Yuan Yuan (苑 苑)

I am currently a postdoctoral researcher at Courant Institute of Mathematical Sciences, New York University, working with Prof. Laure Zanna and Prof. Carlos Fernandez-Granda, and as a member of Multiscale Machine Learning In Coupled Earth System Modeling (M²LInES). Previously, I completed my PhD at FIBLAB, Department of Electronic Engineering, Tsinghua University, advised by Prof. Depeng Jin and Prof. Yong Li. I received my bachelor degree from the Department of Electronic Engineering, Tsinghua University in 2020.

My research centers on developing scalable artificial intelligence methodologies for real-world spatiotemporal systems, with applications in climate dynamics, urban systems, and energy infrastructures. I aim to build robust foundation models that capture complex interdependencies across spatial and temporal scales, enabling intelligent reasoning and decision-making in dynamic and uncertain environments.

Research Interests

• Foundation Models for Real-World Systems: AI foundation models for Earth systems, urban foundation models

• Real-world System Simulation: agent simulation, physical world simulation, and urban/social system simulation

News

Selected Research

Google scholar for full list.

Learning the coupled dynamics of global climate modes

By modelling the world’s climate modes as one interconnected system rather than in isolation, UniCM unlocks the emergent predictability hidden in their teleconnections—extending skilful forecasts across major modes and exposing the inter-mode interactions that precede extreme climate events.

Samudra 2: Scaling Ocean Emulators across Resolutions

Samudra 2 is a multi-resolution neural ocean emulator that runs multi-year global rollouts from 1° down to 1/4°.

Collaborative Deterministic-Probabilistic Learning for Real-World Spatiotemporal Dynamics

CoST pairs a deterministic mean-flow predictor with a diffusion residual for scalable probabilistic spatiotemporal forecasting.

Beyond Model Ranking: Predictability-Aligned Evaluation for Time Series Forecasting

A predictability-aligned diagnostic framework that moves time-series forecasting evaluation beyond leaderboard ranking.

Generative Adaptation of Dynamics to Environmental Shifts via Weight-space Diffusion

DynaDiff adapts dynamical-system predictors to new environments by directly generating expert weights via weight-space diffusion — no fine-tuning needed.

Diffusion Transformers as Open-World Spatiotemporal Foundation Models

We build a foundation model for open-world spatiotemporal learning that successfully scales up diffusion transformers.

WorldMove, a global open data for human mobility

WorldMove is an open access worldwide human mobility dataset. It accurately reflects real-world mobility patterns and ensures authenticity and reliability for urban mobility research.

Unveiling the Power of Noise Priors: Enhancing Diffusion Models for Mobile Traffic Prediction

NPDiff introduces novel noise priors for advancing mobile traffci prediction.

Breaking Data Silos: Towards Open and Scalable Mobility Foundation Models via Generative Continual Learning

MoveGCL is a scalable, privacy-preserving framework that enables training mobility foundation models without sharing raw data, which enables decentralized and progressive model evolution.

UniMove: A Unified Model for Multi-city Human Mobility Prediction

UniMove is a foundational-style mobility model that learns universal spatial representations and city-adaptive movement patterns.

Learning the complexity of urban mobility with deep generative network

DeepMobility integrates micro- and macro-scale mobility dynamics within a single generative architecture, enabling realistic synthetic trajectories and flows.

UniST: A Prompt-Empowered Universal Model for Urban Spatio-Temporal Prediction

We build a universal model for general spatio-temporal prediction and show the benefits of a one-for-all solution in urban contexts.

Spatio-Temporal Few-Shot Learning via Diffusive Neural Network Generation

This diffusion-based framework performs generative pre-training on a collection of model parameters. By generating customized model parameters, we manage to address spatio-temporal few-shot learning.

Spatio-temporal Diffusion Point Processes

We develop a diffusion model to learn spatio-temporal point processes.

Learning to Simulate Daily Activities via Modeling Dynamic Human Needs

We introduce the modeling of dynamic human needs into activity simulation.

Activity Trajectory Generation via Modeling Spatiotemporal Dynamics

ActSTD captures spatiotemporal dynamics underlying activity trajectories by leveraging neural differential equations.

Design and source code from Jon Barron's website