Gallery

The banner shows a manipulator learning how to slide a puck to reach the target (red point) using Reinforcement Learning

Molecular Simulation & Materials Discovery

DNA dance and fluctuation in graphene nanopore

Proline polypeptide translocation in MoS2 nanopore

Ligand-directed conformational changes in Mu-Opioid receptor

Detection of allosteric binding sites of GPCRs with machine learning

Structure and dynamics of water molecules in CNT confinement

Conceptual Adsorb-Agent animation: propose adsorption configurations, evaluate candidates, and refine search using feedback. Schematic, without numerical research results. Source: https://arxiv.org/abs/2410.16658

Conceptual MOFGPT animation: learn MOFid representations, generate material candidates, and guide generation using validity, novelty, and target properties. Abstract lattice, not a generated crystal. Source: https://arxiv.org/abs/2506.00198

PDE Solutions & Accelerated Science

Learning the dynamics of physical systems, generating simulations from language, and discovering scientific equations. Conceptual animations are labeled separately from research demonstrations.

Text2PDE: Latent Diffusion Models for Accessible Physics Simulation. Language prompts guide generated physical dynamics.

Text2PDE — latent diffusion models translate language descriptions into physics simulations.

Hamiltonian Neural Solver: side-by-side evolution of the baseline and HNS, with normalized Hamiltonian tracked over time.

Hamiltonian Neural Solver — comparison with a baseline, showing the evolving physical field and normalized Hamiltonian over time.

Using Deep Learning to generate turbulent wake behind a cylinder

Generative machine learning models can accurately predict the thermal diffusion

Fluid Simulation with Graph Neural Networks

Conceptual LLM-SR animation: propose equation programs, fit parameters, evaluate, and refine. Uses synthetic illustrative data.

Conceptual FactFormer animation showing a spatial field, axis-wise interactions, and combined information. Illustration, not simulation or model output.

Robotics

Learning to act in the physical world: manipulation, locomotion, and autonomous systems.

Drone demonstration in a structured indoor environment. Source footage shown at 4× speed.

Robotic manipulation demonstration: a gripper interacts with a tabletop fixture and camera setup.

Robotic cutting demonstration with an overhead visual collision-detection view. Original footage is labeled 2x speed.

Robot we created in MAIL that learns to Crawl using RL

Conceptual LLM-Craft animation: shape goal, language-guided action planning, and iterative clay deformation. Not experimental footage. Source: https://arxiv.org/abs/2406.08648

AI-Enabled Manufacturing

From design intent to fabrication: translating images and language into manufacturing workflows.

Conceptual Image2Gcode animation: target image, generated toolpath, and layer-by-layer manufacturing. Illustration, not a model output.

Next page

Google Sites

Report abuse