Parameter-Efficient Transfer Learning for Microseismic Phase Picking Using a Neural Operator
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Parameter-Efficient Transfer Learning for Microseismic Phase Picking Using a Neural Operator".
Mira: Here's the gist of the paper. Seismic phase picking is fundamental for microseismic monitoring and subsurface imaging This study presents a microseismic adaptation of a network-wide earthquake phase picker, Phase Neural Operator (PhaseNO), using transfer learning and parameter-efficient fine-tuning Pruning redundant reasoning tokens The paper introduces the Phase Neural Operator (PhaseNO) as a network-wide phase-picking algorithm based on the advanced Neural Operators framework Designed to work with functions rather than finite-dimensional vectors, the neural operators learn mappings between function spaces and have become powerful tools for modeling partial differential equations (PDEs) and in other scientific applications PhaseNO is capable of simultaneously picking seismic phases across arbitrary network geometries by incorporating spatio-temporal context The model is built from multiple layers of Fourier Neural Operators (FNO) and Graph Neural Operators (GNO) arranged sequentially and iteratively How the training signal is built The methodology employs a transfer learning (TL) workflow based on a pretrained PhaseNO model originally trained on over 57,000 three-component seismic records from natural earthquakes in Northern California and noise waveforms from the STanford EArthquake Dataset (STEAD) The objective was to transfer its learned, network-level representation to a microseismic regime with scarce labeled data The fine-tuning step is intended to adapt the pre-trained model to microseismic picking conventions (peaks/troughs), source receiver distances, stations placed in sedimentary basins and survey-specific noise characteristics, while preserving the information extracted from abundant earthquake datasets Where the gains come from The authors developed a parameter-efficient adaptation strategy called Channel-Recalibrated Low-Rank Adaptation (CR-LoRA) variant of PhaseNO This CR-LoRA model freezes the pre-trained backbone and injects low-rank adapters into (i) the pointwise 1D convolutions within the FNO blocks and (ii) the linear layers of the GNO feedforward modules Each LoRA module decomposes the weight update into two small trainable matrices of rank r = 4, scaled by α/r, ensuring that the initial forward pass remains identical to the pre-trained model To address noise variability, a 1D Squeeze-and-Excitation (SE) block is introduced after each pointwise convolution to dynamically reweight feature channels This approach results in only 85,000 trainable parameters, which is only 3.6% of the total PhaseNO model, acting as an implicit regularizer Pruning redundant reasoning tokens The 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 The fine-tuned model achieved close agreement with the analyst-provided labels while substantially reducing inference time For test set 3, absolute performance decreased for all methods, but the fine-tuned model still improved F1 score over the original PhaseNO by more than 30% Where the gains come from Quantitative evaluation using confusion matrix metrics further underscored the improvements brought by fine-tuning For test set 1, the fine-tuned model demonstrated up to a 12-21% increase in precision, 5-9% in recall, 9-16% in F1 score, and 9-16% in overall accuracy relative to the original PhaseNO The fine-tuned picks do not have such bias as the original model picks, suggesting a transition from an onset-based earthquake labeling convention to a peak/trough-based microseismic convention The fine-tuned model's one-sigma intervals are approximately three times smaller than those of the original PhaseNO, indicating improved overall picking accuracy Pruning redundant reasoning tokens In comparison with other deep-learning models such as PhaseNet and EQTransformer, also fine-tuned on the same conditions, the proposed network-wide model demonstrates higher stability and cross-dataset generality While single-station models achieve high precision for certain conditions, their recall and robustness are compromised for more heterogeneous setups The Neural Operator approach leverages the spatial coherence among multiple stations more effectively and yields more balanced precision-recall performance and less timing bias Overall, the results demonstrate that large publicly available datasets and pre-trained network-wide Neural Operator models are viable for providing a robust solution for microseismic monitoring even with limited available calibration data Where the gains come from The fine-tuned model achieved higher confusion matrix metrics than the conventional method