% absolute improvements in all test sets and consistently performing better than STA/LTA and state-of-the-art models. With our adaptation being based on a small calibration set, our proposed workflow is a practical and efficient tool to deploy network-wide models in data-limited microseismic applications.
Figures & tables
Fig. 1: The architecture of the Phase Neural Operator. The network is built from multiple layers of Fourier Neural Operators (FNO) and Graph Neural Operators (GNO) arranged sequentially and iteratively. The up and down projections, denoted by P and Q, respectively, are implemented using neural networks. The model processes seismograms collected from a seismic network with arbitrarily located stations and predicts the P-phase and S-phase arrival time probabilities for each station. In addition to the three channels containing the three-component waveforms, the station locations are encoded as two separate channels, and the relative coordinates ( xi , yi ) between stations are used to learn the edge weights in the graph (adapted after Sun et al. (2023) [ 20 ] ).
Fig. 2: The maps illustrate the locations of datasets D1 (a) and D2 (b) with seismic stations and induced event epicenters. Network stations are marked with triangles. The events occur at an average depth of 2.1 km, ranging from 1.7 to 2.4 km (D1), and 3.1 km, ranging from 2.8 to 3.4 km (D2).
Parameter
Value
Number of FNO layers
8
Number of GNO layers
6
Base hidden dimension ( w )
48
Base Fourier modes ( m )
24
Hidden dimension multipliers
1×,2×,4×,8× (U-shaped)
Fourier modes multipliers
1×,1/2×,1/3×
TABLE I: \chadded Key hyperparameters of the inherited base PhaseNO architecture [ 20 ] .
Fig. 3: Evolution of the average P– and S–phase F1 score during training. Shown are the training and validation F1 curves over 100 epochs, with the validation performance peaking at an F1 score of 0.906 at epoch 55.
Fig. 4: Impact of progressive unfreezing of PhaseNO layers on validation F1 score as a function of the number of trainable parameters. Blue markers show the mean validation F1 ( ±1 SD) over five independent fine-tuning runs for the original PhaseNO backbone with progressively unfrozen FNO/GNO blocks. Full-network fine-tuning (¿2.0M trainable parameters) approaches a validation F1 of 0.90. The green star denotes the proposed LoRA+SE (CR-LoRA) adaptation, which achieves a validation F1 of 0.889±0.013 using only 85K trainable parameters (3.6% of the full model). Despite its parameter efficiency, the LoRA+SE model attains performance comparable to extensive backbone unfreezing. For comparison, PhaseNet (red circle; 268K parameters) yields 0.832±0.022 , whereas EQTransformer (EQT) (magenta circle; 377K parameters) achieves 0.629±0.084 under similar fine-tuning conditions.
Fig. 5: A comparison of performance between the original PhaseNO and the fine-tuned model for P and S arrival picks for the three test sets. We use precision, recall, F1, and accuracy scores to quantify and compare the performance.
Processing Stage
Parameter Settings
1. Pre-Processing
Filter Type: 4th-order Butterworth bandpass
(All components)
Frequency Band: 2.0 – 38.0 Hz
2. P-Wave Picking
Component: Vertical (Z)
(STA/LTA + AIC)
STA Window: 0.5 s
LTA Window: 5.0 s
Trigger Threshold (ON): 3.0
TABLE II: \chadded Parameter configuration for the conventional picking baseline.
Fig. 6: A comparison of performance between the conventional STA/LTA method and the fine-tuned model for P and S arrival picks for the three test sets. We use precision, recall, F1, and accuracy scores to quantify and compare the performance.
P-Arrival
S-Arrival
Model
Pr
Re
F1
Acc
Pr
Re
F1
Acc
Test Set 1
Original
0.68
0.88
0.77
0.77
0.76
0.92
0.83
0.84
Fine-tuned
0.89
0.97
0.93
0.93
0.88
0.97
0.92
0.93
Conventional
0.63
0.78
0.69
0.72
0.34
0.43
0.38
0.49
PhaseNet
0.89
0.28
0.43
0.66
0.88
0.30
0.45
0.67
TABLE III: Performance comparison of seismic phase picking models across three test sets. Metrics for P-arrival and S-arrival include Precision (Pr), Recall (Re), F1-score (F1), and Accuracy (Acc). The original and fine-tuned models are compared against Conventional methods, PhaseNet, and EQ Transformer.
