Predicting EV Liquid-Cooling Radiator Performance with physicsAI

Software used: physicsAI、Inspire、SimLab

Results

Highly efficient performance prediction

• physicsAI prediction results are highly consistent with CFD simulations, with Mean Absolute Error (MAE) below 0.01 and model confidence above 90%.

• Each prediction takes only 10 seconds, compared to 30 minutes for the traditional modeling and solving process, resulting in an efficiency improvement of more than 180 times.

Flexible and diverse design capability

• physicsAI accurately handles variations in different flow channel designs and geometric parameters, including curved channels, narrow channels, compressed designs, and other scenarios, showing excellent generalization capability.

• By combining the parametric modeling capability of Inspire with the script-based batch modeling in SimLab, a large number of design variations can be generated rapidly as training data, greatly accelerating design iterations.

Hardware-friendly and cost-controllable

• With an RTX 4090 GPU, full training for a single physical quantity takes only a few hours, providing low cost and high feasibility.

• physicsAI training efficiency is about 10 times faster than on CPU, making it more adaptable to enterprise resource constraints.

Technical Highlights

Parametric and automated design workflow

• Inspire is used to build parametric models, define the heatsink geometry and design parameters, ensuring flexibility and tunability (see Figure 1).

Figure 1. Inspire parametric model (Source: Hon Hai Precision Industry Co., Ltd.)

• SimLab is used for script-based batch modeling, automatically generating design variations and providing high-quality data for AI training.

Efficient prediction capability

• physicsAI can handle most application scenarios with default parameters (such as width 30, depth 3, learning rate 0.001, etc.), without the need for tedious parameter tuning.

• The training and inference are highly efficient, rapidly generating high-accuracy performance predictions.

Accurate multi-physics prediction

• Supports high-accuracy prediction of Surface Convective Heat Transfer Coefficient (HTC), internal flow velocity field, and pressure field (see Figure 2).

• Focuses on key heatsink parameters such as heat dissipation area and flow characteristics, providing reliable guidance for design optimization.

Figure 2. physicsAI-predicted flow velocity field (Source: Hon Hai Precision Industry Co., Ltd.)

Conclusion

       The application of physicsAI has greatly improved the efficiency and accuracy of EV liquid cooling heatsink design. By combining the parametric modeling of Inspire with the automated modeling capabilities of SimLab, a complete optimization workflow has been established, from geometry design to performance prediction. With its fast training speed, accurate prediction, and strong generalization ability, this technology meets the demands of thermal design for rapid iteration and precise analysis. Looking ahead, by integrating Design of Experiments (DOE) and Multi-Objective Optimization, physicsAI will further enhance heatsink design, shorten development cycles, improve product performance, provide more competitive solutions to the market, and drive continuous innovation and progress in thermal technologies.