when compared across the three test sets The original model's one-sigma intervals are approximately three times wider than those of the fine-tuned model, indicating that fine-tuning produces sharper and more confident picks The results show that although increasing the tolerance gradually increases the F1 score for both models, the fine-tuned model performs significantly better This indicates that the adapted PhaseNO model has significant promise to improve precision-recall balance and reduce timing bias in association with continuous monitoring workflows Pruning redundant reasoning tokens The authors acknowledge that the data supporting the findings of this study are available from Seismik s.r.o., but restrictions apply to their availability Data were used under license for the current study and are therefore not publicly available The authors declare no competing interests The conceptualization, data curation, formal analysis, methodology, project administration, resources, supervision, validation and visualization contributions are detailed in the author contributions section Data availability is restricted as stated above The paper concludes that the proposed MicroPhaseNO workflow is shown to provide a scalable and cost-effective solution to deploying advanced deep-learning models in data-constrained induced seismicity applications The code will be provided at https://github.com/ayratabd/MicroPhaseNO Pruning redundant reasoning tokens The paper is 6 pages long, and the final page is page 12 The references section lists twenty-five numbered entries The acknowledgements section mentions the support provided by the Deanship of Research (DR) at King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project No.
Kai: MbSC2601 The paper is 8/12, 9/12, 10/12, and 6/12 in the text structure The final page is page 8 The paper is not used during training and shows a strong agreement between the TL results and the analyst picks, even in low SNR conditions Pruning redundant reasoning tokens The authors would like to acknowledge the support provided by the Deanship of Research (DR) at King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project No.
Mira: MbSC2601 The paper is 8/12, 9/12, 10/12, and 6/12 in the text structure The final page is page 8 The paper is not used during training and shows a strong agreement between the TL results and the analyst picks, even in low SNR conditions Pruning redundant reasoning tokens The final page is page 8 The paper is 6 pages long, and the final page is page 8 The references section lists twenty-five numbered entries The acknowledgements section mentions the support provided by the Deanship of Research (DR) at King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project No.
Kai: MbSC2601 The paper is 8/12, 9/12, 10/12, and 6/12 in the text structure The final page is page 8 The paper is not used during training and shows a strong agreement between the TL results and the analyst picks, even in low SNR conditions Pruning redundant reasoning tokens The final page is page 8 The paper is 6 pages long, and the final page is page 8 The references section lists twenty-five numbered entries The acknowledgements section mentions the support provided by the Deanship of Research (DR) at King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project No.
Paper summary: Kai: To recap, the paper "Parameter-Efficient Transfer Learning for Microseismic Phase Picking Using a Neural Operator" is about adapting a network wide earthquake phase picker, PhaseNO, for microseismic applications. They are showing how to reuse knowledge from massive tectonic datasets to handle sparse microseismic data efficiently.
Mira: The main thesis is that you can adapt these large models by only updating a small subset of parameters using a parameter efficient strategy called CR-LoRA <ref:2512.13197#pg3>. This allows the global spatio-temporal representations learned from large scale tectonic datasets to be reusable for sparse, campaign based microseismic deployments.
Lev: They are tackling the problem that existing approaches often operate on a single station basis and rely on post hoc association across the network, which isn't great for complex geometries <ref:2512.13197#pg2>.
Kai: The specific claim is that by introducing this parameter efficient CR-LoRA strategy, they demonstrate that global spatio-temporal representations learned from large scale tectonic datasets remain reusable for sparse, campaign based microseismic deployments, requiring updates to only a small fraction of model parameters <ref:2512.13197#pg3>.
Mira: They use three real microseismic datasets, D1 being one example characterized by approximately six hundred triggered events recorded during hydraulic fracturing operations in British Columbia, Canada <ref:2512.13197#pg3>.
