| Abstract Submission No. | ABS-4.1-0089 |
| Title of Abstract | Neuromorphic LSTM-ResUNet Hybrid For Enhanced Denoising And Inversion Of Gravity Gradiometry Data |
| Authors | Kanishk Krishnan* |
| Organisation | Indian Institute of Technology (Indian School of Mines) Dhanbad |
| Address | Harsh Shanti, 68B Patna, Bihar, India Pincode: 801503 E-mail: 21je0448@iitism.ac.in |
| Country | India |
| Presentation | Oral |
| Abstract | Gravity gradiometry provides high-resolution measurements of subsurface density variations, supporting applications in mineral exploration, hydrocarbon prospecting, and tectonic studies. However, the inversion of gravity gradiometry data is hindered by measurement noise, spatial correlations, and the ill-posed nature of the problem. Recent advances in deep learning, particularly LSTM and U-Net architectures, opens possibility for better denoising and structural inversion, yet their implementation on conventional von Neumann architectures faces significant inefficiencies. LSTMs require sequential gate computations with repeated access to large weight matrices and hidden states, creating frequent memory�processor transfers. This bottleneck reduces throughput and increases energy demand, especially for high-dimensional or real-time datasets. We propose a neuromorphic-aware hybrid framework that integrates an LSTM�ResUNet architecture with memristor-based in-memory computing simulation. The novelty lies in explicitly modeling memristor device-level non-idealities nonlinearity, stochasticity, drift, and quantization while exploiting in-memory computing to bypass the von Neumann bottleneck. This design reduces memory traffic, enables parallelized gate computations, and improves energy efficiency, allowing joint evaluation of inversion quality and system performance. In conclusion, by combining hardware-enhanced in-memory computing with deep learning, the proposed approach provides accurate, energy-efficient, and scalable solutions for gravity gradiometry processing. |
| Keywords | Gravity Gradiometry, LSTM, Memristor Computing, Neuromorphic computing, Residual U-Net |