Cross-machine profile prediction and differentiable optimization by learning multi-tokamak databases

The development of high-performance plasma regimes is becoming increasingly important for tokamak operation and scenario design. Conventional operation-space exploration relies on computationally expensive iterative simulations and experimental trial and error. Although machine-learning methods have recently been applied to this problem, many optimization approaches treat the predictive model as a black box. They can therefore require a large number of model evaluations and must be retrained when the objective or operational constraints change. In this work, we develop a neural-network surrogate for multi-tokamak profile prediction using public databases. Device-dependent dimensional input and output variables are transformed into appropriate normalized quantities. These normalizations of input and output variables enable the prediction model to be applied to new devices such as ITER. The surrogate forms a fully differentiable mapping from engineering parameters to plasma profiles and derived performance metrics. Automatic differentiation provides the exact Jacobian of this mapping and enables gradient-based optimization without treating the model as a black box. Under identical objectives and bounds, the proposed method requires substantially fewer model evaluations than other machine-learning approaches. The proposed framework offers an efficient route from multi-machine data to interpretable plasma optimization and reactor scenario design.