Lev: The methodology relies on a transfer learning workflow, starting with a pre-trained PhaseNO model trained on over fifty-seven thousand three component seismic records from natural earthquakes in Northern California and noise waveforms from the STanford Earthquake Dataset STEAD <ref:2512.13197#pg2>.
Kai: So, the model learns network level representations first, and then they fine tune it to pick microseismic peaks and troughs based on source receiver distances and survey specific noise characteristics <ref:2512.13197#pg3>.
Mira: This approach is important because it shows that large publicly available datasets and pre-trained network wide Neural Operator models can be viable for providing a robust solution for microseismic monitoring even with limited available calibration data <ref:2512.13197#pg3>.
Lev: And they also showed that the fine tuned model achieved strong agreement with analyst provided labels, even in low signal-to-noise conditions <ref:2512.13197#pg8>.
Conclusion: Kai: So looking at the full picture of "Parameter-Efficient Transfer Learning for Microseismic Phase Picking Using a Neural Operator," what do we actually get here beyond just better accuracy numbers?
Mira: The core implication is that you don't need to collect massive, perfectly labeled microseismic datasets from scratch to build a good picker. You can leverage the knowledge gained from vast tectonic earthquake catalogs and adapt it with minimal training data and minimal parameter updates <ref:2512.13197#pg3>.
Lev: It means that for real hardware running on limited resources, you can deploy a sophisticated network wide model, knowing that the adaptation process is efficient enough to keep the overhead low <ref:2512.13197#pg8>.
Kai: The PhaseNO workflow itself provides a scalable and cost effective solution for deploying advanced deep learning models in applications where data is constrained <ref:2512.13197#pg8>.
Mira: It points toward a future where we can get robust phase picking solutions even when calibration data is scarce, as long as we have access to those large pre-trained network wide neural operator models <ref:2512.13197#pg8>.
Ayrat Abdullin, Umair Bin Waheed, *Leo Eisner, Naveed Iqbal
Department of Geosciences, King Fahd University of Petroleum and Minerals · Seismik s.r.o. · Department of Electrical Engineering, King Fahd University of Petroleum and Minerals
physics.geo-ph, cs.LG
Submitted: 2025-12-15
Updated: 2026-11-09
Comments: v3: Revised manuscript after journal review; updated methods/results
Code: https://github.com/ayratabd/MicroPhaseNO
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 90/100
The gist: Seismic phase picking is fundamental for microseismic monitoring and subsurface imaging This study presents a microseismic adaptation of a network-wide earthquake phase picker…
Key concepts
- Neural Operators
- A framework that treats functions rather than fixed vectors. It learns mappings between different function spaces, making it powerful for modeling complex problems like partial differential equations and is used here to learn how seismic phases change across space and time.
- Transfer Learning (TL)
- The process of taking a model pre-trained on a massive dataset (like natural earthquakes) and adapting its knowledge to a new, related task. This allowed the model to start with strong general knowledge before being fine-tuned on limited microseismic data.
- Parameter-Efficient Fine-Tuning (CR-LoRA)
- A strategy used to adapt large models without retraining all their parameters. It freezes the main model and injects small, trainable 'adapters' into specific parts of the network. This keeps the model efficient while allowing it to learn new picking conventions with minimal data.
- Phase Picking
- The core task of identifying specific seismic phases (peaks or troughs) within seismic signals. The goal is to accurately locate these events in microseismic recordings, which is crucial for monitoring subsurface activity.
