Authors: The-Gia Leo Nguyen1, Javid Piruzjam2, Thomas Carraro2, Ullrich Köthe1 1 Computer Vision and Learning Lab, Interdisciplinary Center for Scientific Computing (IWR), Heidelberg University, Im Neuenheimer Feld 205, 69120 Heidelberg 2 Applied Mathematics, Institute for Modelling and Computational Science, Faculty of Mechanical and Civil Engineering, Helmut Schmidt University/University of the Federal Armed Forces Hamburg, Holstenhofweg 85, Hamburg, 22043, Germany Abstract The development of advanced energy storage materials, such as batteries, increasingly relies on virtual material testing (VMT) to link intrinsic properties with macroscopic performance. This presentation explores the challenges and methods associated with predicting parameters between different electrochemical battery models, focusing on the transition between the Single Particle Model with Electrolyte (SPME) and the Pseudo Two-Dimensional (P2D) model. Following the simulation-based inference paradigm, we use existing simulation models to generate synthetic training data for generative neural networks. We unify simulations from various battery models, spanning from simplified to more detailed representations, into a single network architecture and training framework. This joint training approach enables knowledge transfer from simplified to more complex models, improving predictive accuracy for high-fidelity simulations despite the limited training data available due to their high computational cost. Using the BayesFlow framework, we predict key battery properties from inputs such as C-rates and Electrochemical Impedance Spectroscopy (EIS), while ensuring reliable uncertainty quantification. In particular, we show that joint training with simplified models (SPME) significantly enhances predictions for more complex systems, including P2D models, by directly solving inverse problems across varying model complexities. Our results highlight the synergy between stochastic modeling, machine learning, and spatially resolved simulations, contributing to the emerging paradigm of virtual materials testing in battery research. These results demonstrate the potential for scalable, multi-model training techniques to reduce computational effort while maintaining robustness of predictions for next-generation battery technologies.