Performing fast mechanical simulations of the multiple length scales of materials is a fundamental topic in computational mechanics. Overcoming this pressing challenge may yield feasible multi-scale coupled analyses of materials and enable the generation of large material (synthetic) datasets to leverage data-driven surrogate modeling [1]. For this purpose, a reduced-order method called Self-consistent Clustering Analysis (SCA) [2] was proposed to achieve a striking balance between accuracy and efficiency. The model reduction is realized in a offline-stage (learning) by means of a clustering-based domain decomposition, being the solution precision and computational cost dependent on the number of clusters. Previous numerical studies demonstrated that the computation cost of the so-called Cluster Interaction Tensors (CITs) increases significantly with the number of clusters and is consequently the bottleneck of the offline-stage. Moreover, in a recent extension called Adaptive Self-consistent Clustering Analysis (ASCA) [3], such tensors are also recomputed in the online-stage (prediction), which further stresses the need of speeding up their computation. In this contribution, we propose a new data-driven surrogate modeling approach to predict the CITs directly from the clusters support functions. The fast generation of a large representative training dataset and the choice of a suitable model architecture are key aspects of such an approach and are thoroughly discussed. Extensive comparisons between this new approach and the full computation of the CITs are presented for different domain resolutions and number of clusters, namely in terms of accuracy and computational time. At last, the efficacy of the new approach is demonstrated in the multi-scale analysis of a heterogeneous material. References: [1] B.P. Ferreira and F.M. Andrade Pires and M.A. Bessa, CRATE: A Python package to perform fast material simulations. Journal of Open Source Software, 2023. [2] Z. Liu and M.A. Bessa and W.K. Liu, Self-consistent clustering analysis: An efficient multi-scale scheme for inelastic heterogeneous materials. Comput Methods Appl Mech Eng, Vol. 306, 2016. [3] B.P. Ferreira and F.M. Andrade Pires and M.A. Bessa, Adaptivity for clustering- based reduced-order modeling of localized history-dependent phenomena. Comput Methods Appl Mech Eng, Vol. 393, 2022.