Prashant Govindarajan

Prashant Govindarajan

PhD Student at Mila-Quebec AI Institute and Polytechnique Montréal

Prashant Govindarajan

Hello and வணக்கம்! I am a fourth-year Computer Engineering PhD candidate at Mila-Quebec AI Institute and Polytechnique Montréal (engineering school of UdéM), working under Sarath Chandar. I am keenly interested in reinforcement learning and large language models for scientific discovery and beyond. In collaboration with Intel, I developed offline and online reinforcement learning methods for crystalline material design using first-principles. I also worked with Ansys on LLMs for automating Computer-Aided Design (CAD). My research is supported by the PBEEE scholarship for international students.

I was previously a dual degree student at the Indian Institute of Technology Madras, where I worked under Balaraman Ravindran and Karthik Raman on target-specific drug design. Besides academics, I like playing chess, frisbee, traveling, and cooking. Lately, I’ve been learning to snowboard and play the flute. Feel free to reach out if you’d like to chat about research or anything beyond 😁! I’m always eager to learn more about crystallography, DFT, and solid-state physics, and would love to exchange ideas, especially on the RL aspects of my work!

I am actively seeking research internships in the areas of LLMs, reinforcement learning, and AI for Science.

Publications

  • Govindarajan, Prashant*, Davide Baldelli*, Jay Pathak, Quentin Fournier, and Sarath Chandar. CADmium: Fine-Tuning Code Language Models for Text-Driven Sequential CAD Design. arXiv preprint arXiv:2507.09792 (2025).
  • Govindarajan, Prashant, Mathieu Reymond, Antoine Clavaud, Mariano Phielipp, Santiago Miret, and Sarath Chandar. CrystalGym: A New Benchmark for Materials Discovery Using Reinforcement Learning. In AI for Accelerated Materials Design-ICLR 2025.
  • Govindarajan, Prashant, Mathieu Reymond, Santiago Miret, Mariano Phielipp, and Sarath Chandar. Crystal Design Amidst Noisy DFT Signals: A Reinforcement Learning Approach. In AI for Accelerated Materials Design-NeurIPS 2024.
  • Govindarajan, Prashant, Mathieu Reymond, Santiago Miret, Antoine Clavaud, Mariano Phielipp, and Sarath Chandar. A Reinforcement Learning Pipeline for Band Gap-directed Crystal Generation. In AI for Accelerated Materials Design-Vienna 2024.
  • Govindarajan, Prashant, Santiago Miret, Jarrid Rector-Brooks, Mariano Phielipp, Janarthanan Rajendran, and Sarath Chandar. Learning Conditional Policies for Crystal Design Using Offline Reinforcement Learning. Digital Discovery (2024).
  • Govindarajan, Prashant, Santiago Miret, Jarrid Rector-Brooks, Mariano Phielipp, Janarthanan Rajendran, and Sarath Chandar. Behavioral Cloning for Crystal Design.” In Workshop on Machine Learning for Materials, ICLR 2023.

News

Projects

Fine-tuning LLMs for automating Computer-aided Designing

With Ansys, we developed CADmium, a new open-source dataset consisting of natural language descriptions for 170k+ 3D CAD objects. We propose a novel LLM fine-tuning workflow with code LLMs, and augment existing evaluation schemes with new metrics.

Reinforcement Learning for Crystal Structure Design

Currently exploring offline and online reinforcement learning methods for learning a policy for sequentially designing crystal structures.

Graph generative models for binding site-specific molecule generation

Designing novel deep graph generative model for generating new ligand molecules that can bind to a given target receptor binding site.

Analysis of drug response and gene expression data of AML cells

Computational methods to identify drug-gene correlations and molecules that can induce leukemic cell maturation.

Incorporating Geometry into Score-Based Model for Crystal Structure Design

Two ways to incorporate crystal symmetry information as an inductive bias into a generative model for crystal structure design.

Analyzing the effects of visual representations in visual navigation tasks

Two ways for encoding visual inputs for navigation task using RL – pretraining a contrastive learning-based SimCLR model and VAE-based generative approach.

Deep generative models for single-cell gene expression analysis

Evaluated state-of-the-art unsupervised deep learning techniques including variational autoencoders for single-cell gene expression data analysis.

Generating drug-like molecules from gene expression signatures using transformer

Designed an attention-based transformer model for de novo generation of drug-like molecules that can induce a desired transcriptomic profile. Accepted as poster at MLCSB COSI, ISMB 2022.

Parallel analyses of canonic polyadic tensor decomposition algorithm

CPU- and GPU-level parallelization of tensor decomposition algorithm using OpenMP and OpenACC.