MUMBAI, India, Oct. 26 -- Intellectual Property India has published a patent application (202621106998 A) filed by Parul University on September 05, 2026, for An Ai-Based Framework For Automated Mathematical Model Generation, Validation, And Optimization.
Inventors include Dr. Snehlata Barde; Dr. Virendra Gawande; Dr. Aarti Sharma; and Ms. Sarika V. Gawande.
The application for the patent was published on October 23, 2026, under issue no. 43/2026.
Abstract: ABSTRACT OF THE INVENTION: The present invention relates to an artificial-intelligence-based framework that fully automates the generation, multi-criteria validation, and continuous optimisation of mathematical models from heterogeneous inputs. Conventional mathematical modelling remains a predominantly manual, expertise-intensive activity that struggles to keep pace with the volume and complexity of modern scientific and industrial data. Existing automated techniques either produce opaque black-box predictors lacking interpretability and physical consistency, or generate symbolic expressions without systematic enforcement of conservation laws, multi-objective trade-offs, or uncertainty quantification. The disclosed framework overcomes these deficiencies through a tightly integrated five- module architecture. A multi-modal input processor ingests numerical datasets, natural-language problem statements, explicit symbolic constraints, and optional domain knowledge graphs. A domain-adapted large language model extracts entities, candidate functional forms, and physical principles, which are combined with automated feature engineering of the numerical data. A hybrid model generation engine then constructs candidate models by interleaving large-language-model guided proposals, genetic-programming symbolic regression, sparse identification of nonlinear dynamics, and neuro-symbolic architecture search. The resulting population of candidates—expressed as closed-form equations, systems of ordinary or partial differential equations, or hybrid neural-symbolic networks—is passed to a multi-criteria validation module. This module simultaneously scores every candidate on predictive accuracy, residual violation of physical laws, structural complexity, and generalisation under distribution shift, thereby constructing a Pareto front of non-dominated solutions. An optimisation and uncertainty-quantification module subsequently refines both continuous parameters and discrete structural choices using Bayesian optimisation and ensemble methods, returning posterior distributions and confidence intervals. A human-in-the-loop interface presents ranked models in mathematical notation together with natural-language explanations, highlights regions of high uncertainty, and permits experts to inject additional constraints or reject specific functional forms. Accepted models are exported in standard formats suitable for simulation, control, and digital-twin environments. Because the entire pipeline operates in a closed loop, newly arriving data or revised constraints trigger incremental updates without requiring a complete restart, enabling continuous adaptation in dynamic settings. Empirical evaluation on benchmark problems drawn from fluid dynamics, chemical kinetics, epidemiology, and structural mechanics demonstrates that the framework recovers governing equations of accuracy comparable to or higher than expert-derived models while reducing human modelling effort by more than an order of magnitude and supplying explicit uncertainty estimates. The invention therefore provides a practical, interpretable, and continuously improving solution for automated mathematical modelling across scientific and engineering domains.
Disclaimer: Curated by HT Syndication.
