A physics-encoded transformer model cut three-component borehole transient electromagnetic forward modeling to milliseconds, delivering a 30,080-fold speedup over conventional finite-volume simulations while keeping a pooled R² of 0.9846 ± 0.0012 across 400 test cases, researchers reported Thursday in Scientific Reports.
The surrogate model takes a three-dimensional conductivity volume as input and predicts response nodes across all stations, components, and time gates. It compresses the conductivity structure with a 3D convolutional encoder and uses a multi-layer, multi-head self-attention transformer to capture interactions along the borehole, across components, and across time gates. Physical descriptors explicitly encode station position, observation time, dimensionless time, component attributes, and sampling indices.
Training data came from batch 3D finite-volume simulations over homogeneous conducting backgrounds, layered media, and typical mine anomalous bodies. Seven baseline and ablation configurations were trained with three random seeds under a common protocol. The pooled score is strongly influenced by the early-time axial response. Component-wise scores, decay curves, and zero-crossing diagnostics characterize agreement with the finite-volume reference for the weaker transverse responses.
Single-inference latency drops to milliseconds, a model-only online speedup of 30,080-fold that excludes preprocessing, database generation, and training.
In a separate development, an international team from Germany, Thailand, India, China, and the United States created a three-component Hubbard system with tunable interactions in the laboratory for the first time. They used a quantum gas microscope to make individual atoms visible. Peter Schauss, who conducts research at the University of Hamburg and in the Cluster of Excellence “CUI: Advanced Imaging of Matter”, led the team.
A third component makes the system significantly more versatile, letting researchers investigate richer quantum mechanical phases.
“Although the Hubbard model is important and simple to write down, it isn’t easy to solve. Theorists have been banging our heads on the two-component Hubbard model for decades, yet there are still major open questions,” said Kaden Hazzard, a Rice University associate professor of physics and astronomy and co-author on the work. “The experiments are a wonderful testing ground for the theoretical ideas.”
The researchers used a quantum gas microscope, which detects each individual atom with spatial resolution. By using atoms in different internal states, the experiment realized various types of fermions, allowing the team to ultimately use three distinct states. This results in a system that, while very different from the electrons on a lattice known from solid-state physics, resembles nuclear physics, where quarks can have three different colors.
A physics-guided deep learning framework has been proposed for rapid one-dimensional transient electromagnetic inversion with uncertainty quantification. The study, authored by Zhao Le, Kaifeng Wu, Haitao Yan, Yonglong Yang, Hui Ye, Wentao Zhang, Shaobiao Liang, and Depu Hu from CCCC Second Highway Consultants Co., Ltd. in Wuhan, Hubei, China, was published in Frontiers in Artificial Intelligence and Applications on October 28, 2026.
Transient electromagnetic methods are widely used in groundwater, mineral exploration, and engineering geophysics; however, conventional 1D TEM inversion is computationally expensive and highly non-unique. The proposed framework integrates forward modeling constraints, data-driven initialization, and regularized inversion to achieve both high efficiency and physical consistency. A large synthetic dataset was generated using a rigorous 1D layered-earth TEM forward operator based on frequency-domain impedance recursion and Gaver-Stehfest numerical Laplace inversion.
A feedforward neural network is trained to predict initial resistivity and thickness models from logarithmic transient responses. The network output is further refined through a Tikhonov-regularized physics-based inversion. Experimental results show the hybrid approach improves inversion stability and interpretability while cutting computational cost by more than ten times compared with conventional methods. The framework also enables preliminary uncertainty quantification via ensemble-based prediction.