DeepONet extends operator learning

Most neural networks learn one input-output rule. DeepONet was built to learn an operator: a rule that turns one whole function into another. (nature.com) That matters in science because many equations are really function machines. A boundary condition, source term or material profile goes in, and a full solution field comes out. (nature.com) Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang and George Em Karniadakis laid out DeepONet in Nature Machine Intelligence in 2021. They based it on a universal approximation theorem for operators, then extended that idea to deep neural networks. (nature.com) The architecture splits the job in two. A branch network reads sampled values of the input function, while a trunk network reads the coordinates where the answer should be evaluated. (nature.com) (docs.nvidia.com) In plain terms, the branch net asks “what function did you give me?” and the trunk net asks “where do you want the answer?” Their outputs are combined to produce the value of the target function at that point. (nature.com) The original paper tested 16 applications. Those included explicit operators such as integrals and fractional Laplacians, and implicit operators tied to deterministic and stochastic differential equations. (nature.com) (osti.gov) One practical advantage is that the input samples and output query points do not have to be the same set of coordinates. NVIDIA’s PhysicsNeMo documentation uses an anti-derivative example where the source function is sampled at one grid and evaluated at different points. (docs.nvidia.com) Nature’s accompanying commentary in 2021 noted another flexibility point: the sensors used for the input function do not have to sit on a regular grid, as long as the same sensor layout is used consistently. (nature.com) That caveat has shaped later work. A 2026 Nonlinear Dynamics paper by Andong Cong, Yuhong Jin, Haiming Yi, Yifan Jiang, Jun Li, Shijun Wang and Lei Hou says standard DeepONet still leans on uniformly sampled inputs in practice, then replaces the usual Fourier step with a nonuniform version to handle irregular grids directly. (link.springer.com) Other extensions push on geometry rather than sampling. Geom-DeepONet, reported in Computer Methods in Applied Mechanics and Engineering, encodes parameterized 3D shapes and predicts full-field solutions on an arbitrary number of nodes. (sciencedirect.com) The application list has widened too. A 2024 Scientific Reports paper examined DeepONet as a surrogate for nuclear-energy digital twins, and a separate 2024 Scientific Reports paper reported about three orders of magnitude speedup for a hypersonic-flow data-assimilation inverse problem. (nature.com 1) (nature.com 2) So the story around DeepONet is less about a single new result than a shift in what researchers ask neural networks to learn. Instead of memorizing one simulation output, they train a model to approximate the simulator itself. (nature.com)