Physics-Informed ML: Fusing Scientific Laws with Machine Learning

From predicting weather and modeling fluids to optimizing financial markets, traditional simulations rely on solving partial differential equations (PDEs) or using data-driven machine-learning models. However, differential equations solvers are often computationally expensive, and pure data-driven approaches struggle with limited or noisy data. Physics-Informed Machine Learning (PI-ML) offers a powerful alternative by embedding known physics of the problem directly into deep learning models, combining the strengths of both worlds.

This talk will introduce Physics-Informed Neural Networks (PINNs) and how we can use various open source python libraries to exploit these innovations that make scientific simulations more accessible and efficient. We’ll explore how they are transforming real-world applications, from fluid simulations in engineering to climate forecasting.

We will see:

• How the "known physics of the problem" can enhance ML models for better generalization and efficiency.
• Glimpses of implementation using Python libraries like PyTorch, Deep-XDE, and NVIDIA PhysicsNeMo (Modulus)
• Case studies where PI-ML models outperform traditional methods.

No deep math knowledge is required. This session is designed to be insightful, engaging, and accessible to ML practitioners, engineers, and researchers curious about Scientific Machine Learning.