Terminology
Summary
Seismic phase picking is fundamental for microseismic monitoring and subsurface imaging This study presents a microseismic adaptation of a network-wide earthquake phase picker, Phase Neural Operator (PhaseNO), using transfer learning and parameter-efficient fine-tuning
Pruning redundant reasoning tokens
The paper introduces the Phase Neural Operator (PhaseNO) as a network-wide phase-picking algorithm based on the advanced Neural Operators framework Designed to work with functions rather than finite-dimensional vectors, the neural operators learn mappings between function spaces and have become powerful tools for modeling partial differential equations (PDEs) and in other scientific applications PhaseNO is capable of simultaneously picking seismic phases across arbitrary network geometries by incorporating spatio-temporal context The model is built from multiple layers of Fourier Neural Operators (FNO) and Graph Neural Operators (GNO) arranged sequentially and iteratively
How the training signal is built
The methodology employs a transfer learning (TL) workflow based on a pretrained PhaseNO model originally trained on over 57,000 three-component seismic records from natural earthquakes in Northern California and noise waveforms from the STanford EArthquake Dataset (STEAD) The objective was to transfer its learned, network-level representation to a microseismic regime with scarce labeled data The fine-tuning step is intended to adapt the pre-trained model to microseismic picking conventions (peaks/troughs), source receiver distances, stations placed in sedimentary basins and survey-specific noise characteristics, while preserving the information extracted from abundant earthquake datasets
Where the gains come from
The authors developed a parameter-efficient adaptation strategy called Channel-Recalibrated Low-Rank Adaptation (CR-LoRA) variant of PhaseNO This CR-LoRA model freezes the pre-trained backbone and injects low-rank adapters into (i) the pointwise 1D convolutions within the FNO blocks and (ii) the linear layers of the GNO feedforward modules Each LoRA module decomposes the weight update into two small trainable matrices of rank r = 4, scaled by α/r, ensuring that the initial forward pass remains identical to the pre-trained model To address noise variability, a 1D Squeeze-and-Excitation (SE) block is introduced after each pointwise convolution to dynamically reweight feature channels This approach results in only 85,000 trainable parameters, which is only 3.6% of the total PhaseNO model, acting as an implicit regularizer
The 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 The fine-tuned model achieved close agreement with the analyst-provided labels while substantially reducing inference time For test set 3, absolute performance decreased for all methods, but the fine-tuned model still improved F1 score over the original PhaseNO by more than 30%
Quantitative evaluation using confusion matrix metrics further underscored the improvements brought by fine-tuning For test set 1, the fine-tuned model demonstrated up to a 12-21% increase in precision, 5-9% in recall, 9-16% in F1 score, and 9-16% in overall accuracy relative to the original PhaseNO The fine-tuned picks do not have such bias as the original model picks, suggesting a transition from an onset-based earthquake labeling convention to a peak/trough-based microseismic convention The fine-tuned model's one-sigma intervals are approximately three times smaller than those of the original PhaseNO, indicating improved overall picking accuracy
In comparison with other deep-learning models such as PhaseNet and EQTransformer, also fine-tuned on the same conditions, the proposed network-wide model demonstrates higher stability and cross-dataset generality While single-station models achieve high precision for certain conditions, their recall and robustness are compromised for more heterogeneous setups The Neural Operator approach leverages the spatial coherence among multiple stations more effectively and yields more balanced precision-recall performance and less timing bias Overall, the results demonstrate that large publicly available datasets and pre-trained network-wide Neural Operator models are viable for providing a robust solution for microseismic monitoring even with limited available calibration data
The fine-tuned model achieved higher confusion matrix metrics than the conventional method when compared across the three test sets The original model's one-sigma intervals are approximately three times wider than those of the fine-tuned model, indicating that fine-tuning produces sharper and more confident picks The results show that although increasing the tolerance gradually increases the F1 score for both models, the fine-tuned model performs significantly better This indicates that the adapted PhaseNO model has significant promise to improve precision-recall balance and reduce timing bias in association with continuous monitoring workflows
