Medha Sawhney

I am a PhD student at the Knowledge Guided Machine Learning Lab advised by Dr. Anuj Karpatne.

I work on generative AI (Diffusion models, VLMs) to tackle complex scientific tasks such as solving and discovering equations through vision-based and multimodal learning.

Broadly, my research lies at the intersection of AI for Science: developing machine learning to accelerate scientific discovery, and Science for AI: using scientific principles to enhance model robustness and generalization. I’ve contributed to diverse domains including physics, geophysics, biology, and aquatic sciences through both independent and collaborative projects. My research works include:

• Diffusion models for solving PDEs in sparse, noisy, and out-of-distribution settings via physics-informed learning, with extensions to super-resolution and inpainting.
• Invertible neural networks & normalizing flows for full waveform inversion in geophysics, exploring latents for reconstruction vs. manifold-learning. (GFI’25)
• Object tracking for biomedical applications, including tiny object tracking and motion detection from noisy complex images, videos for cancer research, and inverse modeling of cellular forces.(MEMTRACK’24 , DLFM’22)
• Foundation models for aquatic science, modeling lake dynamics from real-world sensor data, with a focus on pretraining strategies and handling highly sparse observational data. (LakeFM’25)
• Vision-language models for open-world scene graph generation (OW-SGG’25), biological trait prediction (VLM4Bio’24), and equation discovery from visual inputs, addressing challenges in prompting and generating structured outputs.

During my PhD, I have interned as a Machine Learning Engineer Intern (MLE) at Twitter and NVIDIA. Prior to starting my PhD, I worked as a full-time ML Engineer at HP, and hold a Bachelor’s in Electronics and Communication Engineering from Manipal Institute of Technology, Manipal, India.

Feel free to get in touch if you would like discuss, collaborate and connect :)