Research

1. Symmetry-aware AI for solid-state quantum materials

We develop machine learning frameworks that build physical principles (point-group symmetry, equivariance, local bonding motifs, and exact physical constraints) into the network itself rather than learning them from data, for accurate, data-efficient and interpretable prediction for quantum materials. - Equivariant networks for tensorial and spectral properties: TSENN predicts full frequency-dependent dielectric tensors about 1,000× faster than DFT (Nat. Commun. 2026); equivariant GNNs for tensor properties of crystals (arXiv:2406.03563). - Point-group equivariant graph neural networks for materials (arXiv:2607.16871). - Beyond-atom representations: motif-centric learning (AMDNet, Sci. Adv. 2021), crystal hypergraph convolutional networks (npj Comput. Mater. 2025), and material–motif heterogeneous graphs (Adv. Intell. Discov. 2026).

2. AI-driven inverse design of functional and quantum materials

Moving from property prediction to design: symmetry-respecting, property-steered generative models coupled with first-principles validation in a closed loop. - Symmetry- and property-aware crystal generation with reinforcement learning (SPARC, arXiv:2609.13468), which avoids the low-symmetry “reward-hacking” structures of surrogate-guided generators. - Targets include altermagnets, low-damping magnets, nonlinear-optical and Berry-curvature-dipole materials. - Toward foundation models and agentic workflows that combine equivariant models, generative design, high-throughput first-principles computation, and AI agents.

3. Machine learning for electronic structure across scales

End-to-end and Hamiltonian-based learning of electronic structure, from simple crystals to large twisted and disordered systems. - WANDER bridges deep-learning force fields and Wannier Hamiltonians, reaching twisted bilayers with more than 1,000 atoms at 10³–10⁴× the speed of DFT (npj Comput. Mater. 2025). - Bandformer, a graph Transformer for end-to-end band-structure prediction (arXiv:2411.16483). - Open data: tensorial optical and transport properties of 7,301 materials from automated Wannierization (Sci. Data 2025); physics-constrained density functionals via contrastive learning (Digital Discovery 2023).

4. Multilayer twisted quantum materials

Twisting stacks of three or more two-dimensional layers opens a vast design space of moiré superlattices whose flat bands, topology, chirality, and nonlinear optical responses are absent in the individual layers. We aim to make this space computable and designable. - A universal machine-learning Hamiltonian for twisted multilayers: equivariant models evaluate each layer in its own frame, and a twist-equivariant network learns the interlayer coupling. Trained on large-scale first-principles data spanning thousands of 2D materials, it targets the electronic structure and optical response of moiré cells far beyond the reach of DFT. - Symmetry analysis of arbitrary twisted stacks (point and space groups, chirality, polarity) to identify allowed second-harmonic generation and circular-dichroism responses, guiding the design of chiral and nonlinear-optical moiré materials. Explore twisted stacks in our Twisted Multilayer Builder (see Resources). - Deep-learning-guided twistronics for self-assembled quantum optoelectronics (NSF DMREF), and ideal topological flat bands in moiré heterostructures with type-II band alignment (arXiv:2507.06168).

5. Data-driven design of quantum defects in 2D materials

Symmetry-guided, high-throughput discovery of point defects for qubits, single-photon emitters, and quantum sensors in atomically thin materials. - A local-symmetry design principle identified antisite defect qubits in transition metal dichalcogenides (Nat. Commun. 2022; patent) and more than 40 quantum-defect candidates across binary 2D hosts (arXiv:2405.11379). - Machine learning for defects with persistent-homology features (Chem. Mater. 2025). - The Defect Genome Initiative (Adv. Mater. 2024) and the Roadmap on 2D Materials for Quantum Technologies (2D Mater., guest editor).

6. Disorder, alloys, and complex materials

Machine learning frameworks for configurational, chemical, and spin disorder in multicomponent alloys. - GNN-accelerated Monte Carlo for order–disorder transitions and configurational entropy (npj Comput. Mater. 2024), and ensemble properties of atomically disordered materials (ACS Nano 2025). - A multi-scale framework for coupled chemical, spin, and structural disorder in alloys (arXiv:2607.07456).