Scientific learning · Evidence engineering · Mathematical data science

Designing Scientific Learning and Discovery Systems

We study how limited observations, experiments, and computational resources can yield reliable, testable scientific knowledge that future research can build on. Our current focus is AI for understanding mechanisms and guiding design in molecular and materials science.

Explore our research See selected work

Learn better. Create better evidence. Discover deeper. Design forward.

One connected scientific system

Our research connects the design of scientific learning systems, the creation and acquisition of evidence, and mathematical and data-science exploration.

Explore the connections

One connected scientific system

Open a concept to explore its role.

EvidenceCreate and acquire what matters

Choose observations, perturbations and experiments that can distinguish competing explanations and change the next learning step.

Explore evidence engineering

Evidence informs the learner; the learner helps choose new evidence.

Design the learner together

SpaceRepresent what matters

Organise scientific states, coordinates and scales so that relevant structure, symmetries and uncertainty can be represented.

Explore representation design

InteractionExpose meaningful relations

Design how information is exchanged, transformed and combined, from local relationships to collective and cross-scale effects.

Explore interaction design

LearningUpdate from evidence

Design learning goals, feedback and constraints together to study how the system changes as evidence arrives.

Explore learning dynamics

Each choice changes what the others can express, combine and learn.

Research aims to test against evidence

MechanismAn explanation to challenge

Seek explanations with consequences that can be tested against alternatives. A model pattern alone does not establish a mechanism.

Explore what evidence can establish

DesignA proposal to test

Propose candidates or interventions under scientific constraints, then use new evidence to test whether they achieve the intended effect.

Explore discovery and design

Tests of proposed mechanisms and designs return new evidence, revised questions and constraints to the learner.

01 · Selected work

Research in action.

Selected work connects representation learning, cross-scale analysis, multimodal generation and active discovery. Published papers and accepted conference work are identified below.

2026 · Journal article
An electron-density point-cloud framework for robust protein-ligand interaction prediction
Robust representation and interaction learning
Yujian Liu^#, Yutong Wang^#, Qingquan Wang^#, Meitang Peng, Yuan Chen, Yuechuan Lin, Dongxu Shen, Xiaoli Liu, Shidang Xu^*, Bin Liu
Nature Communications, 17(1), 7424 (2026)
Paper details for An electron-density point-cloud framework for robust protein-ligand interaction prediction Publisher record for An electron-density point-cloud framework for robust protein-ligand interaction prediction Download BibTeX for An electron-density point-cloud framework for robust protein-ligand interaction prediction
E-CloudBind: from electron-density representations to binding-affinity prediction.
Figure 2 · Yujian Liu et al., Nature Communications 17, 7424 (2026). © The Author(s) 2026. Full figure reproduced; display size and image format only are adjusted. · Published figure · CC BY 4.0
View full-size figure
2026 · Conference paper · Accepted
PyraE2E: Enhancing End-to-End WSI Analysis via Cross-Scale Super-Resolution
Cross-scale learning for whole-slide images
Yuechuan Lin^#, Yujian Liu^#, Weipeng Zhang, Yanyu Fan, Zikang Wang, Dongxu Shen, Liqin Fei, Xiaoli Liu, Shidang Xu^*
European Conference on Computer Vision, 2026 — Accepted
Paper details for PyraE2E: Enhancing End-to-End WSI Analysis via Cross-Scale Super-Resolution Download BibTeX for PyraE2E: Enhancing End-to-End WSI Analysis via Cross-Scale Super-Resolution
PyraE2E: cross-scale supervision connects patch sampling, representation learning and slide prediction.
Yuechuan Lin et al. · ECCV 2026 authors’ poster, overview panels (a–b). Cropped and resized from the full poster. · Full conference poster
View full-size figure
2026 · Conference paper · Accepted
AnyAvatar: High-Fidelity Gaussian Head Avatars under Uncalibrated Camera Settings
3D avatar reconstruction with uncalibrated cameras
Yujian Liu^#, Dongxu Shen^#, Haoran Li^#, Yuting Liu, Chuang Chen, Xinyi Jiang, Zhupeng Jiang, Peng Cao^*, Shidang Xu^*, Xiaoli Liu^*
ACM International Conference on Multimedia, 2026 — Accepted
Paper details for AnyAvatar: High-Fidelity Gaussian Head Avatars under Uncalibrated Camera Settings Download BibTeX for AnyAvatar: High-Fidelity Gaussian Head Avatars under Uncalibrated Camera Settings
AnyAvatar: camera-pose refinement, Gaussian head representation and novel-view rendering.
Yujian Liu et al., AnyAvatar authors' project figure. Full figure reproduced; display size and image format only are adjusted. Noncommercial reuse. · Authors’ project figure · CC BY-NC 4.0
View full-size figure
2025 · Conference paper
SyncAnimation: A Real-Time End-to-End Framework for Audio-Driven Human Pose and Talking Head Animation
Audio-driven multimodal generation
Yujian Liu^#, Shidang Xu^#, Jing Guo, Dingbin Wang, Zairan Wang, Xianfeng Tan, Xiaoli Liu^*
Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence (IJCAI-25), 1657–1665 (2025)
Paper details for SyncAnimation: A Real-Time End-to-End Framework for Audio-Driven Human Pose and Talking Head Animation Publisher record for SyncAnimation: A Real-Time End-to-End Framework for Audio-Driven Human Pose and Talking Head Animation Download BibTeX for SyncAnimation: A Real-Time End-to-End Framework for Audio-Driven Human Pose and Talking Head Animation
SyncAnimation: audio-conditioned pose, expression and progressive avatar rendering.
Figure 2 · Yujian Liu et al., SyncAnimation, arXiv:2501.14646v1 (2025). Full figure reproduced; display size and image format only are adjusted. · Authors’ paper figure · CC BY 4.0
View full-size figure
2021 · Journal article
Self-Improving Photosensitizer Discovery System via Bayesian Search with First-Principle Simulations
Active learning and a self-improving discovery loop
Shidang Xu, Jiali Li, Pengfei Cai, Xiaoli Liu, Bin Liu^*, Xiaonan Wang^*
In Journal of the American Chemical Society
Paper details for Self-Improving Photosensitizer Discovery System via Bayesian Search with First-Principle Simulations Publisher record for Self-Improving Photosensitizer Discovery System via Bayesian Search with First-Principle Simulations Download BibTeX for Self-Improving Photosensitizer Discovery System via Bayesian Search with First-Principle Simulations
Self-improving photosensitizer discovery: Bayesian search connects uncertainty, predicted properties and active learning.
Discovery-system diagram · Shidang Xu et al., J. Am. Chem. Soc. 143 (2021), 19769–19777. Existing author-website image; only display size and format are adjusted. · Published article · ACS author reuse policy
View full-size figure

