AI-Teaching Assistant

  • Developed scripts to scrape structured data from the textbook/ lectures using OpenAI API to obtain training Question-Answer pairs.
  • Fine-tuned Llama 3.2-11B using LoRA to develop a domain specific expert. Carried out hyperparameter optimization using Optuna and WANDB.
  • Developed Retrieval-Augmented Generation (RAG) based on several embedding model’s to leverage the fine-tuned multimodal Large Language model (LLM).
  • Aimed at improving education with providing a specialized AI assistant for targeted academic support.
  •  Demo AI-TA available here. Paper submission to a reputed conference is in progress. 

Vision Transformer to learn physics

  • Applying Vision Transformers (ViT) to learn physics from simulation images.
  • Using generative AI to learn coupled emergent spatiotemporal multiphysics phenomena.
  • Link to GitHub repo.

Multimodal transformers

  • Developing a Vision Language model (VLM) to explain the physical simulations and its mathematics.
  • Integrating fine-tuned Llama 3.2-11B for language processing with a vision transformer coupled via cross attention between LLaMA embeddings and visual tokens.

Multiphysics modeling using Finite Element method (FEM)

  • Modeling of emergent phenomena involving multiscale/multiphysics interactions using advanced numerical techniques and high-performance scientific computing.
  • Developing phenomenological description of the constituent processes, numerical formulation of the governing equations
  • Developing computational strategies and parallel code infrastructure for handling multi-million degrees of freedom
  • The code framework is in C++ and uses open source libraries like dealii (FE library), PETSc, Trilinos etc.
  • Applications include solid-state battery, shell, neurons

Physics informed neural network (PINNs)

  • Used PINNs to approximate the solution of a general type of partial differential equations.
  • Implemented using PyTorch and compared the results with numerical solvers based solution.