In the afternoon of June 25, 2026, Professor Yue Liu from the School of Computer Engineering and Science at Shanghai University delivered an academic lecture entitled “Interpretable and Controllable Generative Models and Their Applications in Materials Science” in Academic Activity Room 526 of Xingjian Building, at the invitation of Professor Zhibin Sun. Focusing on trustworthy applications of generative models in complex materials systems, the lecture presented the research group’s progress in interpretable and controllable modeling, materials-domain knowledge embedding, and multi-model collaborative intelligent design, outlining an AI-enabled pathwayspanning data accumulation, knowledge discovery, and candidate-material design.Generative models are attracting increasing attention in materials science because of their strong capabilities in pattern learning and data modeling. Materials systems, however, are simultaneously constrained by composition, structure, properties, and physical and chemical laws. Purely data-driven models may therefore produce results that are difficult to explain, difficult to control, andinsufficiently reliable in practical applications. Professor Liu emphasized that embedding domain knowledge, prior scientific principles, and explicit constraints into machine-learning models is essential for improving both scientific credibility and engineering usability.At the data level, the lecture introduced an interpretable and controllable variational autoencoder (VAE). Designed for common challenges in materials research, including small datasets, imbalanced distributions, and shortages of high-quality samples, the model uses latent-space representation and controlled generation to augment scarce data and optimize data distributions, thereby improving scale, diversity, and validity. This approach helps address the high cost and long cycle of experimental data acquisition while providing a stronger foundation for property prediction and candidate screening.At the knowledge level, the research group has developed a large language model (LLM)embedded with domain knowledge and equipped it with a domain-knowledge-constrained error-correction mechanism. The framework improves the model’s ability to identify and extract materials entities, relations, and specialized semantics. It also enables knowledge dispersed across publications, databases, and experimental records to be mined, connected, and organized more systematically, supporting materials knowledge-base construction, scientific question formulation, and hypothesis generation.At the design level, the lecture highlighted a multi-model collaborative system integrating chemical-formula generation, crystal-structure generation, and property prediction. Under physical and chemical constraints, the system efficiently generates and screens candidate solid-state electrolyte materials. In this framework, generative modeling is no longer limited to producing samples; it works together with structural modeling, performance evaluation, and rule-based validation to form a closed-loop workflow for intelligent materials discovery.The lecture demonstrated three major roles for interpretable and controllable generative models in materials research: strengthening the data foundation through augmentation, facilitating scientific discovery through knowledge extraction, and accelerating materials design through constrained
generation and collaborative screening. The work offers new approaches to improving the trustworthiness, controllability, and transferability of AI models in materials science, whilesupporting a transition from experience-based trial and error toward coordinated innovation drivenby data, knowledge, and models.Professor Yue Liu is a professor and doctoral supervisor at the School of Computer Engineering and Science, Shanghai University. She has conducted research for 25 years on machine-learning generalization, usability, interpretability, and applications. Since 2014, she has collaborated with Professor Shi Siqi to develop a new paradigm for energy-storage materials design that integrates algorithms, data, knowledge, and experiments. Her research has been published in journals including National Science Review, Acta Materialia, and Pattern Recognition, and has contributed to thedevelopment and industrial application of an electrochemical energy-storage materials design platform. The lecture further illustrated the potential of deep interdisciplinary integration between artificial intelligence and materials science.（By Zhibin Sun）