Deep learning for distributed acoustic sensing-based microseismic candidate-window classification in enhanced geothermal systems: A comparative study of twelve architectures

Microseismic event discrimination in distributed acoustic sensing data is critical for monitoring enhanced geothermal system operations, yet classical detectors based on short-term to long-term energy ratios, in the simple validation-tuned form evaluated here, fail to discriminate genuine seismic arrivals from structured noise. This paper presents a comparative study of twelve artificial intelligence models based on deep learning, spanning five architecture families—convolutional neural networks, transformers, recurrent networks, state-space models, and graph neural networks—benchmarked against three classical baselines on 3974 manually labeled distributed acoustic sensing recordings from a field stimulation campaign. We contribute an open-source labeling tool for multichannel seismic data, and evaluate complete 1.2-second triggered candidate-window classification rather than end-to-end continuous-stream detection. A preliminary temporal-context ablation is reported but not treated as a repeated-seed conclusion. The squeeze-and-excitation residual network and the plain residual network are statistically comparable; the squeeze-and-excitation variant has the highest numerical F1-score (the harmonic mean of precision and recall), 98.50%, and an area under the receiver operating characteristic curve of 99.46%, while the plain residual network has slightly stronger balanced accuracy and noise recall under the exported reviewer metrics. A selective state-space model achieves 95.76% F1-score, providing a competitive baseline for this architecture family in this dataset. A graph neural network exploiting fiber spatial topology achieves 95.47% F1-score with only 94,290 parameters. Gradient-weighted class activation mapping provides qualitative evidence that the convolutional models focus on primary-wave moveout patterns. To stress-test external transfer, all twelve saved Stage 14 models are evaluated without fine-tuning on Utah Frontier Observatory for Research in Geothermal Energy April 2024 data, a different acquisition. The external set contains 4073 candidate windows (2024 event windows and 2049 non-event windows). The native records have 1496 traces and 4800 samples at 4000 samples per second; the same preprocessing function was used for event and non-event windows: the first 361 traces were retained, the records were downsampled to 2400 samples at 2000 samples per second using anti-aliased polyphase resampling, and per-channel z-score (zero mean, unit variance) normalization was applied. The best external model is the convolutional neural network–gated recurrent unit hybrid with F1-score 0.9930, balanced accuracy 0.9931, and Matthews correlation coefficient 0.9863. We interpret this as zero-shot evidence of matched-pipeline external robustness, not as universal cross-site validation or deployment-ready generalization. A three-seed retraining of the top two residual convolutional models with full random-state control further confirms that the small numerical F1 advantage of SE-ResNet over ResNet is within initialization variance: the pooled inter-seed standard deviation ( σ ≈ 0.0036 F1) is approximately four times the inter-model mean difference (0.0009 F1), and paired t -tests and Wilcoxon signed-rank tests fail to reject the null of no difference on F1, balanced accuracy, and MCC.