Research direction 03

Machine learning

Multiscale electron–phonon materials

I use machine learning as a structural engine that extends electron and phonon workflows beyond small-cell DFT to realistic length, time, temperature, and configuration scales.

High-throughput discovery
Active learning
ML potentials
Multiscale dynamics

Featured work

Learning across materials and scales.

These projects use data and learned models to reach structures, dynamics, and functions that direct first-principles calculations cannot access alone.

Stress and energy changes during machine-learning molecular dynamics with atomistic snapshots of a layer-by-layer phase transition in titanium oxide — Stress and energy steps track successive transformed layers in a 165,888-atom trajectory. Figure 3 from Liu et al., Nature Communications 15, 3079 (2024), CC BY 4.0.

01 Nature Communications · 2024

Active learning reveals a layer-by-layer phase transition.

Reconstructive transitions require chemical bonds to break and reform, while direct first-principles molecular dynamics cannot reach the full length and time scales of the process.

On-the-fly active learning, a moment-tensor potential, and advanced sampling enabled 165,888-atom trajectories. They show the β→λ transition in Ti₃O₅ nucleating in the ab plane and advancing layer by layer along c.

Active learning
ML potential
Large-scale MD

View publication

Transfer-learning workflow from untwisted molybdenum ditelluride data to neural-network force prediction and relaxed twisted bilayers — Transfer learning connects untwisted training data with large-scale relaxation of twisted MoTe₂. Figure 5 from Mao et al., Communications Physics 7, 262 (2024), CC BY 4.0.

02 Communications Physics · 2024

Transfer learning reaches realistic moiré scales.

Transfer learning adapts a DeePMD neural-network potential trained on untwisted MoTe₂ to relax moiré supercells down to a 1.1° twist angle, extending beyond the 2.88° limit reached with large-scale plane-wave DFT.

The relaxed structures bridge atomic reconstruction, first-principles electronic bands, and a continuum model, revealing robust Chern bands and strong strain-induced gauge fields across experimentally relevant twist angles.

Transfer learning
Twisted MoTe₂
Moiré relaxation

View publication

Pixels → device → network

→

±V

→

sense · remember · preprocess

Original concept illustration: one bias-controlled device connects optical sensing with neuromorphic preprocessing.

03 Advanced Materials · 2023

One device, two modes of visual intelligence.

A bias-switchable architecture converts a single device between fast photodetection and neuromorphic vision sensing. The same platform can acquire optical information, retain it, and perform early-stage preprocessing.

This dual-mode behavior offers a compact route from sensing to memory and neural-network-ready data without separating each function into a different component.

Neuromorphic vision
Photodetection
In-sensor computing

View publication

Tree-structured genetic programs with hoist mutation, subtree mutation, and crossover operations used in symbolic regression — Tree-structured programs and the mutation and crossover operators used in symbolic regression. Figure 1 from Zhao et al., The European Physical Journal Plus 138, 643 (2023). © The Author(s) 2023, under exclusive licence to Società Italiana di Fisica and Springer-Verlag GmbH Germany, part of Springer Nature.

04 The European Physical Journal Plus · 2023

Symbolic regression makes materials models interpretable.

Interpretable symbolic regression yields compact, transferable models of Vickers hardness and fracture toughness using only the bulk and shear moduli as inputs.

Coupled with machine-learned elastic moduli, these equations create an efficient high-throughput route to screen for materials that combine superhardness with fracture resistance.

Symbolic regression
Interpretable ML
Mechanical properties

View publication

Workflow for automated high-throughput screening and classification of topological phonons — Automated collection, stability filtering, topology diagnosis, and database construction. Figure 1 from Li et al., Nature Communications 12, 1204 (2021), CC BY 4.0.

05 Nature Communications · 2021

Computation becomes a reusable discovery platform.

A high-throughput workflow combined phonon data, automated stability filters, band-crossing searches, and topological classification across more than 10,000 materials.

The resulting database demonstrates an essential ML foundation: turn domain knowledge into structured, reproducible data that can guide the next calculation and the next experiment.

High throughput
Scientific data
Automation

View publication

Continue across the research program.

Learned structural models become most powerful when they connect lattice dynamics with correlated electronic function.

Research direction 01Phonons Research direction 02Electrons

Let’s connect

Interested in ideas, collaborations, and academic opportunities.

Start a conversation

Google Scholar ↗GitHub ↗ORCID ↗