The Mechanical and Artificial Intelligence Lab (MAIL) at Carnegie Mellon University develops machine learning methods for understanding and predicting physical systems. We connect numerical simulation, scientific data, and physical principles to study fluid dynamics, transport, and other processes central to mechanical engineering.

Transformer-based PDE modeling is a foundational thread of our research. Operator Transformer (OFormer), introduced in 2022 and published in TMLR in 2023, learns PDE solution operators using self-attention and cross-attention, with flexibility in how input data and prediction locations are sampled. Our subsequent Scalable Transformer for PDE Surrogate Modeling (FactFormer, NeurIPS 2023) uses axial factorization to reduce the cost of attention for multidimensional physical fields, demonstrated on high-resolution 2D flows and 3D smoke simulations. Together with our graph-based surrogates, these methods connect flexible operator learning with scalable physical prediction.

We also seek interpretable explanations. LLM-SR (ICLR 2025) combines language-model scientific knowledge with numerical search to discover equations from data. LLM-SRBench (ICML 2025) provides a benchmark for evaluating this emerging approach to scientific discovery.

Complementary work in physics-informed diffusion and super-resolution reconstructs fine-scale flow structure from coarse observations. Together, these research directions address three connected needs: faster simulations, richer information from limited data, and mathematical models that scientists can inspect and test.

For methods, papers, and project details, click here.

Discovering a useful material requires searching a vast space of possible structures and compositions. At MAIL, we develop models that connect molecular representations to properties, generate promising candidates, and guide computational screening. Our research spans polymers, metal–organic frameworks, catalysts, and molecular systems relevant to drug discovery.

Self-supervised methods such as MolCLR learn useful representations from unlabeled molecules, while TransPolymer and MOFormer adapt transformer models to polymers and metal–organic frameworks. MOFGPT (2025) moves from property prediction to property-guided generation, combining language modeling with reinforcement-learning feedback to propose new MOF candidates.

Our recent work also explores agents that organize scientific reasoning and computation. Adsorb-Agent (2026) uses language-model reasoning to propose adsorption configurations for atomistic evaluation. AgentD (2026) coordinates specialized tools for molecular generation, property prediction, and early-stage drug-discovery workflows.

Across these projects, the goal is to make exploration more efficient while keeping proposed structures and predicted properties connected to physical evaluation. For methods, datasets, and publications, click here.

Robots must connect what they see, what they sense, and what they do. At MAIL, we study learning-based manipulation in tasks where physical interactions are difficult to model, including pouring liquids, sculpting clay, and handling food. Our methods combine demonstrations, learned dynamics, and feedback from the environment.

Human demonstrations collected through intuitive teleoperation provide a foundation for learning. We investigate diffusion policies, Action Chunking with Transformers (ACT), and multimodal sensing to represent action sequences and respond to changing task conditions.

Our research brings together three capabilities:

• Learning manipulation skills from demonstrations and physical interaction.

• Connecting high-level goals to actions through perception and planning.

• Combining visual and tactile information to improve policy learning.

LLM-Craft uses language-model reasoning and feedback to guide clay sculpting from shape descriptions and semantic goals. It complements SculptBot, which learns the dynamics of clay deformation. VITaL (ICRA 2025) explores visuo-tactile pretraining, including how tactile experience during training can improve policies that use only vision at deployment.

Across these projects, we investigate how learned representations and physical feedback can make robot behavior more adaptable and data-efficient.

For projects, demonstrations, and publications, click here.

At MAIL, we develop AI methods that connect design, process monitoring, and decision-making in additive manufacturing. Our research spans material-extrusion printing and monitoring for metal additive manufacturing, with a focus on using data and physical feedback to improve fabrication workflows.

Generative and Agentic Manufacturing Control

LLM-3D Print (2025) combines visual monitoring with language-model reasoning to diagnose printing problems and plan corrective actions. The PrinterChat project presents this approach to interactive printer control. Image2Gcode (2026) addresses a complementary challenge: using a diffusion transformer to generate printer-ready G-code from 2D images, connecting design inputs directly to fabrication instructions.

Real-Time Monitoring and Control

We study how images and sensor measurements reveal changes in build quality. Our work includes feedback-driven monitoring of extrusion printing and diffusion-based optical image super-resolution for layerwise monitoring in laser powder bed fusion. Better observations support defect analysis and informed process decisions.

Predictive Modeling and Process Optimization

We investigate models that connect process conditions to manufacturing outcomes and explore how learned representations can transfer across machines, materials, and operating conditions. AMGPT complements these models by retrieving additive-manufacturing literature to support domain-specific questions and engineering knowledge access.

Open-Source Tools and Data Repositories

We share project resources, demonstrations, and available code to support reproducibility and follow-on research. For publications and links to these resources, Click here.

At MAIL, we combine molecular simulation, statistical learning, and language models to understand biological interactions and support molecular discovery. Our research connects sequence, structure, and dynamics, from biomolecule–material interfaces to protein behavior and computational drug-discovery workflows.

Biomolecular Interactions and Recognition

We use molecular dynamics to investigate how DNA and other biomolecules interact with synthetic materials. Statistical learning identifies collective variables that summarize complex trajectories, helping explain molecular recognition and binding at these interfaces.

Protein Dynamics and Small Molecule Interactions

Proteins change shape as they function and interact with other molecules. We analyze simulation trajectories to identify reaction coordinates and conformational transitions, turning high-dimensional motion into more interpretable descriptions of protein behavior.

Machine Learning Frameworks for Protein Analysis

Our models examine residue-level dynamics and sequence representations to study protein function, flexibility, and disorder. These complementary views help connect local molecular behavior with larger structural changes.

Sequence-Based Binding Affinity Prediction

Protein and ligand language representations provide a route to property prediction when detailed structures are unavailable. Building on this computational perspective, AgentD (2026) coordinates retrieval, molecular generation, property prediction, refinement, and protein–ligand evaluation within an early-stage drug-discovery workflow.

Tissue Engineering and Organoid Modeling

Our interdisciplinary collaborations also examine how engineered tissues reproduce biological organization. Recent work on extracellular-matrix-incorporated airway organoids (Biomaterials, 2026) studies apicobasal tissue polarity, extending the lab’s collaborative interests beyond molecular-scale systems.

For research details and recent publications, Click here.