Multi-Beam Writing Correction for Circuit-Induced Beam Shift

The patent replaces the conventional method of creating complex correction tables through manual measurement and data processing with a neural network-based automated system. The neural network learns the relationship between beam positions and correction amounts from training data, automatically generating correction values without requiring complex manual table creation. This substitutes the mechanical/data-processing system with an intelligent system that simplifies the workflow while maintaining or improving correction precision.

The patent creates a virtual model of the beam writing system through the neural network, which copies and learns from training data representing the complex interactions between multiple beams and the writing medium. This virtual model then generates correction values that replicate the effects of complex physical measurements without requiring actual physical measurements for every correction scenario.

The patent performs preliminary correction by calculating beam position displacements caused by Coulomb effects before the actual writing process. The system computes correction values based on the planned beam irradiation pattern and applies these corrections in advance, ensuring that beams are positioned accurately despite the repulsive forces that will occur during multi-beam writing. This preliminary correction prevents position errors rather than correcting them after they occur.

The patent applies preliminary anti-action by calculating and applying correction values that counteract the expected Coulomb repulsion effects before they occur. The system predicts the beam position displacement caused by inter-beam repulsion and applies equal and opposite corrections to the beam positions, effectively neutralizing the harmful effect before it can degrade writing accuracy.

The patent changes the approach from using fixed correction parameters to using a neural network that can adapt its parameters based on input conditions. The system takes into account various parameters including beam positions, irradiation patterns, and circuit characteristics, and dynamically generates appropriate correction values. This allows the same system to adapt to different circuit layouts and writing conditions without requiring manual reconfiguration.

The patent implements self-service by enabling the system to automatically correct for its own circuit-specific variations and field effects. The neural network learns from training data that includes the specific characteristics of the control circuits and blanking aperture array substrate, allowing the system to self-correct for variations in electric and magnetic fields generated by its own components without requiring external calibration or adjustment.