Generates atomic structures of disordered materials such as glasses that match desired target properties, and refines them into stable low-energy configurations. Also introduces new datasets of amorphous materials to support this kind of design.

Fills in missing fine-grained, time-varying traffic conditions on road networks, such as travel speeds on individual road segments during specific time periods, when observations are sparse both across segments and within each segment. Estimates the full range of likely conditions for each segment and time rather than a single value.

Forecasts future traffic flow across a road network while also estimating how uncertain each prediction is, combining transportation theory with learned models to capture how congestion shifts traffic between connected roads over time.

Learns rich and compact representations of vehicle GPS trajectories that capture both how vehicles move and why they travel. Removes redundant trajectory points and adds travel purpose information from roads and nearby places, improving accuracy and efficiency across downstream tasks.

Learns from vehicle GPS trajectories in a way that transfers across different geographic regions and different prediction tasks without retraining, removing the need to keep separate specialized models. Handles each task by treating it as recovering hidden parts of a trajectory, so one model serves many tasks even with limited data.

Recovers the missing points in sparse movement trajectories to restore detailed paths, while needing only a small amount of dense training data and generalizing across trajectories recorded at different sampling rates.

A single model for vehicle GPS trajectories that handles many tasks such as travel time estimation, trajectory recovery, and trajectory prediction, instead of maintaining a separate model for each. It stays accurate even when trajectories are sparse or only part of their features are available, by learning to rebuild dense and complete trajectories from incomplete ones.

Adapts large language models to understand movement trajectories so they can recognize how people move and infer the purposes behind their travel. A single model handles a range of trajectory analysis tasks across different settings.

Surveys existing methods for learning reusable representations of movement trajectories and brings them together into a single modular framework with shared code, so that new methods can be built, compared, and evaluated on the same basis.

Learns representations of paths in a road network by combining the network structure with text that describes physical and regional context, which earlier methods left out. The combined view improves accuracy on tasks such as ranking paths and estimating travel time, including settings with little or no labeled data.

Estimates the travel time between an origin and a destination and measures how uncertain each prediction is, by first inferring the likely route the trip will follow and then producing reliable confidence intervals that reflect how individual road segments contribute to the overall uncertainty under varying conditions.

Analyzes human mobility data such as sequences of visited locations to learn why people visit places and what travel preferences they have, and uses large language models to support tasks like predicting future locations.

Estimates the travel time between an origin and a destination at a given departure time by learning from many past trips that connect the same pair of locations. It first infers a likely route between the two points and then predicts the travel time along it, improving accuracy for map-based navigation services.

Learns general-purpose representations of movement trajectories from unlabeled data that capture both travel behavior and spatial and temporal patterns. The learned representations avoid task-specific bias so they transfer well across many downstream tasks.

Learns vector representations of locations from spatial-temporal trajectories that capture not only where places are but also when people visit them, since the time of visit reflects what a place is used for. These time-aware representations are pre-trained without labels and improve a range of downstream location-based prediction tasks.

Pre-trains location representations from movement trajectories that capture how the meaning of a place changes with its surrounding context and the time of visit, leading to more accurate prediction of a user's next location.

Predicts the future paths of vehicles for autonomous driving by inferring driver intentions such as lane changes and by modeling how nearby vehicles influence each other. It produces several likely future trajectories rather than a single guess, which improves prediction accuracy on real driving datasets.

Lets multimodal large language models analyze movement trajectories across different regions and tasks without any training, by turning raw trajectories into combined image and text inputs that keep their spatial and temporal structure. A single setup then handles many trajectory analysis tasks.

Generates drug molecules that fit a target protein's three-dimensional structure, keeping the molecules shaped to sit within the protein's surface and consistent with both its overall form and fine details. This produces more accurate candidate molecules for structure-based drug discovery.

Fills in missing values in traffic data by guiding the generation process to stay close to the available observations, adjusting how strongly it follows them for each location and time so that places with few observations are reconstructed more accurately.

Improves forecasting of multivariate time series by adapting to how patterns change over time and by modeling the relationships between the different variables while reducing the influence of noisy ones.

Generates synthetic vehicle trajectories that stay on real road networks and carry road-level information, helping address the shortage of public trajectory data caused by privacy concerns. Each generated point is tied to a road segment and a moving rate, so the trajectories are realistic and ready for downstream trajectory mining tasks.

Generates realistic synthetic check-in sequences that capture how people move and visit places over time and space, so that researchers can study and build on this kind of data without relying on scarce real records that are limited by privacy concerns.

Reconstructs the missing points in sparse GPS trajectories so the recovered paths stay on the road network. It combines the fine details of each individual trip with broader travel patterns shared across many trips to fill in the gaps more accurately.

Estimates traffic conditions at locations that have no sensor by learning from the sensors that are available, and handles newly added locations without needing to retrain the model.

Learns general-purpose representations of user check-in sequences from location-based services by jointly capturing where users go and when they go there. The learned representations transfer across different downstream tasks that rely on understanding user mobility.

Learns general-purpose representations of human check-in sequences that capture spatial and temporal patterns of where people go over time, so the same learned features transfer across many mobility tasks. Removes the need to hand-design how the data is varied during training, giving stronger and more reliable results than task-specific models.

Forecasts many future steps of a time series at once and adapts to how the patterns in the series change over time. A training scheme that judges whether predictions look realistic and follow naturally from the recent input helps keep long-range forecasts accurate.

Forecasts future values of a time series over multiple steps ahead by adapting to how temporal patterns change over time and by capturing patterns that appear at different time scales. Tested on climate and energy consumption data, it gives more accurate forecasts than earlier methods.

Forecasts time series far into the future by jointly capturing patterns at different time scales and the repeating structure across them. It does this while keeping computation efficient, so it stays fast and accurate even over very long horizons.