Artificial Intelligence (AI) is revolutionizing a wide range of fields, yet its application in scientific domains remains constrained by fundamental challenges: the scarcity of labeled data, the demand for expert-level reasoning, and the need for interpretable results that align with scientific rigor. Scientific discovery requires AI systems that not only excel in performance but also integrate seamlessly into workflows and provide insights that experts can trust and build upon. This thesis explores a spectrum of innovative approaches to bridge the gap between AI and scientific domains, addressing these challenges through a combination of self-supervised learning, expert-guided systems, and foundational model specialization. I first tackle the problem of limited labeled data with novel self-supervised learning approaches. By leveraging spatial relationships in remote sensing data, I demonstrate significant gains in sample efficiency for ecological monitoring. Next, I integrate expert interaction into the AI pipeline through systems like EeLISA, which accelerates ecological research by facilitating rapid, expert-verified segmentation of eelgrass wasting disease. To further improve learning under limited supervision, I introduce Critic Loss, a novel training paradigm that enhances model calibration and enables more effective active learning. Finally, I present advanced frameworks like GEM-RAG and AiSciVision, which specialize large multimodal models for scientific applications, demonstrating expert-like reasoning capabilities and achieving state-of-the-art results across ecological, marine, and satellite imagery tasks. These methods are validated through diverse scientific use cases, from invasive species monitoring to large-scale studies of marine ecosystems. They showcase the potential of combining foundational AI models, domain expertise, and self-supervised techniques to advance both theoretical understanding and practical impact. Looking forward, this work paves the way for further development of adaptive AI systems and their applications in science, leveraging the continued evolution of powerful foundational models to democratize AI for critical scientific challenges.