HAWFNO: A hybrid attention-weighted Fourier neural operator for full-field prediction and hot-spot screening of parametric conjugate heat transfer

Parametric conjugate heat transfer (CHT) sets the temperature field and the local hot-spot risk in thermal-management systems, where heat-source placement, source strength, inlet velocity, inlet temperature and convection regime change together during design. High-fidelity computational fluid dynamics (CFD) is accurate but too costly for sweeping the many candidate layouts of early design. Using two representative cases—a channel cooled around three distributed heat sources and a simplified transformer-like cooling section—this study tests whether a neural-operator surrogate can deliver rapid full-field temperature prediction and hot-spot screening. The proposed hybrid attention-weighted Fourier neural operator (HAWFNO) maps geometry, heat-source distribution and operating conditions to the temperature field through a Fourier-neural-operator backbone with multi-scale inputs and direction-aware feature fusion. On the three-heat-source case, covering Richardson numbers from 0.001 to 60 , it attains an MAE, RMSE, R 2 and peak-temperature error of 0.059 K, 0.213 K, 0.9548 and 0.086 K, reducing MAE, RMSE and peak-temperature error by 15.7 % , 7.4 % and 25.9 % over the best baseline; on the transformer-like case the MAE drops by 42.2 % . After training, one field takes about 0.004 s against about 150 s per CFD solve ( ∼ 3.75 × 10 4 faster), so a thousand candidates needing about 41.7 h of CFD are screened in about 4 s. The surrogate thus suits early-stage layout screening and hot-spot assessment, while shortlisted designs still require CFD verification.