dynamic-systems-DL

What the repository collects

dynamic-systems-DL is not a software package but a curated resource list. The README describes it as a collection of resources that combine dynamical systems, control, and deep learning, and it invites pull requests to add good papers plus contact for research cooperation. The content is organized as a long annotated bibliography grouped by topic. The first major section, Differential Equations in Deep Learning, splits into general architectures such as recurrent neural networks for multivariate time series and deep equilibrium models, then into Neural ODEs with dozens of entries like the original Neural Ordinary Differential Equations paper, Augmented Neural ODEs, Latent ODEs, and Stable Neural Flows. It continues with Neural SDEs, Neural CSDEs, Neural CDEs, and Normalizing Flows, each with arXiv or conference links and sometimes code links. A separate section on energy based models covers Hamiltonian and Lagrangian neural networks, with references to Hamiltonian Neural Networks, Symplectic ODE-Net, and Deep Lagrangian Networks. The structure lets a reader enter from either the differential equations side or the physics-informed side and find the canonical papers in each subfield. For a graduate student or a practitioner surveying the field, the value is the grouping: related papers sit together instead of being scattered across years and venues, which shortens the time spent hunting for the right starting point.

Control, theory, and PDE sections

Beyond the core differential equation papers, the list covers Deep Learning Methods for Differential Equations, including PDE-Net, Fourier Neural Operator, and work on learning parametric PDEs from sparse data with graph neural networks. A Deep Control section lists Differentiable MPC and Neural Lyapunov model predictive control, connecting the reading list to classical control theory rather than leaving it purely in machine learning. The Dynamical System View of Deep Learning section is the most theoretical, with subsections on recurrent neural networks, theory and perspectives, optimization, signals and systems, and others. Here the entries range from early work like Gradient and Hamiltonian Dynamics Applied to Learning in Neural Networks from 1996 to newer optimal control formulations of deep learning and stability analyses of continuous-time recurrent networks. The optimization subsection alone lists Hamiltonian descent methods, port-Hamiltonian training approaches, and implicit differentiation for hyperparameter search. By collecting these, the repository makes the argument that deep learning can be studied as a dynamical system or a control problem, and it gives the reader the citations to follow that line of thought without assembling the map themselves. The optimization subsection alone lists Hamiltonian descent methods, a port Hamiltonian training approach, and implicit differentiation for hyperparameter search, which shows how far the reading list reaches within a single theme.

Software pointers and scope

The README closes with a Software and Libraries section that points to concrete tools for readers who want to move from reading to implementation. On the Python side it names torchdyn as a library for neural differential equations, torchdiffeq for differentiable ODE solvers with GPU support and constant-memory backpropagation, torchsde for stochastic differential equation solvers, torchSODE for a block-diagonal ODE solver, and JAX MD for differentiable physics. On the Julia side it names DiffEqFlux and NeuralNetDiffEq, the latter implementing ODE, SDE, and PDE solvers via deep neural networks, with support for stiff and non-stiff problems and adaptive time stepping. A short Websites and Blogs section links the Scientific ML blog by Chris Rackauckas and the SciML project, and a Slides section links implicit layers and ODE talk materials. The repository carries no license file and no primary language in its metadata, which matches its nature as a reading list rather than a codebase. Its scope is broad but coherent: everything listed sits at the intersection of continuous-time systems and learning, and the consistent annotation style makes the list readable as a single document rather than a dump of links. The Software and Libraries section also names torchdyn, torchdiffeq, torchsde, and JAX MD on the Python side, giving a reader concrete tools to move from reading to implementation.

Editorial conclusion

The dynamic-systems-DL repository has no listed license and no primary language in its metadata, and it was last updated on GitHub on 2026-08-24 with 31 stars recorded.