AI/ML Projects
AI/ML Projects
Embedded Files
Predicting Aerosol Exposure in Speech-Induced Flow using Convolutional Neural Network
Predicting Aerosol Exposure in Speech-Induced Flow using Convolutional Neural Network
Computational fluid dynamics (CFD) simulations, while crucial for understanding complex physical phenomena, are often computationally expensive. This necessitates the development of efficient data-driven surrogate models. Image-to-image regression is a key technique in this area, enabling the prediction of flow or scalar fields from input data representations. This work is motivated by a physical problem where two people engaged in a conversation are involved, and if either of them is affected by a respiratory virus, how likely is the other person to get affected. Direct Numerical Simulation (DNS) for this problem was carried out by using an in-house solver “MEGHA-5" for different configurations of speaker, i.e, various axial separations and vertical offset between the two speakers. Data obtained from that analysis is used for carrying out this study for a fixed axial separation of 55D (where D denotes the characteristic diameter calculated from an ellipitical orifice used to model the speakers’s mouth) and vertical offsets are 0D (no offset), 3D, 5D, 8D, and 10D relative to a reference axis, and each offset here corresponds to a distinct configuration. This study is aimed at predicting the transport of aerosol exposure to the non-infected person using ML techniques such as Convolutional Neural Network(CNN). The model was initially trained on the dataset for individual configurations and then on a combined dataset, consisting of all the configurations and excluding just one configuration (55D axial separation & 5D vertical offset), to be used for making inference and evaluating the model’s performance against the actual DNS data and validating the results. We present a robust UNet-based deep learning framework for high-fidelity regression on multi-channel CFD data. UNet architecture is well-suited for this task due to its encoder-decoder structure and skip connections, which capture contextual information and preserve spatial details. The proposed architecture enhances the standard UNet with several improvements
Author: Bhavesh Srivastava (M.Tech)
A Chatbot for Scientific Literature Survey using RAG
A Chatbot for Scientific Literature Survey using RAG
An intelligent chatbot designed to streamline scientific literature surveys utilizes Retrieval-Augmented Generation (RAG) to seamlessly integrate information retrieval with generative AI. Orchestrated through LangChain and deployed via a Streamlit frontend, the system processes uploaded PDFs by extracting text with PyPDF2 and parsing visual data using the Unstructured library. To ensure comprehensive data capture, text is segmented using LangChain’s Text Splitter, while extracted images and figures are summarized into text via OpenAI's gpt-4o-mini. This processed information is then converted into high-dimensional vectors using OpenAI Embeddings and indexed in a FAISS vector database for rapid similarity search. Upon receiving a user query, the most relevant context is retrieved and fed into ChatGroq’s LLaMA-3-8B model to generate accurate, context-aware responses. Ongoing development involves exploring HuggingFace integrations, fine-tuning models for specialized scientific domains, and establishing a persistent database for enhanced long-term information retention.
Author: Yash Naiwar (M.Tech Reseach)
Agentic AI framework for automating Transitional Boundary Layer Research
Agentic AI framework for automating Transitional Boundary Layer Research
The process begins with an intelligent knowledge retrieval system tailored for engineering problem-solving, utilizing RAG technology to scan and extract academic texts. Complex user queries, such as specific aerodynamic parameters, are decomposed into multiple targeted sub-queries and processed through a hybrid search mechanism. By combining vector and BM25 ranking with a Weaviate database, the system retrieves the most relevant textbook chunks and empirical correlations. This data is validated, enriched with metadata, and fed into a Claude-powered LLM solver to generate highly structured, step-by-step mathematical baselines and verified sources. Armed with these theoretical foundations, the system seamlessly transitions into computational fluid dynamics (CFD) automation for high-fidelity digital simulations. Following the digital validation, the workflow extends into the physical realm by employing Python scripts to directly control wind tunnel hardware and precision sensors for empirical testing. The collected experimental data then undergoes rigorous machine learning analysis to identify complex aerodynamic patterns and process high-frequency signals. Finally, the system acts as a comprehensive synthesis engine, integrating the initial RAG-driven literature review, CFD simulation data, and physical experimental findings into an automatically generated manuscript that includes everything from raw data tables to finished, publication-ready figures.
Author: Priyanshi Dubey (M.Tech)
Page updated
Report abuse