What if AI learned physics to cool electric motors?

Electric motors are everywhere: cars, drones, robots. But they have an invisible enemy: heat. If they overheat, permanent magnets lose strength and the motor fails. Predicting the exact temperature inside every motor part is tough. Classic methods use physics equations: fast but imprecise at high power. Pure neural nets are accurate but don't explain why.

Enter OLTEM, a hybrid model blending the best of both worlds. Researchers took a lumped-parameter thermal network (LPTN), rooted in physics, and embedded it inside a recurrent neural network. The model thus "understands" thermal laws and learns from real data.

Two tricks make OLTEM special. First: an attention mechanism (SC-SE) that adapts network weights based on current temperature. Second: a power-loss sub-network respecting physical constraints, like heat never being negative.

Tested on a public Paderborn University dataset, OLTEM beats both a baseline neural net and a CNN-RNN model. Maximum error stays under 8-9°C even in extreme conditions. Plus, it keeps interpretability: engineers can see heat flow between stator, windings, and magnets.

This physics-informed AI will help design more compact, efficient, and safer motors. Next step? Testing transfer to other machines and cutting errors during abrupt load changes.

OLTEM (Lumped Thermal and Deep Neural Model) addresses the fundamental limit of Lumped Parameter Thermal Networks (LPTN): inability to capture strong nonlinearities at high power density without continuous manual calibration. The architecture embeds the LPTN into a discrete-time recurrent state-space formulation: x_{k+1} = A(θ)x_k + B(θ)u_k, where θ includes thermal conductances, inverse capacitances, and power losses. These parameters are not fixed but predicted by three neural sub-modules.

The key innovation is the State-Conditioned Squeeze-and-Excitation (SC-SE) block. Unlike classic SE attention weighting channels globally, SC-SE conditions weights on the instantaneous thermal state x_k, allowing the model to dynamically modulate the importance of thermal paths (e.g., stator-rotor conductance) as operating regimes change.

The loss module uses a deep MLP with Softplus activation to enforce non-negativity (P_loss ≥ 0), a critical physical constraint often violated by free regressions. Training minimizes MSE on four target temperatures (magnets, yoke, teeth, windings) with L2 regularization on θ for stability.

Results on the Paderborn dataset (WLTP profile and dynamic cycles): average MSE < 4 °C² across all components; max absolute error 4.3–8.2 °C. Ablation study: removing SC-SE raises MSE by 12-18%; removing the constrained loss module raises it by 9-15%. Complementarity is statistical: SC-SE captures spatial dynamics, the loss module captures thermal excitation accuracy.

Limitations: short prediction horizon (0.1 s steps); generalization to unseen geometries requires fine-tuning. Future work: geometric embedding for zero-shot transfer learning and quantization for automotive MCU deployment.