Gradient-based optimization of complex nanoparticle heterostructures enabled by deep learning on heterogeneous graphs

Area of Science:

• Materials Science
• Nanotechnology
• Computational Chemistry

Background:

• Deep learning (DL) applications in nanomaterial design are limited by data representation and availability.
• Core-shell upconverting nanoparticles (UCNPs) offer promising applications in biosensing, microscopy, and 3D printing due to their unique light-emitting properties.

Purpose of the Study:

• To overcome data limitations and leverage DL for optimizing the nonlinear optical properties of UCNPs.
• To develop a DL framework for the inverse design of UCNPs with enhanced emission characteristics.

Main Methods:

• Generated a large dataset (>6,000) of UCNP emission spectra using kinetic Monte Carlo simulations.
• Trained a heterogeneous graph neural network with a physically informed representation of UCNP nanostructures.
• Employed gradient-based optimization on the trained network to discover new UCNP designs.

Main Results:

• Identified UCNP structures with a predicted 6.5x higher emission under 800-nm illumination compared to existing ones.
• Demonstrated the effectiveness of the DL approach in predicting and optimizing UCNP performance.
• Established design principles for UCNP heterostructures.

Conclusions:

• The developed DL framework successfully optimizes UCNP nonlinear optical properties.
• This work provides a roadmap for DL-based inverse design of nanomaterials.
• The findings enable the creation of UCNPs with superior light emission for diverse technological applications.