The goal of the interdisciplinary seminar series Computation & Data at HSU is to bring together researchers and foster exchange on the development of algorithms, methods and software. The seminar series is typically scheduled for the last Wednesday every month, 16:00-17:00, with 1 presentation per hybrid session (digital and at HSU). Immediately after the seminar series, the HPC Café take place.
Feel free to subscribe the seminar newsletter by sending an e-mail to info-hpc-bw@hsu-hh.de with the subject line „Subscription Seminar Computation & Data”.
Efficient data driven estimation, prediction and controll is increasingly central to battery research and battery management systems (BMS). Battery state estimation models must run fast enough for real-time monitoring, control, and large-scale design exploration. In his presentation, Dominic Karnehm will highlight neural operators and their potential for generalizability, as well as related topics in HPC, using the example of thermal state determination in Li-ion batteries. He presents the results of Paper “Generalizable Fourier Neural Operator for Estimation of Lithium-Ion Battery Temperature Distribution”, which was written during his research stay at Aalborg University, Denmark.
The study tests two models, Fourier Neural Operator (FNO) and Parameter-Embedded FNO (PE-FNO), to predict temperature distribution of cylindrical battery cells under different operating conditions. The models are trained using simulated data produced by an electro-thermal PDE model and a limited amount of measurement data. After training, the models take current, voltage, state of charge (SOC), surface and cooling temperature as time-series inputs and outputs the temperature profile over space and time. From an HPC perspective, FNO is efficient because it performs its primary computations in the frequency domain using the fast Fourier transform (FFT) and inverse fast Fourier transform (iFFT). This transforms the problem of capturing system dynamics into fast matrix operations that run efficiently on GPUs and other accelerators.
To extend performance beyond a single calibrated cell, PE-FNO integrates Channel-Attention Parameter Embedding (CAPE), injecting PDE parameters. In this study density and specific heat capacity are implemented for evaluation of the approach. This turns the surrogate into a specified solver: one trained model can generalize across a parameter space without retraining, reducing the expensive loop of “simulate, recalibrate and resimulate” that typically dominates studies related to data driven models with limited amount of experimental data. While parameter embedding trades a smaller amount of accuracy for flexibility, both operators achieve low errors on simulated cycles, and transfer learning enables strong agreement with experimental surface/core temperatures. From an HPC systems perspective, the results quantify an end-to-end speed benefit: on a workstation where the PDE solver runs on CPU and the neural operators execute on GPU, FNO and PE-FNO are ~6× and ~5× faster than the conventional PDE solver, respectively, for comparable temperature-field outputs. This acceleration supports real-time BMS deployment and enables high-throughput parametric sweeps, uncertainty studies, and digital-twin workflows where classical solvers are often the computational bottleneck.