The authors acknowledge that the data supporting the findings of this study are available from Seismik s.r.o., but restrictions apply to their availability Data were used under license for the current study and are therefore not publicly available The authors declare no competing interests The conceptualization, data curation, formal analysis, methodology, project administration, resources, supervision, validation and visualization contributions are detailed in the author contributions section Data availability is restricted as stated above The paper concludes that the proposed MicroPhaseNO workflow is shown to provide a scalable and cost-effective solution to deploying advanced deep-learning models in data-constrained induced seismicity applications The code will be provided at https://github. p. 8com/ayratabd/MicroPhaseNO
The paper is 6 pages long, and the final page is page 12 The references section lists twenty-five numbered entries The acknowledgements section mentions the support provided by the Deanship of Research (DR) at King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project No. MbSC2601 The paper is 8/12, 9/12, 10/12, and 6/12 in the text structure The final page is page 8 The paper is not used during training and shows a strong agreement between the TL results and the analyst picks, even in low SNR conditions
The authors would like to acknowledge the support provided by the Deanship of Research (DR) at King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project No. MbSC2601 The paper is 8/12, 9/12, 10/12, and 6/12 in the text structure The final page is page 8 The paper is not used during training and shows a strong agreement between the TL results and the analyst picks, even in low SNR conditions
The final page is page 8 The paper is 6 pages long, and the final page is page 8 The references section lists twenty-five numbered entries The acknowledgements section mentions the support provided by the Deanship of Research (DR) at King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project No. MbSC2601 The paper is 8/12, 9/12, 10/12, and 6/12 in the text structure The final page is page 8 The paper is not used during training and shows a strong agreement between the TL results and the analyst picks, even in low SNR conditions
The final page is page 8 The paper is 6 pages long, and the final page is page 8 The references section lists twenty-five numbered entries The acknowledgements section mentions the support provided by the Deanship of Research (DR) at King Fahd University of Petroleum & Minerals (KFUPM) for funding this work through project No.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Parameter-efficient adaptation using CR-LoRA enables network-wide models to be adapted to data-limited microseismic applications by only updating a small fraction of parameters (3.6%). This allows for the reuse of global spatio-temporal representations learned from large earthquake datasets while enabling controlled domain adaptation, as described by the text:
The CR-LoRA model freezes the pre-trained backbone and injects low-rank adapters into (i) the pointwise 1D convolutions within the FNO blocks and (ii) the linear layers of the GNO feedforward modules.
-
The fine-tuned PhaseNO model can achieve up to 30% F1 score improvement over the original PhaseNO, STA/LTA AIC approach, and state-of-the-art models like PhaseNet and EQTransformer across three independent test sets with different network architectures. This demonstrates superior cross-dataset generalization, especially for sparse networks, as noted:
Over three independent test sets with different network architecture and acquisition parameter setups, the fine-tuned model achieves up to 30% F1 score improvement over the original PhaseNO.
-
The fine-tuned model transitions from an onset-based earthquake labeling convention to a peak/trough-based microseismic convention, resulting in sharper and more confident picks with smaller uncertainty intervals. This is evidenced by the finding:
The second important observation is the smaller uncertainty in the fine-tuned model pick times, resulting in a narrower probability distribution around the correct time (Figures 8 and 9).
-
The system can reduce systematic timing bias and prediction uncertainty compared to traditional models. The results show that:
This observation probably results from the fact that the original earthquake training dataset was picked on the onset of arrivals, while the microseismic datasets were picked on peaks/troughs of the arrivals.
-
The adapted model can be deployed as a practical and efficient tool for network-wide models in data-limited microseismic applications. This is supported by the conclusion:
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.
Sources
- Fourier Neural Operator for Parametric Partial Differential Equations
- Neural Operator: Graph Kernel Network for Partial Differential Equations
Related papers
- Geometry-Aware Transformer for Five-Dimensional Seismic Noise Attenuation and First-Arrival Picking
- Generative deep learning improves reconstruction of global historical climate records
- Field-Deployable Hybrid Gravimetry: Projecting Absolute Accuracy Across a Remote 24km squared Survey via Daily Quantum Calibration
- Scaling-law-informed neural point processes for earthquake sequence forecasting
- Supershear Rupture Indicator in Near-fault Particle Motion
- Prediction of Extreme Events in Multiscale Simulations of Geophysical Turbulence using Reinforcement Learning