^* Corresponding author ^# Co-first author (equal contribution)

Browse all publications

02 · Research system

Three contributions. One discovery system.

Our research connects the design of scientific learning systems, the creation and acquisition of evidence, and mathematical and data-science exploration.

Pillar IThe scientific learner

Design how scientific learning systems work

Foundational Scientific Learning-System Design

How should a scientific system learn?

Representation
System architecture
Scientific evaluation

Explore Pillar I

Pillar IIThe evidence environment

Create evidence that changes what can be learned

Evidence Engineering and Closed-Loop Discovery & Design

What evidence should the system create next?

Active acquisition
Perturbation
Closed loops

Explore Pillar II

Pillar IIIThe mathematical lens

Explore mathematics and data science for scientific discovery

Mathematical Data Science and Frontier Exploration

What mathematical and data-science methods can reveal important structure and open new scientific possibilities?

Mathematical modelling
Scientific computing
Data analysis

Explore Pillar III

Enter the research system Explore code, data and resources

03 · Scientific questions

Scientific questions that shape our methods

We investigate mechanistic questions in molecular and materials science to advance scientific understanding, guide design, and develop AI methods. We draw on observations, experiments, and simulations across diverse systems and scales to test and refine these methods.

A porous material and vesicle, an abstract folded protein ribbon, and layered image, signal, and relational patterns share a blue framework. The forms and patterns are illustrative. — A conceptual illustration of materials and delivery systems, proteins and sequences, and multimodal scientific data as settings for developing and testing learning methods.

Molecular science

How molecular structure, electronic states, interactions, and transformations shape behaviour and function.

Materials science

How composition, structure, interfaces, and processing shape material properties and responses across organic, inorganic, metallic, and hybrid systems.

04 · People and culture

Different training. Shared scientific questions.

Our research programme connects scientific learning, evidence creation and mathematical questions.

Our team works across biomedical engineering, pharmacy, chemistry, materials, and machine learning. Individual profiles document each researcher’s current role, training, and research interests.

Meet the people Explore group moments

Xu Lab members at the 2025 welcome dinner for the incoming class. — Welcome dinner for the incoming class · 2025

05 · Latest news

From the lab.

Recent entries from our group record, with their original event dates.

September 2026 · Lab update
Hongyuan Zhou joins the group as a master's student
September 2026 · Lab update
Dr. Chi Zhihe joins the group as a postdoctoral researcher
10 Jun 2026 · Lab update
Graduation Ceremony 2026

All news

Xu Lab · How we research, learn, and grow

Curiosity, independent thought, and work that lasts.

We share how questions take shape, how evidence changes our minds, and how we learn through building together. Explore the principles that guide research and growth in Xu Lab.

Explore our culture

06 · Work with us

Bring a scientific question worth stress-testing.

We welcome postdoctoral researchers, PhD candidates and collaborators interested in model architectures, learning algorithms, evidence acquisition and scientific intelligence. Molecular, biological and material systems offer rich settings for developing and testing these methods.

Join or collaborate Meet the team