Additive manufacturing (AM) and advanced materials testing increasingly demand predictive simulation frameworks that are not only high-fidelity but also agile enough to enable fast and close to real-time decision-making, control, and design optimization. Traditional numerical approaches offer accuracy but are hindered by computational cost, limiting their integration into iterative design and process control loops. This minisymposium will focus on next-generation computational paradigms that unify large and small language models (LLMs/SLMs), Software 3.0 concepts, and AI-enhanced numerical methods to achieve concurrent calibration, training/learning, and solving of governing physics equations relevant to additive manufacturing. This minisymphosium will focus on Software 3.0, where models, solvers, and AI components co-evolve during runtime, enabling the direct embedding of reduced-order modelling (ROM) into AM simulation software, streamlining the offline–online bottleneck. Examples include tensor decomposition methods such as TAPS1 and INN-TD2, interpretable symbolic surrogates from Ex-HiDeNN3, and scalable convolutional hierarchical deep networks (C-HiDeNN-TD) for ultra-large-scale PDEs4. These approaches eliminate or drastically reduce offline data generation, offering speedups of several orders of magnitude while retaining mechanistic interpretability critical for certification in safety-critical industries. The minisymposium will also explore AI-enhanced multi-level variational multiscale frameworks that couple macro-, meso-, and micro-scale physics with data-driven inference. Topics include Digital Twin5 for defect prediction, physics-informed AM modelling6, and computational models leveraging differentiable solvers7 for AM and gradient-based inverse design. Novel workflows that jointly calibrate models from in-situ monitoring data, update parameters on-the-fly using various deep learning architecture8, and immediately propagate changes through multiscale solvers will be emphasized. Particular attention will be given to the interaction between LLM/SLM technologies and AM modelling: large models for multi-modal data fusion9, code generation, and simulation orchestration; smaller, domain-specialized models for real-time inference embedded in machine controllers. These capabilities align with Software 3.0’s vision of self-improving, self-documenting computational ecosystems. Through case studies in various AM techniques and mechanical testing, such as l