MUMBAI, India, July 24 -- Intellectual Property India has published a patent application (202641084947 A) filed by National Institute Of Technology-Warangal on July 10, 2026, for System And Method For Adaptive Neural Architecture And Operator-Rank Optimization In Physics-Informed Deep Operator Networks.
Inventors include Sarthak Sharma; D. Srinivasachary A; and Manjubala Bisi.
The application for the patent was published on July 17, 2026, under issue no. 29/2026.
Abstract: TITLE: SYSTEM AND METHOD FOR ADAPTIVE NEURAL ARCHITECTURE AND OPERATOR-RANK OPTIMIZATION IN PHYSICS-INFORMED DEEP OPERATOR NETWORKS The present invention discloses a physics-informed neural operator system and method for solving forward and inverse problems governed by partial differential equations through automated neural architecture and rank optimization as shown in figure-1. The system comprises branch and trunk subnetworks, a structural parameter controller, and differentiable gating modules that generate neuron, skip, and operator-rank gates using sigmoid-based transformations. Structured neuron masking, adaptive residual connections, and rank-controlled branch–trunk basis modulation enable dynamic adjustment of network width, depth, and operator rank during training. A physics-informed loss function enforces governing equations, boundary conditions, initial conditions, and data consistency. An alternating bi-level optimization strategy updates network weights in an inner loop while adapting structural parameters in an outer loop after a warm-up phase, with gate regularization ensuring stable convergence. Post-training thresholding yields a compact neural operator model with reduced redundancy. The framework enables simultaneous solution reconstruction and physical parameter estimation under sparse and noisy data, achieving improved accuracy, robustness, and computational efficiency across forward and inverse PDE learning tasks.
Disclaimer: Curated by HT Syndication.
