Interpretable Deep Learning Paradigm for Airborne Transient Electromagnetic Inversion

TL;DR

The paper tackles ATEM inversion under substantial noise by introducing a unified, interpretable deep-learning framework built on disentangled representation learning that separates observations into a signal factor $Z_s$ and a noise factor $Z_n$, and performs inversion using $Z_s$ under physics-guided constraints. It combines an encoder $E$, an inversion decoder $G_r$, and a data decoder $G_s$ with a mutual information estimator $CLUB$ to enforce complete disentanglement, while swapping factors during training validates the integrity of the separation and incorporates forward-model-based losses $L_{physic}$ via $F(m)$. The approach demonstrates accurate inversions on synthetic data from the resistivity model database (RMD) and robust handling of field data with significant environmental noise, yielding improved lateral resolution compared to traditional regularized inversion. Field tests on USGS AeroTEM data at Leach Lake Basin show the method processes all data points, producing coherent, high-resolution subsurface images that identify faults and stratigraphic features, closely aligning with, and in some cases enriching, conventional interpretations.

Abstract

The extraction of geoelectric structural information from airborne transient electromagnetic(ATEM)data primarily involves data processing and inversion. Conventional methods rely on empirical parameter selection, making it difficult to process complex field data with high noise levels. Additionally, inversion computations are time consuming and often suffer from multiple local minima. Existing deep learning-based approaches separate the data processing steps, where independently trained denoising networks struggle to ensure the reliability of subsequent inversions. Moreover, end to end networks lack interpretability. To address these issues, we propose a unified and interpretable deep learning inversion paradigm based on disentangled representation learning. The network explicitly decomposes noisy data into noise and signal factors, completing the entire data processing workflow based on the signal factors while incorporating physical information for guidance. This approach enhances the network's reliability and interpretability. The inversion results on field data demonstrate that our method can directly use noisy data to accurately reconstruct the subsurface electrical structure. Furthermore, it effectively processes data severely affected by environmental noise, which traditional methods struggle with, yielding improved lateral structural resolution.

Figures (10)

• Figure 1: The overall framework, the data is encoded by the encoder $E$ into signal and noise factors. The inversion decoder $G_r$ decodes the signal factor to obtain accurate inversion results. $G_s$ is used to decode the combination of signal and noise factors into the signal, ensuring the correct disentangled representation. The loss function incorporates physical information to impose physical constraints.
• Figure 2: (a) represents the architecture of the encoder E, (b) represents the architectures of the inversion decoder Gr and the data decoder Gs, and (c) shows the stacked block structure within the encoder and decoder.
• Figure 3: The 30-layer RMD model (on the left), forward response (in the middle), and the data after noise addition (on the right).
• Figure 4: (a) Inversion results using clean signals, (b) Inversion results using noisy signals, (c) Forward modeling response results using the inversion model from (a), (d) Forward modeling response results using the inversion model from (b).
• Figure 5: (a) RMSPE statistics of the inversion results for clean signals across the entire test set, (b) RMSPE statistics computed from the forward modeling response of the inversion results in (a) and the clean signals, (c) RMSPE statistics of the inversion results for noisy signals across the entire test set, (d) RMSPE statistics computed from the forward modeling response of the inversion results in (c) and the clean signals.