1: \begin{abstract}

2: We introduce a new class of hybrid preconditioners for solving parametric linear systems of equations. 

3: The proposed preconditioners are constructed by hybridizing the deep operator network, namely DeepONet, with standard iterative methods. 

4: Exploiting the spectral bias, DeepONet-based components are harnessed to address low-frequency error components, while conventional iterative methods are employed to mitigate high-frequency error components.

5: Our preconditioning framework comprises two distinct hybridization approaches: direct preconditioning (DP) and trunk basis (TB) approaches. 

6: In the DP approach, DeepONet is used to approximate an action of an inverse operator to a vector during each preconditioning step. 

7: In contrast, the TB approach extracts basis functions from the trained DeepONet to construct a map to a smaller subspace, in which the low-frequency component of the error can be effectively eliminated. 

8: Our numerical results demonstrate that utilizing the TB approach enhances the convergence of Krylov methods by a large margin compared to standard non-hybrid preconditioning strategies. 

9: Moreover, the proposed hybrid preconditioners exhibit robustness across a wide range of model parameters and problem resolutions.

10: \end{abstract}

11: