Designing an AI-Driven Research Pipeline: Calibration, Synthetic Data, Latent Learning, and Dynamic Models

When the twin problems of data scarcity and model drift in battery-management research are tackled, a promising hypothesis emerges: that a tight integration of physics-based simulation, synthetic data generation, and modern deep learning techniques can outperform static equivalent-circuit models (ECMs) in tracking battery aging. Based on this hypothesis, a four-stage research framework has been constructed. Stage 1 -Calibration: the high-order PyBaMM electrochemical model is calibrated using a two-phase iRACE optimization, first estimating non-degradation parameters and subsequently degradation-related ones, thereby establishing a trustworthy ground-truth simulator. Stage 2 – Synthetic Data Generation: once calibrated, the model is used to simulate thousands of randomized cycling scenarios – varying load, temperature, and depth-of-discharge – spanning the full degradation trajectory down to 80 % state-of-health. This process yields a large, diverse dataset suitable for training robust neural models. Stage 3 – Representation Learning: a dual-path autoencoder is employed, in which a lightweight CNN reconstructs the healthy voltage signal while a Transformer encodes the residual (aging-related signal) into a compact, interpretable latent space -demonstrating the advantages of architectural specialization. Stage 4 – Dynamic ECM Adaptation is divided into two complementary strategies: 4.1 Reinforcement Learning, where the adaptation of ECM parameters is framed as a sequential decision-making problem aimed at minimizing prediction error while satisfying operational constraints; and 4.2 Physics-Informed Neural Networks (PINNs), where the latent aging code is mapped directly into ECM parameters through a network constrained by the underlying physical equations, ensuring that predictions remain consistent with the governing electrical laws. Taken together, these stages form a generalizable pipeline: define the problem through physical modeling, generate synthetic yet realistic data, learn disentangled latent representations, and dynamically adapt models using hybrid AI–physics approaches—offering a roadmap not only for battery systems, but for a broad class of cyber-physical research problems.