Seminar, Hosted by Argonne National Laboratory
Data and the Limits of Scientific Machine Learning
University of Chicago
Abstract
Mansi Sakarvadia studies how to make scientific machine learning useful when high-quality data are expensive. The talk examines where neural surrogates trained on inexpensive discrete data fail to capture continuous phenomena, and how training on data of differing quality can improve performance for a given cost. It also considers how information supplied at inference time can support discovery when a language model’s prior knowledge is uncertain.