MUMBAI, India, March 13 -- Intellectual Property India has published a patent application (202641024920 A) filed by Vishnu Institute Of Technology, Bhimavaram, Andhra Pradesh, on March 3, for 'a method and device based on mathematical modeling and optimization of complex systems using machine learning and artificial intelligence.'
Inventor(s) include Dr. M. Vijaya Lakshmi; Dr. R. S. Durga Rao; and Dr. K. Ramesh,.
The application for the patent was published on March 13, under issue no. 11/2026.
According to the abstract released by the Intellectual Property India: "The present invention relates to a method and device for mathematical modeling and optimization of complex dynamical and multi-physics systems by integrating state-of-the-art machine learning and artificial intelligence techniques with rigorous mathematical frameworks derived from recent scientific literature. The invention addresses critical limitations in conventional computational tools, which treat physics-based modeling and data-driven AI as separate, disconnected pipelines, resulting in physically inadmissible predictions, poor uncertainty quantification, and inability to handle real-time multi-objective optimization under dynamic operating conditions. The core of the invention is a layered, co-designed computational architecture comprising five principal components: (i) a Physics-Informed Neural Network (PINN) engine that enforces governing ordinary and partial differential equations as composite loss terms using automatic differentiation and an adaptive augmented Lagrangian weighting scheme; (ii) a Gaussian Process Regression (GPR) uncertainty quantification layer that provides calibrated Bayesian posterior predictive distributions over model outputs, enabling risk-informed decision-making; (iii) a Transformer-based temporal attention module for processing high-frequency sequential sensor data with causal self-attention mechanisms that respect physical causality; (iv) a multi-objective Reinforcement Learning optimizer based on Proximal Policy Optimization with entropy regularization and Pareto-archive replay buffers for identifying Pareto-optimal design solutions under competing engineering objectives; and (v) a federated learning coordination layer supporting privacy-preserving distributed model training across geographically separated data nodes using homomorphic encryption-based secure gradient aggregation with Byzantine fault tolerance. The device embodiment of the invention incorporates a hardware accelerator unit (FPGA or GPU-based) implementing 8-bit quantized neural network inference with pipelined dataflow architecture achieving latency below 10 milliseconds, a real-time drift detection engine using sliding-window statistical tests for automatic model retraining, and an explainability module generating SHAP-based feature attribution maps for physically interpretable model diagnostics. An online learning module with Kalman filter-based learning rate adaptation enables continuous model updating in non-stationary environments without catastrophic forgetting. The invention has been validated on multiple benchmark complex systems including the Lorenz chaotic attractor, IEEE 14-bus power network, and finite element structural models, demonstrating 38-65% improvement in prediction accuracy and 2.1x to 4.7x reduction in optimization convergence time over comparable state-of-the-art systems. Application domains encompass structural health monitoring, electrical power grid optimization, chemical process control, biomedical signal analysis, and autonomous systems navigation. The modular API design enables domain experts to specify custom physical constraint equations, objective functions, and learning schedules without requiring expertise in AI or embedded hardware programming, substantially lowering the barrier to deployment in real-world engineering environments."
Disclaimer: Curated by HT Syndication.
