Automated Design Optimization of TMS Coils Using Machine Learning

Abstract

Transcranial Magnetic Stimulation (TMS) is a non-invasive brain stimulation technique used to treat neurological disorders including treatment-resistant depression, obsessive compulsive disorder, and post-stroke motor impairment. Its therapeutic effectiveness depends critically on the precise positioning of a magnetic coil on the patient’s scalp. Knowing exactly where the stimulation electric field reaches inside the brain– for any given coil position– requires solving Maxwell’s equations over a detailed head mesh. The standard tool, Finite Element Method (FEM) simulation via SimNIBS, takes approxi mately three minutes per position on a normal workstation. This delay makes interactive, real-time positioning supervision impossible in a clinical setting. This project addresses the problem from the perspective of Automation and Industrial Computing. A surrogate modelling approach directly analogous to the Digital Twin con cept used in Industry 4.0 is applied. A 3D Convolutional Neural Network (TMS_CNN_3D) is trained on 50 automatically generated FEM simulations. It learns the mapping from a 16-scalar coil configuration vector to a complete 64 × 64 × 64 electric field volume. A real-time Human-Machine Interface (HMI) integrates the trained model, allowing an operator to explore coil positions interactively. The CNN achieves a Pearson correlation of R = 0.887 and a Mean Absolute Error of 2.19% on a normalised field scale on the validation set (10 simulations). Inference runs in approximately 150ms on a standard Intel Core i5 CPU– a factor of 1,200 faster than FEM– meeting the 200ms real-time supervision target. No GPU is required.