AI-ENHANCED COMPUTATIONAL FRAMEWORKS FOR COMPLEX PHYSICAL SYSTEMS: 3D MICROSTRUCTURE RECONSTRUCTION, QUANTUM NON-UNITARY DYNAMICS, AND REINFORCEMENT LEARNING BASED RADIOTHERAPY OPTIMIZATION

Abstract

Artificial intelligence (AI) and machine learning (ML) have emerged as transformative tools for modeling and optimizing complex physical systems across diverse scientific and engineering domains. This article synthesizes recent advances in three computationally intensive application areas: (1) AI-driven 3D microstructure reconstruction from limited imaging data, (2) computational simulation of quantum non-unitary dynamics in open quantum systems, and (3) reinforcement learning (RL)–based optimization of radiotherapy treatment planning. In microstructure reconstruction, generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models enable synthesis of statistically equivalent 3D volumes from single 2D cross-sections, achieving relative misorientation errors of approximately 3×10⁻² and generation of 10⁸-voxel volumes in seconds. For quantum non-unitary dynamics, hybrid quantum algorithms, tensor-network methods, and stochastic unravelings extend classical and quantum computational reach to systems of ~1000 spins, with algorithmic complexity bounds of O(t²‖L‖/ε) for repeated-interaction embeddings. In radiotherapy, deep reinforcement learning (DRL) agents employing DQN, PPO, and actor-critic architectures demonstrate plan score improvements (e.g., 120→141 for IMPT replanning), parotid dose reductions of ~7 Gy, NTCP reductions of −5.46%, and plan generation times of 3.3–12.4 minutes. Collectively, these advances illustrate how AI-enhanced computational frameworks address high-dimensional inverse problems, accelerate physics-based simulations, and automate complex decision-making in safety-critical applications. This review provides a unified perspective on methods, quantitative benchmarks, and future directions for AI integration in computational physical sciences.