Scientific Machine Learning
Scientific machine learning fuses data-driven models with the governing equations, symmetries, and constraints of physical systems.
A hybrid discipline
Scientific machine learning (SciML) sits between classical numerical simulation and general-purpose machine learning. Pure simulation starts from known equations and integrates them forward; pure machine learning starts from data and fits a flexible function. SciML combines both: it uses known physics to constrain what a network can learn, and uses data to fill gaps where the physics is incomplete or too expensive to resolve.
The motivation is practical. Many physical models are accurate but slow, or fast but crude. A trained surrogate can approximate an expensive solver in milliseconds, and a physics-informed loss can keep a data-driven model from producing results that violate conservation of mass, momentum, or energy.
The main families
- Physics-informed networks that add the residual of a differential equation to the training loss
- Neural operators that learn mappings between whole functions rather than between points
- Differentiable simulators whose solver itself supports gradient-based learning and inversion
- Equation discovery methods that recover symbolic laws directly from data
- Hybrid closure models that correct a coarse physical model with a learned residual
Why constraints help
A neural network with millions of parameters can fit almost any finite dataset, including its noise. Encoding physics reduces the space of admissible solutions, which improves generalization from limited data and makes extrapolation less reckless. Constraints can be soft (penalty terms) or hard (architectures that satisfy a law exactly by construction).
Where it is used
SciML is applied to fluid dynamics, materials modeling, climate, and plasma physics. In fusion, learned surrogates and operators are studied for accelerating equilibrium and transport calculations for design machines such as the Hyperion breeder, a D-T spherical tokamak whose design studies rely on many repeated solver evaluations. Such models accelerate design sweeps; they do not replace verified physics or claim measured performance from hardware that is not yet built.