Model Variant
F0
F1
F2
F3
F4
Mean
SD
CR-LoRA
0.884
0.891
0.884
0.896
0.884
0.888
0.005
LoRA-only
0.890
0.881
0.897
0.869
0.883
0.884
0.009
SE-only
0.845
0.849
0.842
0.812
0.831
0.836
0.013
TABLE IV: \chadded Ablation study: 5-fold cross-validation F1 scores. The best fold for each variant (used for test evaluation) is shown in bold.
Model Variant
D1
D2
D3
CR-LoRA
0.929
0.930
0.594
LoRA-only
0.929
0.928
0.589
SE-only
0.910
0.870
0.537
TABLE V: \chadded Ablation study: test-set F1 scores (mean of P-wave and S-wave F1) for the best checkpoint of each variant. Bold values indicate the best score for each test set.
Fig. 7: Example waveforms from test set #1 demonstrate the probability results for the original (red and blue graphs) and fine-tuned (orange and cyan graphs) PhaseNO models with different signal-to-noise ratios (SNR). The red and orange graphs represent the P-pick probabilities, whereas the blue and cyan graphs represent the S-pick probabilities. The green and magenta vertical bars represent the analyst picks for the P and S waves, respectively. The results for the fine-tuned model are more accurate and closer to the analyst picks for low SNR values, thereby highlighting the robustness and accuracy achieved through the fine-tuning process.
Fig. 8: Comparison of differences in arrival times for picks (top row) and the calculated prediction uncertainty based on the probability function of each pick (bottom row) on test set #1 of two deep learning models: the original PhaseNO (left) and the fine-tuned PhaseNO (right). The top row of plots shows a histogram of time residuals of P-wave picks (red) and S-wave picks (blue), i.e., differences between model predictions and analyst picks. Each plot provides information about the mean ( μ ), standard deviation ( σ ), and mean absolute error (MAE) of all true positive picks with a time residual below 0.5 s. Note that the fine-tuned model has the mean residual closer to zero, and its average uncertainty is roughly three times smaller compared to the original PhaseNO, showing improved overall picking accuracy.
Fig. 9: Comparison of picking arrival time differences (top row) and predicted uncertainty computed from the probability function of a corresponding pick (bottom row) on the test set #2 for two deep-learning models: the original PhaseNO (left) and the fine-tuned PhaseNO (right). The top-row histograms show time residuals for P-wave picks (red) and S-wave picks (blue), defined as the difference between the model-predicted arrivals and the analyst picks. Each panel reports the mean ( μ ), standard deviation ( σ ), and mean absolute error (MAE) for all true-positive (TP) picks with time residuals below 0.5 s. The bottom-row histograms display the distribution of 1-sigma probability intervals for TPs in each model. Similar to the D1 dataset performance, the fine-tuned model’s mean residual is closer to zero, and its average uncertainty interval is about three times smaller than that of the original PhaseNO, indicating improved overall picking accuracy.
Fig. 10: Sensitivity of the overall F1 score with respect to the true-positive time-tolerance threshold for both the original and fine-tuned models over the three test sets. \chadded The evaluation ranges from a strict tolerance of 0.1 s to a relaxed 1.0 s, evaluated in 0.1 s increments. For all test data, it is evident that the fine-tuned model significantly outperforms the original PhaseNO over the entire range \chreplaced (from 0.1 to 1.0 seconds). Although increasing the tolerance gradually increases the F1 score for both models, the fine-tuned model performs significantly better \chadded even under the strictest 0.1 s criterion.
