Aparna Chandramowlishwaran

Research

I study how to build AI models that learn the physics of a system and generalize across conditions, scales, and domains. My research is in scientific machine learning (also known as AI for Science), where we build neural operators, transformer-based surrogates, and foundation models to address these challenges. A growing direction in my group is mechanistic interpretability of these models to understand the representations and computations they learn, and whether their predictions remain faithful to the underlying physics.

My background in high-performance computing informs how we build these models. Architectures are designed with scalability in mind. We design algorithms and GPU kernels for computational bottlenecks such as attention, memory-efficient inference, and scalable training to enable these models to operate at the scale and speed required for practical scientific applications.

Questions I ask: How do we build a model that's faithful to the underlying physics? How do we make it fast enough to matter in practice?

Representative projects include BubbleML, NUCLEUS-MoE, Bubbleformer, and Mosaic Flows.