21 April 2026 in IEEE Transactions on Consumer Electronics
Solving the EEG forward problem is essential for EEG source localization but remains computationally expensive for realistic head models when using conventional numerical methods such as the Boundary Element Method. In this paper, a Boundary Potential-Informed Neural Network (BPINN) framework is proposed to efficiently solve the boundary electric potential. The proposed framework embeds boundary potential constraints into the loss function of the neural network to ensure physical consistency and introduces a solid-angle-based geometric transformation layer to extract spatial geometric features and reduce learning complexity. The effectiveness of the proposed framework is validated on both a single-sphere conductor with an analytical solution and a realistic head model with complex geometry. Numerical results demonstrate high accuracy, good generalization, and significant acceleration through transfer learning when system configurations vary. These results indicate that the proposed BPINN provides an efficient and flexible alternative for EEG forward problem.