An AI-Driven Framework for Directed Self-Assembly Toward Design-Technology Co-Optimization

Directed self-assembly (DSA) provides a promising route for sub-10-nm patterning, but its integration into design-technology co-optimization (DTCO) remains limited by the difficulty of translating self-assembly process constraints into manufacturable layout solutions. Here, an artificial-intelligence-driven closed-loop simulation framework for DSA is presented to convert experimentally constrained DSA process knowledge into a process-constrained design-space map for DSA-aware layout decomposition. The framework integrates a three-dimensional Ohta-Kawasaki model with Bayesian optimization and convolutional neural network screening, enabling automated model calibration using experimental racetrack-template data without manual tuning. It efficiently identifies optimal model parameters, achieving a mean simulation-experiment error below 1.5 nm while maintaining morphological consistency. In addition, four multimodal large language models are further evaluated for DSA morphology screening and are shown to adapt to new image modalities and defect criteria without additional task-specific fine-tuning, relying solely on text-based rules and few-shot examples. The calibrated model enables target-constrained inverse design of experimentally validated template windows, translating DSA constraints into a manufacturable solution-space map for layout decomposition. Its extension to peanut-shaped templates without additional parameter calibration further supports DSA-aware layout decomposition and DTCO. Overall, this work establishes an intelligent closed-loop workflow linking experiment, simulation, and design, providing a practical route toward DTCO.