← Research

Physics-Informed ML & Neural Operators

This pillar develops machine learning architectures that respect the physics of the systems they model, rather than treating them as black boxes. Recent work includes PINTO, a physics-informed transformer neural operator for learning generalized PDE solutions across arbitrary initial and boundary conditions; the first PINN model for simulating turbulent flow behind wind turbines (WAKE-NET, with Shell India); deformable neural operators for SAR object detection; and CNDiff, a conditional diffusion model for time series forecasting presented at ICML 2025. This is also the natural home for the lab’s newer Foundation Model Development direction — multi-GPU training of LLMs and Vision Transformers for scientific data.

Publications in this pillar

1. Multi-turbine turbulent wake simulation using physics-informed neural networks

   Azhar Gafoor CTP, Sumanth Kumar Boya, Rishi Jinka, Abhineet Gupta, Deepak N. Subramani · Physics of Fluids, 38(5), 055108 · 2026

2. Mask2Cause: Causal Discovery via Adjacency Constrained Causal Attention

   Omar Muhammad, Pasupuleti Dhruv Shivkant, Deepak N. Subramani · arXiv preprint · 2026

3. DNOD: Deformable Neural Operators for Object Detection in SAR Images

   GVS Mothish, J. Rishi, Shobhit Kumar Shukla, Deepak Subramani · Transactions on Machine Learning Research · 2025 ★ J2C Certification — Top 3% of Accepted Papers

4. PINTO: Physics-informed transformer neural operator for learning generalized solutions of partial differential equations for any initial and boundary condition

   Sumanth Kumar Boya, Deepak N. Subramani · Computer Physics Communications, Vol. 315, 109702 · 2025

5. A physics-informed neural network for turbulent wake simulations behind wind turbines

   Azhar Gafoor CTP, Sumanth Kumar Boya, Rishi Jinka, Abhineet Gupta, Ankit Tyagi, Suranjan Sarkar, Deepak N. Subramani · Physics of Fluids, 37(1), 015110 · 2025

6. Conditional Diffusion Model with Nonlinear Data Transformation for Time Series Forecasting

   Rishi J, Mothish GVS, Deepak Subramani · International Conference on Machine Learning (ICML) 2025, Vancouver, Canada · 2025 ★ CORE2023 Ranking: A*. Acceptance rate = 26.9%

7. Discrete Residual Loss Functions for Training Physics-Informed Neural Networks

   Rishi J, Sumanth Kumar Boya, Azhar Gafoor, Deepak Subramani · ICCS 2025, Singapore · 2025

8. On the Training Efficiency of Shallow Architectures for Physics Informed Neural Networks

   Rishi, J., Gafoor, A., Kumar, S., Subramani, D. · ICCS 2024. Lecture Notes in Computer Science, vol 14834. Springer, Cham · 2024

All Physics-Informed ML & Neural Operators publications on the Publications page →