Han Liu / Research program
We develop the foundations and systems for AI that can discover how the world works—and learn how to build within it.
From discovery to invention, our research asks how AI can reason about the physical world, design interventions, and improve from verifiable outcomes. Our central thesis: each research cycle should improve both what an AI knows and how it chooses what to do next.
A research architecture for scientific intelligence
Verifiable outcomes → Reusable experience → Better decisions
Research agenda
How can an agent get better at deciding what to investigate next?
We study agents that retain useful experience, revise hypotheses, and improve their use of scientific tools through feedback. Memory, reasoning, adaptation, and in-context computation provide the foundations. We evaluate improvement through the quality of subsequent decisions and measured outcomes: better predictions, more informative experiments, and more effective designs.
What must a model understand to guide scientific discovery?
We develop models of scientific objects, processes, and dynamics, from genomes, proteins, and cells to molecular systems and the changing universe. We investigate how these representations can support prediction, uncertainty quantification, and reasoning about interventions. Biology and astronomy provide complementary settings for studying these questions.
How can each experiment make the next one more informative?
We develop methods and systems for connecting scientific agents with molecular design, physics-based simulation, robotics, and laboratory instruments. Our focus is the full Design–Build–Test–Learn cycle: choosing experiments, executing them, preserving measurement context, and using successes and failures to guide the next design.
The unifying research agenda
Understand what is possible.
Learn how to build it.
Scientific discovery and engineering design share a central challenge: reason about the world, act on it, observe the consequences, and improve from verifiable outcomes.
Evidence informs the next hypothesis.
Progress means stronger explanations, calibrated predictions, and more informative experiments.
Measured performance informs the next design.
Progress means better performance, lower cost, and designs that meet physical and practical constraints.
Our ambition is to make experience compound. Existing contributions advance individual parts of this agenda—from genomic models and protein-design policies to research agents and simulation. Connecting them into systems that learn across discovery and engineering cycles is the research opportunity ahead.
Selected contributions
Representative work with our students and collaborators, spanning scientific representations, protein design policies, research agents, and embodied execution.
From pretrained representations of DNA to efficient modeling across species. DNABERT-2 pairs a more efficient genome model with the Genome Understanding Evaluation benchmark, making model quality and computational cost jointly testable.
Genome foundation models trained on large metagenomic assemblies, with released models up to four billion parameters. The work connects genomic representation learning with sequence generation, supported by public code and model checkpoints.
A joint-embedding predictive approach to single-cell transcriptomics. Predicting latent cellular representations from partial gene-expression observations helps the model learn features that are robust to missing measurements.
Trained on measured protein outcomes, PEV learns a policy for choosing single-residue substitutions or STOP. It scores candidate edits in parallel from one parent encoding, exploring how experimental experience can inform new protein-design decisions.
A prototype AI collaborator for the science of science, developed with Dashun Wang and collaborators. It brings language models into analytical workflows, supporting research iteration and reproducibility.
A unified simulation runtime for embodied agents, connecting task specifications, robot control, planning, evaluation, and trajectory collection. High-level commands execute through robot actions, producing structured multimodal records for learning and evaluation.
Flagship programs
Liu helps lead complementary efforts that connect this research agenda to experimental infrastructure and scientific communities.
Co-PI · AI lead
DREAM is being developed as a national cloud laboratory for protein engineering. The program connects AI-driven design, cell-free protein synthesis, automated measurements, and iterative learning. Our work on its AI systems and data infrastructure aims to make experimental evidence a foundation for better models and scientific decisions.
Explore the DREAM Cloud Lab ↗Director · Northwestern University
The center brings together researchers across disciplines to advance foundation models and generative AI, connecting fundamental research with problems in science and engineering.
Explore CFMG ↗Co-lead · NSF–Simons SkAI Institute
TEMPI extends the scientific world-model agenda to astronomy: learning representations of physical dynamics from time-series observations of the changing universe.
Explore TEMPI ↗Intellectual foundations
Our current program grows out of contributions to nonparametric graphical models, nonconvex optimization, and high-dimensional inference. These ideas remain central: learning structure, quantifying uncertainty, and making reliable decisions from incomplete evidence.
Selected recognition
PECASE · Alfred P. Sloan Fellowship · IMS Tweedie New Researcher Award
ASA Noether Young Scholar Award · NSF CAREER Award
