Modern scientific and engineering challenges, from understanding cell growth to predicting material failure and crack formation under stress, require complex modeling and expensive experiments. While machine learning has demonstrated remarkable potential to accelerate scientific discovery for highly complex systems and reduce costs, its adoption in scientific research remains limited by a crucial bottleneck: the shortage of labeled training data. Obtaining large quantities of labeled data for scientific problems is often itself prohibitively costly, time-consuming, and sometimes physically impossible. This data scarcity creates two additional challenges: trained models often fail when applied to new conditions or experimental contexts, and the reasoning behind their predictions remains opaque, limiting confidence in the results as well as the ability to leverage those results to develop new scientific knowledge. Solving this small data problem by taking advantage of information about how the systems change with time will unlock the potential of machine learning to achieve higher performance with limited labeled datasets. This will ultimately accelerate innovation across chemistry, materials science, biology, and engineering, advancing technologies from battery development to manufacturing innovation by reducing costs, enhancing safety, and improving performance through AI-assisted automation and discovery. This project develops a unified framework for enabling machine lear