Model-based Battery Management Systems - Where Solvers Fail

Fast charging is being heavily researched for the widespread implementation of lithium-ion batteries for electric vehicles. However, charging at high currents accelerates several parasitic reactions that lead to the degradation of the cell, affecting its lifetime. It is possible to study material degradation mechanisms and predict their impact on capacity loss under several operating conditions using physics-based multi-scale battery models. These models can be integrated with battery management systems (BMSs) to control the cell’s performance and to design novel charging protocols that enable safe and optimal cell performance and suppress cell degradation. Our group has successfully applied  BMS^2,3 based on a physics-based battery model to improve life and reduce charging time for different batteries (as shown in Figure 1). This seminar will present some results from our group for cells, modules, and packs. The talk will also include the theoretical development of pulse profiles as predicted by optimal control of phase-field models.

Model-based BMS algorithms require fast and efficient production codes that can predict and estimate battery parameters in real-time and control the battery’s performance under different loads. The theory is reasonably well developed and defined for one or all of (a) ordinary differential equations (ODEs), (b) elliptic partial differential equations (PDEs) with homogenous boundary conditions, and (c) well-conditioned linear equations. Solvers, optimizers, and software have been well developed for the same by optimizing/utilizing one or all of (a) high-performance computing, (b) GPU acceleration, (c) adaptive time stepping, (d) adaptive mesh refining, (e) estimation, (f) parallel computing, (g) efficient Jacobian/Hessian/adjoint calculation, and (h) sparse linear algebra. When physics-based models are taken to BMS - solvers and optimizers can (and often) fail. A Google search (aided by Generative AI) on “where solvers fail” returns four common scenarios (not all of them are accurate or relevant).

Speaker:Venkat Subreamanian, University of Texas