Projects

AI-AFM Assisted Structure Prediction of Protein Complexes

• Developing novel view synthesis methods for 3D NeRF reconstruction of protein structure using Diffusion Models.
• Designed a GPU-accelerated Virtual AFM utilizing volume rendering for efficient synthetic data generation.
• Generated a large-scale synthetic dataset of multi-view AFM images for over 550,000 3D protein structures.

LLM-Powered Guidance for Defect Detection in AFM Microscopic Images

• Created an LLM-powered user interface using LangChain, LangGraph and Streamlit to guide users in identifying and resolving defects in AFM images.

Cell Shape and Defect Detection in AFM Microscopic Images

• Created an LLM-powered user interface using LangChain, LangGraph and Streamlit to guide users in identifying and resolving defects in AFM images.
• Analyzed zero-shot performance of Vision-Language Models (VLMs), including GPT-4o, Gemini, and LLaVA, for cell shape classification.
• Leveraged transfer learning to enhance YOLOv3-based object detection for cell shape detection in AFM images, achieving a 43% accuracy improvement.
• Achieved up to 60× speed improvement in AFM scanning probe traversal using intelligent vision-based navigation over manual navigation.

Deep Learning for High-resolution 3D Structural Topology Optimization

• Designed and implemented a PSP-U-Net architecture in Keras for Structural Topology Optimization.
• Developed an efficient and scalable multigrid-style training for high-resolution (128 × 128 × 128) 3D structures.
• Achieved 5x training speedup at high-resolution by leveraging distributed training with multi-node, multi-GPU setup.

Deep Learning for Structural Topology Optimization

• Designed a framework of multiple 3D CNNs to perform end-to-end topology optimization.
• Created a dataset of 60K high-resolution (128 × 128 × 128) 3D voxelized structures and accelerated the generation pipeline using GNU parallel.