Predictive Intelligence Lab

Current Research

  • Foundation models for physical simulation

Our vision is to create foundation models that can transform physical simulation across multiple scientific domains. These models aim to learn from vast and diverse datasets to develop a deep understanding of physical phenomena. By leveraging the power of machine learning and big data, we’re working towards AI systems that can generalize across different physical problems, adapting their knowledge to new scenarios with minimal additional training. This approach has the potential to transform fields like climate science, materials engineering, and fluid dynamics by providing faster, more accurate, and more flexible simulation tools. Our goal is to develop AI that can assist scientists and engineers in tackling complex physical problems, accelerating scientific discovery and technological innovation.

Recent Publications:

Micrometer: Micromechanics Transformer for Predicting Mechanical Responses of Heterogeneous Materials

Aurora: A Foundation Model of the Atmosphere


  • Physics informed neural networks and neural operators

We are pioneering the integration of physical laws and domain knowledge into machine learning models through physics-informed neural networks and neural operators. This approach aims to create AI systems that don’t just learn from data, but also respect fundamental physical principles. By embedding scientific knowledge directly into our models, we can achieve more accurate and efficient simulations of complex physical systems, even with limited data. Our research focuses on developing novel architectures and training strategies that can handle a wide range of physical phenomena, from simple mechanical systems to complex fluid dynamics. The ultimate goal is to bridge the gap between traditional scientific computing and modern machine learning, creating tools that combine the best of both worlds to solve challenging scientific and engineering problems.

Recent Publications:

Bridging Operator Learning and Conditioned Neural Fields: A Unifying Perspective

PirateNets: Physics-informed Deep Learning with Residual Adaptive Networks

Δ-PINNs: physics-informed neural networks on complex geometries

An expert’s guide to training physics-informed neural networks


  • Generative models in science and engineering

Our research explores the potential of generative models to not just analyze but also synthesize scientific data and engineering designs. We are developing AI systems that can learn the underlying patterns and structures in complex scientific datasets and use this knowledge to generate new, physically consistent scenarios or designs. This work has far-reaching implications across multiple fields, from materials science to drug discovery. By enabling AI to generate novel scientific hypotheses or engineering solutions, we aim to accelerate the pace of innovation and discovery. Our vision is to create AI partners for scientists and engineers, capable of suggesting new experiments, predicting outcomes of complex systems, and even proposing innovative designs that humans might not have considered.

Recent Publications:

Guided autoregressive diffusion models with applications to PDE simulation

Pde-refiner: Achieving accurate long rollouts with neural PDE solvers

Variational autoencoding neural operators


  • Uncertainty quantification and sequential decision making

In the realm of uncertainty quantification and sequential decision making, we are working to develop AI systems that can make reliable predictions and informed decisions in uncertain environments. This is crucial in scientific and engineering contexts where data may be scarce, noisy, or incomplete. Our research focuses on creating methods that can not only provide predictions but also quantify the confidence in those predictions. We’re also developing strategies for sequential decision making, allowing AI systems to adapt their approach as new information becomes available. The goal is to create robust, trustworthy AI tools that can assist in complex decision-making processes in fields ranging from climate science to autonomous systems, enhancing our ability to navigate uncertain and dynamic environments.

Recent Publications:

Composite Bayesian Optimization In Function Spaces Using NEON–Neural Epistemic Operator Networks

Scalable Bayesian optimization with randomized prior networks

Output-weighted sampling for multi-armed bandits with extreme payoffs


  • Engineering Applications

Our overarching aim is to translate cutting-edge AI research into practical engineering solutions. We are collaborating with domain experts to apply our advancements in AI and machine learning to real-world engineering challenges across a broad spectrum of domains. This includes areas such as healthcare, where we’re developing tools for more accurate medical diagnostics and personalized treatment plans; environmental science, where we’re creating models for better climate and pollution predictions; and industrial design, where we’re using AI to optimize complex systems and processes. By bridging the gap between theoretical AI advancements and practical engineering needs, we aim to accelerate innovation, improve efficiency, and tackle some of the most pressing technological challenges of our time. Our vision is to create a new generation of AI-powered engineering tools that can work alongside human experts, augmenting their capabilities and opening new frontiers in scientific and technological progress.

Recent Publications:

Disk2Planet: A Robust and Automated Machine Learning Tool for Parameter Inference in Disk-Planet Systems

Accelerated design of architected materials with multifidelity Bayesian optimization

Feasibility of Vascular Parameter Estimation for Assessing Hypertensive Pregnancy Disorders

Physics-informed neural networks to learn cardiac fiber orientation from multiple electroanatomical maps