Research Summary

My research lies at the intersection of physics and artificial intelligence (AI), known as AI4Physics and Physics4AI.

In my work on Physics4AI, I design physical neural networks that harness the inherent dynamics of physical systems to perform machine learning tasks more efficiently and robustly. My work on AI4Physics focuses on interpretable machine learning for modeling and understanding complex physical systems such as coral reefs, origami systems and viscoelastic materials.

Besides, I study topological metamaterials with nontrivial topological features, combining theoretical analysis and computational modeling to design metamaterials with robust, defect-immune behaviors and novel dynamic responses.

Image credit: DALL-E 3

Research Interests

Physics4AI

Physical neural networks are systems where physical components interact in a network-like structure, often to perform computations, process information, or respond to external stimuli. These networks can be electrical, fluidic, mechanical, or even biological in nature. This approach not only leverages the natural laws governing physical phenomena but also enables faster, energy‐efficient computation by directly integrating hardware and physics-based constraints into neural network architectures.

Mechanical neural networks are a subclass of physical networks implemented using mechanical elements (e.g., springs). Instead of artificial neurons and weights represented in software, mechanical neural networks use physical mechanical components—such as springs, beams, masses, and dampers—to store and process information. We develop the learning rule for the mechanical neural networks in the static regime, satisfying local rules where only local information is used to update the networks.

Wave-based physical neural networks offer distinct advantages over static physical neural networks, especially in terms of speed, parallelism, and computational capability. It allows information to be encoded by multiple waveforms and travel rapidly through the network. Since waves naturally propagate through multiple channels simultaneously, they can perform high-dimensional parallel processing. We also develop the learning rule for the mechanical neural networks beyond the static regime, satisfying local rules where only local information is used to update the networks.

AI4Physics

By incorporating optimization and classical models, I aim to produce algorithms that offer transparent insights into the underlying mechanisms of natural phenomena. This dual strategy of fusing physics with data-driven methods ultimately bridges the gap between theory and practice—providing models that are both high-performing and explainable, with broad applications in complex systems such as materials science, biological systems and beyond.

Structural dynamics is a field of engineering that studies the response of structures to dynamic loads, including forces from wind, earthquakes, impacts, and vibrations. Traditionally, structural dynamics has relied on physics-based models derived from first principles. However, data-driven methods are now revolutionizing the field by incorporating experimental data and computational techniques. We use the simulation data and experimental data to establish the interpretable data-driven model for structural dynamics, including wave dynamics in topological metamaterials, vibration in origami structures and impact loading on viscoelastic materials.

Coral reefs harbor great biodiversity, support 25% of all marine life and hold immense ecological and economic value. Reef-building corals are inherently sessile organisms, however, motion is an important behavioral trait of coral. Corals use their polyps and tentacles to capture nutrients, expel waste, and interact with their surroundings. This motion is typically slow and driven by ciliary action or muscle contractions. Soft corals and some branching hard corals can sway in response to water currents. The hydrodynamic interaction between corals and water flow affects their nutrient uptake and structural integrity. Notwithstanding the importance of inherent temporal and spatial multi-scale features of coral motion, its quantitative properties and modeling still remain challenging and unexplored. We use the long-period observation, numerical analysis and theoretical/data-driven modeling to describe and understand the coral motion biologically and physically.

Soft/Flexible Topological Metamaterials

Topological elastic metamaterials offer insight into classic motion law and open up opportunities in quantum and classic information processing. Topologically protected wave propagation possesses prominent applications in quantum computation and communication field due to its remarkable characteristic: the robust defect-immune transport. Recently, significant research efforts devoted in phononic topological insulators provide a new way to manipulate sound propagation, such as vibration isolation and particle manipulation. Soft materials are capable of deformation so that they possess a large tunability. The combination of soft materials with high-frequency topological states offers unprecedented opportunities, which requires insight exploration. We design a soft metamaterial with honeycomb pattern with reversible topological phase and dynamically tunable topological states.

Recently, valley—the degenerate yet inequivalent energy extrema in momentum space—has emerged as a dimension in manipulating waves in electronics, photonics, and phononics. In graphene and transition metal dichalcogenides (TMD), the valley Hall effect has been studied for the promising applications in information carrier and storage. As the concept of valley is introduced into the classic system, the photonic and phononic valley crystals have also been proposed, showing valley-dependent energy transportation. Likewise, various designs of elastic valley metamaterials have been reported based largely on the two different types: TMD-inspired hexagonal lattices, and triangular lattices with triangle-like scatterers. We design a novel spiral elastic valley metamaterials by introducing the combination of soft matrix and hard sprial, which is able to have topological transition and propagation of topological valley states.

Topological origami is a fascinating field that combines principles of origami (paper folding) with topology and mechanics to design structures with programmable shape transformations, tunable mechanical properties, and even novel material behaviors. It is widely used in metamaterials, robotics, and deployable structures. Topological protection can guide waves in a controlled manner, preventing unwanted vibrations or impacts. We design various topological systems using Kresling origami and Miura-folded origami, realizing on-demand topological states, chiral Landau level and topological pumping.