Interview with AAAI Fellow Yan Liu: Machine Learning for Time Series

The rise of machine‑learning‑driven time‑series analysis marks a decisive shift from traditional statistical techniques to deep learning architectures. Early work relied on state‑space and Granger causal models, but the advent of recurrent networks such as LSTMs and graph convolutional networks expanded capabilities to handle complex temporal and spatial dependencies. Recent research, epitomized by Liu’s foundation model, leverages massive pre‑training to perform forecasting, anomaly detection, and generative tasks without extensive domain‑specific data, echoing the broader trend of large‑scale models reshaping AI across modalities.

Physics‑informed time‑series models represent the next frontier, marrying data‑driven learning with hard scientific laws. By integrating partial‑differential‑equation constraints directly into model architectures, these systems maintain physical plausibility while still benefiting from the flexibility of neural networks. This hybrid approach is especially valuable in fields where observations are sparse—such as subsurface transport flow, climate dynamics, or molecular simulations—enabling more reliable predictions and faster convergence with limited training samples. Liu’s PINFDiT framework exemplifies how plug‑in physics modules can refine predictions without retraining, addressing long‑standing challenges like missing values and multi‑resolution data.

Industry adoption is accelerating as firms recognize the operational advantages of adaptable, high‑generalization models. Companies like Google, Amazon, and Salesforce are building solutions atop foundation models for demand forecasting, supply‑chain optimization, and real‑time monitoring. The broader implication is a move toward "precise AI," where sensor‑rich time‑series data provide a concrete bridge between digital inference and the physical world. Over the next decade, we can expect tighter integration of physics‑aware AI, more zero‑shot capabilities, and expanded cross‑disciplinary collaborations that will drive scientific breakthroughs, smarter infrastructure, and more resilient economic systems.