Numerous studies have shown that the machine-learning picker PhaseNet produces accurate P and S picks on local earthquake signals, but its performance can degrade sharply on teleseismic signals. To address this limitation, we present a reproducible MsPASS workflow that (i) enables scalable data preparation and management for large seismic archives and (ii) supports standardized PhaseNet training and inference. We assembled a control dataset of 1.6 million waveforms linked to teleseismic P-wave picks made by analysts at the USArray Array Network Facility (ANF). The control dataset confirms that the PhaseNet model trained on regional signals performs poorly on these data. We then trained PhaseNet from scratch on the training split of the ANF control dataset and evaluated it on a non-overlapping held-out test split, increasing P-pick recall by 741.5% and yielding 683.9% more picks within a 0.1s residual window. We also evaluated PhaseNet across different model sizes on both CPUs and GPUs. Increasing the model size by about 120 times improved precision and recall by 15.6% and 23.2%, respectively. However, the scaled model reduced inference throughput by 87.2% on an NVIDIA A100 GPU and by 97.3% on a 128-core high-performance CPU node. These results indicate that scaling PhaseNet is more practical on GPUs than on CPUs, and that simply enlarging the model is not an efficient way to achieve large accuracy gains.
Jinxin Ma, Yinzhi Wang, Gary L. Pavlis +1
Texas Advanced Computing Center, The University of Texas at Austin · Department of Earth and Atmospheric Sciences, Indiana University, Bloomington, IN 47405
Earthquake detection and seismic phase picking are fundamental yet challenging tasks in seismology due to low signal-to-noise ratios, waveform variability, and overlapping events. Recent deep-learning models achieve strong results but rely on large datasets and heavy backpropagation training, raising concerns over efficiency, interpretability, and sustainability. We propose GreenPhase, a multi-resolution, feed-forward, and mathematically interpretable model based on the Green Learning framework. GreenPhase comprises three resolution levels, each integrating unsupervised representation learning, supervised feature learning, and decision learning. Its feed-forward design eliminates backpropagation, enabling independent module optimization with stable training and clear interpretability. Predictions are refined from coarse to fine resolutions while computation is restricted to candidate regions. On the Stanford Earthquake Dataset (STEAD), GreenPhase achieves excellent performance with F1 scores of 1.0 for detection, 0.98 for P-wave picking, and 0.96 for S-wave picking. This is accomplished while reducing the computational cost (FLOPs) for inference by approximately 83% compared to state-of-the-art models. These results demonstrate that the proposed model provides an efficient, interpretable, and sustainable alternative for large-scale seismic monitoring.
Yixing Wu, Shiou-Ya Wang, Dingyi Nie +5
Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California · Center for Earth and Environmental Studies, National Central University · Department of Computer Science & Information Engineering, National Central University
Inaccurately labeled training data, or "label noise", poses a significant threat to the integrity of supervised machine learning models. This corruption directly degrades performance by teaching the model erroneous mappings between features and labels, which leads to poor generalization and reduced accuracy on properly labeled validation and test data. Current seismological applications mainly rely on large-scale training sets or data augmentation to reduce the label-noise impact, which can be labor-intensive and costly. Here, we introduce a Label Noise-Contrastive Robust Learning (LaNCoR) approach that can effectively handle noisy labels in seismic signal processing tasks, without requiring large-scale training datasets. In this approach, the input waveform feature and label representation distributions are aligned in the feature space to correct mislabeling and reduce its impact on the training process. We present LaNCoR's performance on the task of P-phase arrival-time picking of real microseismic data using two baseline models and training approaches. Our results indicate that LaNCoR can improve performance by up to 28.8% across performance metrics. This approach holds great promise for model training in seismology and geosciences.
Sen Li, Xu Yang, S. Mostafa Mousavi +5
Department of Earth and Planetary Sciences, Harvard University, Cambridge, MA, USA. · School of Computer Science and Technology, China University of Mining and Technology, Xuzhou, Jiangsu, China · School of Mines, China University of Mining and Technology, Xuzhou, Jiangsu, China +1