Interview with AAAI Fellow Yan Liu: machine learning for time series

2026-03-19 · Source: ΑΙhub · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Emerging Technologies & Innovation · Depth: Advanced, long

Summary

AAAI Fellow Yan Liu, a Professor at the University of Southern California, discusses her extensive research in machine learning for time series and spatiotemporal data analysis. Her work, spanning over 20 years, has evolved from statistical models to advanced neural networks like LSTMs, RNNs, and GCNs, culminating in general-purpose time series foundation models capable of zero-shot forecasting and anomaly detection. Liu highlights a recent breakthrough in developing physics-informed time series foundation models, which integrate physical principles and partial differential equations to enhance accuracy and adaptability, especially in data-limited scientific domains. These models are being applied across diverse fields including Earth science transport flow modeling, climate applications, structural biology for drug discovery, and transportation systems for traffic and demand forecasting.

Key takeaway

For AI Scientists and Research Scientists developing predictive models for complex systems, integrating physics knowledge into time series foundation models offers a significant advantage. Your models can achieve higher accuracy and better generalization, particularly in data-limited scientific domains like Earth science or structural biology. Focus on developing models that conform to physical principles and PDEs to unlock breakthroughs in next-generation simulation and accelerate scientific discovery.

Key insights

Physics-informed foundation models enable accurate, adaptable time series analysis, especially in data-scarce scientific and real-world applications.

Principles

• Integrate physics knowledge into AI models.
• Prioritize generalization for diverse time series tasks.
• Address missing values and multi-resolution data.

Method

Develop general-purpose time series foundation models that incorporate physics principles and PDEs, allowing for zero-shot or few-shot adaptation to specific domains and scientific constraints.

In practice

• Apply physics-informed AI to Earth science transport flow.
• Utilize foundation models for climate and weather forecasting.
• Enhance drug discovery with molecule physics constraints.

Topics

• Machine Learning for Time Series
• Time Series Foundation Models
• Physics-Informed AI
• Spatiotemporal Data Analysis
• Causal Analysis

Best for: AI Scientist, Research Scientist, AI Researcher, Machine Learning Engineer, Data Scientist

Related on AIssential

• #356 The Forecast for Time Series Forecasts with Rami Krispin, Senior Manager of Data Science at Apple — Time series foundation models enable scalable forecasting for massive datasets, moving beyond traditional methods.
• Spatio-Temporal Models in AI: Teaching AI to Understand Space and Time — Spatio-Temporal AI models learn interconnected changes across space and time, crucial for dynamic real-world predictions.
• Phys-JEPA: Physics-Informed Latent World Models for Multivariate Time-Series Forecasting — Phys-JEPA improves multivariate time-series forecasting by imposing physical consistency directly on latent states and transitions.
• Interventional Time Series Priors for Causal Foundation Models — CausalTimePrior generates synthetic interventional time series data to train foundation models for time series causal inference.
• Semantics-Enhanced Retrieval-Augmented Time Series Forecasting — Time series forecasting improves by combining numerical similarity with semantic context from textual descriptions.
• The brewing GenAI data science revolution — Foundation models are extending beyond text/image to transform tabular and time-series data analysis.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by ΑΙhub.