NeuroMANCER - Differentiable Programming Library for Data-driven Modeling and Control

Explore the open-source differentiable programming library NeuroMANCER in this one-hour talk by Jan Drgona at the Alan Turing Institute. Discover how this PyTorch-based library solves parametric constrained optimization problems, performs physics-informed system identification, and enables parametric model-based optimal control. Learn about the integration of machine learning with scientific computing to create end-to-end differentiable models and algorithms that incorporate prior knowledge and physics. Gain insights into the library's focus on research, rapid prototyping, and streamlined deployment, as well as its extensibility and interoperability within the PyTorch ecosystem. Examine tutorial examples demonstrating the use of physics-informed neural networks for solving and estimating parameters of differential equations, learning to optimize methods with feasibility restoration layers, and implementing differentiable control algorithms for learning constrained control policies in nonlinear systems.