BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Wuyang Chen (Simon Fraser University) DTSTART:20241101T220000Z DTEND:20241101T233000Z DTSTAMP:20260407T223725Z UID:AppliedMath/56 DESCRIPTION:Title: Towards Data-Efficient OOD Generalization of Scientific Machine Learn ing Models\nby Wuyang Chen (Simon Fraser University) as part of SFU Ma thematics of Computation\, Application and Data ("MOCAD") Seminar\n\nLectu re held in K9509 and Hybrid.\n\nAbstract\nIn recent years\, there has been growing promise in coupling machine learning methods with domain-specific physical insights to solve scientific problems based on partial different ial equations (PDEs). However\, there are two critical bottlenecks that mu st be addressed before scientific machine learning (SciML) can become prac tically useful. First\, SciML requires extensive pretraining data to cover diverse physical systems and real-world scenarios. Second\, SciML models often perform poorly when confronted with unseen data distributions that d eviate from the training source\, even when dealing with samples from the same physical systems that have only slight differences in physical parame ters. In this line of work\, we aim to address these challenges using data -centric approaches. To enhance data efficiency\, we have developed the fi rst unsupervised learning method for neural operators. Our approach involv es mining unlabeled PDE data without relying on heavy numerical simulation s. We demonstrate that unsupervised pretraining can consistently reduce th e number of simulated samples required during fine-tuning across a wide ra nge of PDEs and real-world problems. Furthermore\, to evaluate and improve the out-of-distribution (OOD) generalization of neural operators\, we hav e carefully designed a benchmark that includes diverse physical parameters to emulate real-world scenarios. By evaluating popular architectures acro ss a broad spectrum of PDEs\, we conclude that neural operators achieve mo re robust OOD generalization when pretrained on physical dynamics with hig h-frequency patterns rather than smooth ones. This suggests that data-driv en SciML methods will benefit more from learning from challenging samples. \n LOCATION:https://researchseminars.org/talk/AppliedMath/56/ END:VEVENT END:VCALENDAR