The Mechanical and Artificial Intelligence Lab (MAIL) at Carnegie Mellon University is dedicated to advancing research at the intersection of Mechanical Engineering and Data Science. Our mission is to harness cutting-edge machine learning (ML) and deep learning (DL) techniques to model, understand, and predict complex physical phenomena central to mechanical systems.

We focus on a wide range of topics that include multiphysics transport, fluid dynamics, heat transfer, and other foundational areas in mechanical engineering. These systems are often governed by nonlinear, multiscale interactions that are traditionally challenging to model using first-principles approaches alone. By integrating data-driven methods with physics-based frameworks, we aim to bridge this gap and accelerate scientific discovery and engineering design.

With the rapid proliferation of sensors, the Internet of Things (IoT), and cloud-based infrastructure, we are witnessing an unprecedented influx of heterogeneous, high-dimensional data streams. This explosion of sensory data presents both an opportunity and a challenge: how can we extract actionable insights and construct reliable models from sparse, noisy, and incomplete datasets?

Our lab is developing novel algorithms for learning from sparse spatiotemporal data—enabling accurate inference and prediction in scenarios where traditional modeling would fail. These tools are particularly relevant for engineering resilient infrastructure systems, where sensor coverage may be limited but decision-making must remain robust.

Applications of our research span smart cities, energy systems, autonomous vehicles, and next-generation manufacturing. By leveraging the synergy between mechanical engineering principles and AI methodologies, MAIL aims to redefine how we sense, model, and control complex physical environments. For more information,  click here

Another key research thrust in the Mechanical and Artificial Intelligence Lab (MAIL) is the acceleration of materials discovery using state-of-the-art machine learning methodologies. Traditional approaches—whether experimental or computational—are often prohibitively time-consuming and resource-intensive when exploring vast design spaces for novel materials, particularly for energy and biomedical applications.

By leveraging deep learning, we aim to dramatically reduce the time and cost associated with identifying and optimizing materials with desired physical, chemical, or functional properties. Our work focuses on developing predictive models that can learn from existing materials databases, simulate material behavior, and suggest new candidates with tailored performance metrics.

In particular, we are interested in combining domain knowledge with modern AI techniques such as generative modeling, graph neural networks, and active learning frameworks. These tools enable efficient navigation of the materials design space and allow for intelligent data acquisition, guiding both simulation and experimental efforts.

Our goal is to create a closed-loop, AI-driven pipeline that accelerates the discovery of next-generation materials for energy storage, catalysis, biomedical devices, and more. For more detailed information on our methods, datasets, and current projects, please click here

At the Mechanical and Artificial Intelligence Lab (MAIL), we are pioneering new approaches to robotic intelligence by combining deep learning, reinforcement learning (RL), and transformer-based architectures to create robots that behave in increasingly intelligent, adaptive, and human-like ways.

Traditional control methods in robotics—while effective for well-defined tasks—often struggle in the presence of noise, uncertainty, or complex environmental dynamics. Our lab is addressing this limitation by building AI-driven control systems that can learn from demonstration, adapt through interaction, and make decisions under real-world constraints.

We are especially focused on the integration of Diffusion Policies, Adversarial Cloned Transformers (ACT), and multimodal visuotactile sensing to enhance robotic perception and policy learning. These advanced architectures allow robots to:

• Learn from limited demonstrations using diffusion-based generative models to model and predict action sequences

• Develop transformer-based policies that scale with complex spatiotemporal patterns in data

• Fuse vision and tactile sensing to understand physical interactions in a more nuanced and human-like manner

Our research has shown that these models enable persistent, creative, and constraint-aware decision-making in robotic agents—qualities essential for real-world deployment. Whether it's manipulating delicate objects, adapting to unseen environments, or coordinating between multiple sensor modalities, our AI-driven policies are pushing the boundaries of what autonomous systems can achieve.

We envision a future where robots possess the intelligence to reason, feel (through sensors), and act with purpose. Through our work, we aim to close the gap between human cognition and machine autonomy.

For more information on our projects and recent publications,  click here

At the Mechanical and Artificial Intelligence Lab (MAIL), we are at the forefront of integrating advanced AI methodologies with Additive Manufacturing (AM) processes to enhance part quality, process reliability, and accessibility. Our research focuses on leveraging machine learning (ML) and deep learning (DL) techniques to address the complexities inherent in AM, particularly in metal-based processes.

Real-Time Monitoring and Control

We are developing AI-powered systems capable of real-time monitoring and control of the melt pool dynamics during AM processes. By employing computer vision and DL models, such as convolutional neural networks (CNNs) and vision transformers, we can detect and classify anomalies like porosity, keyholing, and spatter formation. These insights enable adaptive control strategies that adjust process parameters on-the-fly to mitigate defects and ensure consistent build quality.

Predictive Modeling and Process Optimization

Our lab is also focused on creating predictive models that correlate process parameters with resultant part properties. Utilizing physics-informed neural networks (PINNs), we incorporate fundamental physical laws into our ML models to predict outcomes like melt pool geometry and thermal gradients. This approach enhances the accuracy of simulations and aids in optimizing process parameters for desired material properties.

Open-Source Tools and Data Repositories

To foster collaboration and accelerate innovation in the AM community, we are committed to developing open-source software tools and establishing comprehensive data repositories. These resources are designed to facilitate the adoption of AI techniques in AM, providing researchers and practitioners with accessible platforms for process monitoring, defect detection, and quality assurance. For more information, Click here

At the Mechanical and Artificial Intelligence Lab (MAIL), we are pioneering the integration of advanced AI methodologies with molecular simulations to unravel the complexities of biomolecular interactions and dynamics. Our interdisciplinary approach combines molecular dynamics (MD) simulations with machine learning (ML) and statistical learning techniques to gain deeper insights into biological systems, with applications spanning from fundamental biophysics to drug discovery.

1. Biomolecular Interactions and Recognition

Understanding how biomolecules such as DNA interact with synthetic materials is crucial for applications in biosensing and nanotechnology. We employ MD simulations to generate detailed trajectories of these interactions. To distill meaningful patterns from the high-dimensional data, we apply statistical learning methods to identify collective variables and reaction coordinates. This fusion of statistical mechanics and machine learning enables us to capture the essential dynamics governing biomolecular recognition processes.

2. Protein Dynamics and Small Molecule Interactions

Proteins exhibit complex conformational changes that are vital to their function. To study these dynamics, especially in the context of small molecule interactions, we analyze extensive MD simulation data. By implementing dimensionality reduction techniques and ML models, we identify key reaction coordinates that describe protein conformational landscapes. This approach aids in elucidating mechanisms of protein function and facilitates the identification of potential therapeutic targets.

3. Machine Learning Frameworks for Protein Analysis

We have developed novel ML frameworks to characterize the dynamic behavior of amino acids within proteins. For instance, our models can classify residues based on their conformational switching behavior, distinguishing between stable and unstable switches. Such classifications provide insights into the roles of specific residues in protein function and stability, enhancing our understanding of protein mechanics at the residue level.

4. Sequence-Based Binding Affinity Prediction

Traditional methods for predicting binding affinities between proteins and ligands often rely on 3D structural data, which can be resource-intensive to obtain. To address this, we've introduced sequence-based models that leverage protein and ligand sequences to predict binding affinities. These models utilize advanced language representations to capture the underlying biochemical interactions, offering a scalable and efficient alternative for in-silico drug discovery. For more information, Click here