MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202511130200 A) filed by Indian Institute Of Technology Banaras Hindu University, Varanasi on December 22, 2025, for A Continual Learning System And Method For Molecular Property Prediction.
Inventors include Sanjay Kumar Singh; and Sakshi Ranjan.
The application for the patent was published on July 03, 2026, under issue no. 27/2026.
Abstract: The present invention provides a continual learning system for molecular property prediction comprising a processor (102) that includes a preprocessing module (103) configured to receive a molecular input data in a predefined format and perform canonicalization, tokenization, and embedding generation, a multi-task continual learning module (104) including at least one pre- trained transformer-based molecular language module that is sequentially trained over a plurality of molecular property prediction tasks, a forget recall mechanism module (105) to selectively suppress a plurality of obsolete molecular representations while retaining task-relevant representations during sequential learning, and a Pareto optimization module (106) to balance stability and plasticity by optimizing a plurality of competing objectives associated with a knowledge retention and a knowledge acquisition through a hierarchical gradient aggregation, and an evaluation and inference module (107) configured to perform a real-time molecular property prediction and quantify catastrophic forgetting by utilizing a plurality of predefined continual-learning performance metrics. Figure 1 on sheet no. 1 of the drawings may accompany the abstract when published.
Disclaimer: Curated by HT Syndication.
