Searcharxiv⌕ Search

arXiv subjects

Tuwei Liao

Publications and source records attributed to Tuwei Liao.

2 recordsLinked to original sources

Scaling laws in machine learning on non-ideal experimental data enable inverse design of radiative-cooling coatings

Materials informatics increasingly targets multicomponent materials, yet experimental datasets inevitably contain non-ideal variations arising from imperfect uniformity and reproducibility in composition, coating formation, and measurement. Whether machine-learning prediction accuracy scales with amount of data under these conditions remains unclear. Here, we develop an automated platform that produced and optically characterized 11,520 radiative-cooling coatings spanning nine material and processing parameters. Forward and inverse neural networks, and a spectrum-conditioned diffusion model were trained on subsets of 10 to 10,000 samples. All models exhibited power-law scaling, with prediction errors decreasing with dataset size. Experimental inverse-design validation showed that larger datasets produced spectra closer to a broadband-emitter target. The diffusion model generated multiple candidate parameter sets, allowing selection among alternative solutions based on ease of fabrication, cost, and scalability. These results demonstrate predictable learning-curve scaling in non-ideal experimental data and provide a quantitative basis for estimating the benefit of additional experimental data, and thus the cost-effectiveness of materials-informatics.

cond-mat.mtrl-sci↗

Data-driven design of multilayer hyperbolic metamaterials for near-field thermal radiative modulator with high modulation contrast

The thermal modulator based on the near-field radiative heat transfer has wide applications in thermoelectric diodes, thermoelectric transistors, and thermal storage. However, the design of optimal near-field thermal radiation structure is a complex and challenging problem due to the tremendous number of degrees of freedom. In this work, we have proposed a data-driven machine learning workflow to efficiently design multilayer hyperbolic metamaterials composed of $α$-MoO$_{\rm 3}$ for near-field thermal radiative modulator with high modulation contrast. By combining the multilayer perceptron and Bayesian optimization, the rotation angle, layer thickness and gap distance of the multilayer metamaterials are optimized to achieve a maximum thermal modulation contrast ratio of 6.29. This represents a 97% improvement compared to previous single layer structure. The large thermal modulation contrast is mainly attributed to the alignment and misalignment of hyperbolic plasmon polaritons and hyperbolic surface phonon polaritons of each layer controlled by the rotation. The results provide a promising way for accelerating the designing and manipulating of near-field radiative heat transfer by anisotropic hyperbolic materials through the data-driven style.

physics.optics↗